release: 1.0.2

docker builder update

.cargo/config.toml

@@ -0,0 +1,2 @@
+[alias]
+eval = "run -p evaluations --"

.gitignore

@@ -10,6 +10,9 @@ result
 data
 database
 
+evaluations/cache/
+evaluations/reports/
+
 # Devenv
 .devenv*
 devenv.local.nix
@@ -21,3 +24,4 @@ devenv.local.nix
 .pre-commit-config.yaml
 # html-router/assets/style.css
 html-router/node_modules
+.fastembed_cache/

CHANGELOG.md

@@ -1,6 +1,26 @@
 # Changelog
 ## Unreleased
 
+## 1.0.2 (2026-02-15)
+- Fix: edge case where navigation back to a chat page could trigger a new response generation
+- Fix: chat references now validate and render more reliably
+- Fix: improved admin access checks for restricted routes
+- Performance: faster chat sidebar loads from cached conversation archive data
+- API: harmonized ingest endpoint naming and added configurable ingest safety limits
+- Security: hardened query handling and ingestion logging to reduce injection and data exposure risk
+
+## 1.0.1 (2026-02-11)
+- Shipped an S3 storage backend so content can be stored in object storage instead of local disk, with configuration support for S3 deployments.
+- Introduced user theme preferences with the new Obsidian Prism look and improved dark mode styling.
+- Fixed edge cases, including content deletion behavior and compatibility for older user records.
+
+## 1.0.0 (2026-01-02)
+- **Locally generated embeddings are now default**. If you want to continue using API embeddings, set EMBEDDING_BACKEND to openai. This will download a ONNX model and recreate all embeddings. But in most instances it's very worth it. Removing the network bound call to create embeddings. Creating embeddings on my N100 device is extremely fast. Typically a search response is provided in less than 50ms. 
+- Added a benchmarks create for evaluating the retrieval process
+- Added fastembed embedding support, enables the use of local CPU generated embeddings, greatly improved latency if machine can handle it. Quick search has vastly better accuracy and is much faster, 50ms latency when testing compared to minimum 300ms.
+- Embeddings stored on own table.
+- Refactored retrieval pipeline to use the new, faster and more accurate strategy. Read [blog post](https://blog.stark.pub/posts/eval-retrieval-refactor/) for more details.
+
 ## Version 0.2.7 (2025-12-04)
 - Improved admin page, now only loads models when specifically requested. Groundwork for coming configuration features.
 - Fix: timezone aware info in scratchpad

Cargo.toml

@@ -5,8 +5,9 @@ members = [
   "api-router",
   "html-router",
   "ingestion-pipeline",
-  "composite-retrieval",
-  "json-stream-parser"
+  "retrieval-pipeline",
+  "json-stream-parser",
+  "evaluations"
 ]
 resolver = "2"
 
@@ -39,9 +40,11 @@ serde_json = "1.0.128"
 serde = { version = "1", features = ["derive"] }
 sha2 = "0.10.8"
 surrealdb-migrations = "2.2.2"
-surrealdb = { version = "2", features = ["kv-mem"] }
+surrealdb = { version = "2" }
 tempfile = "3.12.0"
-text-splitter = "0.18.1"
+text-splitter = { version = "0.18.1", features = ["markdown", "tokenizers"] }
+tokenizers = { version = "0.20.4", features = ["http"] }
+unicode-normalization = "0.1.24"
 thiserror = "1.0.63"
 tokio-util = { version = "0.7.15", features = ["io"] }
 tokio = { version = "1", features = ["full"] }
@@ -53,7 +56,7 @@ url = { version = "2.5.2", features = ["serde"] }
 uuid = { version = "1.10.0", features = ["v4", "serde"] }
 tokio-retry = "0.3.0"
 base64 = "0.22.1"
-object_store = { version = "0.11.2" }
+object_store = { version = "0.11.2", features = ["aws"] }
 bytes = "1.7.1"
 state-machines = "0.2.0"
 fastembed = { version = "5.2.0", default-features = false, features = ["hf-hub-native-tls", "ort-load-dynamic"] }
@@ -77,7 +80,7 @@ implicit_clone = "warn"
 redundant_clone = "warn"
 
 # Security-focused lints
-integer_arithmetic = "warn"
+arithmetic_side_effects = "warn"
 indexing_slicing = "warn"
 unwrap_used = "warn"
 expect_used = "warn"
@@ -87,7 +90,7 @@ todo = "warn"
 
 # Async/Network lints
 async_yields_async = "warn"
-await_holding_invalid_state = "warn"
+await_holding_invalid_type = "warn"
 rc_buffer = "warn"
 
 # Maintainability-focused lints
@@ -106,6 +109,8 @@ wildcard_dependencies = "warn"
 missing_docs_in_private_items = "warn"
 
 # Allow noisy lints that don't add value for this project
-manual_must_use = "allow"
 needless_raw_string_hashes = "allow"
 multiple_bound_locations = "allow"
+cargo_common_metadata = "allow"
+multiple-crate-versions = "allow"
+module_name_repetition = "allow"

Dockerfile

@@ -1,15 +1,15 @@
 # === Builder ===
-FROM rust:1.86-bookworm AS builder
+FROM rust:1.89-bookworm AS builder
 WORKDIR /usr/src/minne
 RUN apt-get update && apt-get install -y --no-install-recommends \
     pkg-config clang cmake git && rm -rf /var/lib/apt/lists/*
 
 # Cache deps
 COPY Cargo.toml Cargo.lock ./
-RUN mkdir -p api-router common composite-retrieval html-router ingestion-pipeline json-stream-parser main worker
+RUN mkdir -p api-router common retrieval-pipeline html-router ingestion-pipeline json-stream-parser main worker
 COPY api-router/Cargo.toml ./api-router/
 COPY common/Cargo.toml ./common/
-COPY composite-retrieval/Cargo.toml ./composite-retrieval/
+COPY retrieval-pipeline/Cargo.toml ./retrieval-pipeline/
 COPY html-router/Cargo.toml ./html-router/
 COPY ingestion-pipeline/Cargo.toml ./ingestion-pipeline/
 COPY json-stream-parser/Cargo.toml ./json-stream-parser/
@@ -30,8 +30,8 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
     libgomp1 libstdc++6 curl \
   && rm -rf /var/lib/apt/lists/*
 
-# ONNX Runtime (CPU). Change if you bump ort.
-ARG ORT_VERSION=1.22.0
+# ONNX Runtime (CPU). Keep in sync with ort crate requirements.
+ARG ORT_VERSION=1.23.2
 RUN mkdir -p /opt/onnxruntime && \
     curl -fsSL -o /tmp/ort.tgz \
       "https://github.com/microsoft/onnxruntime/releases/download/v${ORT_VERSION}/onnxruntime-linux-x64-${ORT_VERSION}.tgz" && \

README.md

@@ -10,13 +10,13 @@
 
 ## Demo deployment
 
-To test _Minne_ out, enter [this](https://minne-demo.stark.pub) read-only demo deployment to view and test functionality out.
+To test _Minne_ out, enter [this](https://minne.stark.pub) and sign in to a read-only demo deployment to view and test functionality out.
 
 ## Noteworthy Features
 
 - **Search & Chat Interface** - Find content or knowledge instantly with full-text search, or use the chat mode and conversational AI to find and reason about content
 - **Manual and AI-assisted connections** - Build entities and relationships manually with full control, let AI create entities and relationships automatically, or blend both approaches with AI suggestions for manual approval 
-- **Hybrid Retrieval System** - Search combining vector similarity, full-text search, and graph traversal for highly relevant results
+- **Hybrid Retrieval System** - Search combining vector similarity & full-text search
 - **Scratchpad Feature** - Quickly capture thoughts and convert them to permanent content when ready
 - **Visual Graph Explorer** - Interactive D3-based navigation of your knowledge entities and connections
 - **Multi-Format Support** - Ingest text, URLs, PDFs, audio files, and images into your knowledge base

api-router/src/error.rs

@@ -20,6 +20,9 @@ pub enum ApiError {
 
     #[error("Unauthorized: {0}")]
     Unauthorized(String),
+
+    #[error("Payload too large: {0}")]
+    PayloadTooLarge(String),
 }
 
 impl From<AppError> for ApiError {
@@ -67,6 +70,13 @@ impl IntoResponse for ApiError {
                     status: "error".to_string(),
                 },
             ),
+            Self::PayloadTooLarge(message) => (
+                StatusCode::PAYLOAD_TOO_LARGE,
+                ErrorResponse {
+                    error: message,
+                    status: "error".to_string(),
+                },
+            ),
         };
 
         (status, Json(error_response)).into_response()
@@ -132,6 +142,10 @@ mod tests {
         // Test unauthorized status
         let error = ApiError::Unauthorized("not allowed".to_string());
         assert_status_code(error, StatusCode::UNAUTHORIZED);
+
+        // Test payload too large status
+        let error = ApiError::PayloadTooLarge("too big".to_string());
+        assert_status_code(error, StatusCode::PAYLOAD_TOO_LARGE);
     }
 
     // Alternative approach that doesn't try to parse the response body

api-router/src/lib.rs

@@ -6,7 +6,7 @@ use axum::{
     Router,
 };
 use middleware_api_auth::api_auth;
-use routes::{categories::get_categories, ingress::ingest_data, liveness::live, readiness::ready};
+use routes::{categories::get_categories, ingest::ingest_data, liveness::live, readiness::ready};
 
 pub mod api_state;
 pub mod error;
@@ -26,9 +26,13 @@ where
 
     // Protected API endpoints (require auth)
     let protected = Router::new()
-        .route("/ingress", post(ingest_data))
+        .route(
+            "/ingest",
+            post(ingest_data).layer(DefaultBodyLimit::max(
+                app_state.config.ingest_max_body_bytes,
+            )),
+        )
         .route("/categories", get(get_categories))
-        .layer(DefaultBodyLimit::max(1024 * 1024 * 1024))
         .route_layer(from_fn_with_state(app_state.clone(), api_auth));
 
     public.merge(protected)

api-router/src/routes/ingest.rs

@@ -6,6 +6,7 @@ use common::{
         file_info::FileInfo, ingestion_payload::IngestionPayload, ingestion_task::IngestionTask,
         user::User,
     },
+    utils::ingest_limits::{validate_ingest_input, IngestValidationError},
 };
 use futures::{future::try_join_all, TryFutureExt};
 use serde_json::json;
@@ -19,7 +20,7 @@ pub struct IngestParams {
     pub content: Option<String>,
     pub context: String,
     pub category: String,
-    #[form_data(limit = "10000000")] // Adjust limit as needed
+    #[form_data(limit = "20000000")]
     #[form_data(default)]
     pub files: Vec<FieldData<NamedTempFile>>,
 }
@@ -29,10 +30,41 @@ pub async fn ingest_data(
     Extension(user): Extension<User>,
     TypedMultipart(input): TypedMultipart<IngestParams>,
 ) -> Result<impl IntoResponse, ApiError> {
-    info!("Received input: {:?}", input);
+    let user_id = user.id;
+    let content_bytes = input.content.as_ref().map_or(0, |c| c.len());
+    let has_content = input.content.as_ref().is_some_and(|c| !c.trim().is_empty());
+    let context_bytes = input.context.len();
+    let category_bytes = input.category.len();
+    let file_count = input.files.len();
+
+    match validate_ingest_input(
+        &state.config,
+        input.content.as_deref(),
+        &input.context,
+        &input.category,
+        file_count,
+    ) {
+        Ok(()) => {}
+        Err(IngestValidationError::PayloadTooLarge(message)) => {
+            return Err(ApiError::PayloadTooLarge(message));
+        }
+        Err(IngestValidationError::BadRequest(message)) => {
+            return Err(ApiError::ValidationError(message));
+        }
+    }
+
+    info!(
+        user_id = %user_id,
+        has_content,
+        content_bytes,
+        context_bytes,
+        category_bytes,
+        file_count,
+        "Received ingest request"
+    );
 
     let file_infos = try_join_all(input.files.into_iter().map(|file| {
-        FileInfo::new_with_storage(file, &state.db, &user.id, &state.storage)
+        FileInfo::new_with_storage(file, &state.db, &user_id, &state.storage)
             .map_err(AppError::from)
     }))
     .await?;
@@ -42,12 +74,12 @@ pub async fn ingest_data(
         input.context,
         input.category,
         file_infos,
-        user.id.as_str(),
+        &user_id,
     )?;
 
     let futures: Vec<_> = payloads
         .into_iter()
-        .map(|object| IngestionTask::create_and_add_to_db(object, user.id.clone(), &state.db))
+        .map(|object| IngestionTask::create_and_add_to_db(object, user_id.clone(), &state.db))
         .collect();
 
     try_join_all(futures).await?;

api-router/src/routes/mod.rs

@@ -1,4 +1,4 @@
 pub mod categories;
-pub mod ingress;
+pub mod ingest;
 pub mod liveness;
 pub mod readiness;

common/Cargo.toml

@@ -16,7 +16,7 @@ tracing = { workspace = true }
 anyhow = { workspace = true }
 thiserror = { workspace = true }
 serde_json = { workspace = true }
-surrealdb = { workspace = true, features = ["kv-mem"] }
+surrealdb = { workspace = true }
 async-openai = { workspace = true }
 futures = { workspace = true }
 tempfile = { workspace = true }
@@ -45,7 +45,11 @@ tokio-retry = { workspace = true }
 object_store = { workspace = true }
 bytes = { workspace = true }
 state-machines = { workspace = true }
+fastembed = { workspace = true }
 
 
 [features]
-test-utils = []
+test-utils = ["surrealdb/kv-mem"]
+
+[dev-dependencies]
+surrealdb = { workspace = true, features = ["kv-mem"] }

common/migrations/20250503_215025_initial_setup.surql

@@ -14,6 +14,9 @@ CREATE system_settings:current CONTENT {
     query_model: "gpt-4o-mini",
     processing_model: "gpt-4o-mini",
     embedding_model: "text-embedding-3-small",
+    voice_processing_model: "whisper-1",
+    image_processing_model: "gpt-4o-mini",
+    image_processing_prompt: "Analyze this image and respond based on its primary content:\n - If the image is mainly text (document, screenshot, sign), transcribe the text verbatim.\n - If the image is mainly visual (photograph, art, landscape), provide a concise description of the scene.\n - For hybrid images (diagrams, ads), briefly describe the visual, then transcribe the text under a Text: heading.\n\n Respond directly with the analysis.",
     embedding_dimensions: 1536,
     query_system_prompt: "You are a knowledgeable assistant with access to a specialized knowledge base. You will be provided with relevant knowledge entities from the database as context. Each knowledge entity contains a name, description, and type, representing different concepts, ideas, and information.\nYour task is to:\n1. Carefully analyze the provided knowledge entities in the context\n2. Answer user questions based on this information\n3. Provide clear, concise, and accurate responses\n4. When referencing information, briefly mention which knowledge entity it came from\n5. If the provided context doesn't contain enough information to answer the question confidently, clearly state this\n6. If only partial information is available, explain what you can answer and what information is missing\n7. Avoid making assumptions or providing information not supported by the context\n8. Output the references to the documents. Use the UUIDs and make sure they are correct!\nRemember:\n- Be direct and honest about the limitations of your knowledge\n- Cite the relevant knowledge entities when providing information, but only provide the UUIDs in the reference array\n- If you need to combine information from multiple entities, explain how they connect\n- Don't speculate beyond what's provided in the context\nExample response formats:\n\"Based on [Entity Name], [answer...]\"\n\"I found relevant information in multiple entries: [explanation...]\"\n\"I apologize, but the provided context doesn't contain information about [topic]\"",
     ingestion_system_prompt: "You are an AI assistant. You will receive a text content, along with user context and a category. Your task is to provide a structured JSON object representing the content in a graph format suitable for a graph database. You will also be presented with some existing knowledge_entities from the database, do not replicate these! Your task is to create meaningful knowledge entities from the submitted content. Try and infer as much as possible from the users context and category when creating these. If the user submits a large content, create more general entities. If the user submits a narrow and precise content, try and create precise knowledge entities.\nThe JSON should have the following structure:\n{\n\"knowledge_entities\": [\n{\n\"key\": \"unique-key-1\",\n\"name\": \"Entity Name\",\n\"description\": \"A detailed description of the entity.\",\n\"entity_type\": \"TypeOfEntity\"\n},\n// More entities...\n],\n\"relationships\": [\n{\n\"type\": \"RelationshipType\",\n\"source\": \"unique-key-1 or UUID from existing database\",\n\"target\": \"unique-key-1 or UUID from existing database\"\n},\n// More relationships...\n]\n}\nGuidelines:\n1. Do NOT generate any IDs or UUIDs. Use a unique `key` for each knowledge entity.\n2. Each KnowledgeEntity should have a unique `key`, a meaningful `name`, and a descriptive `description`.\n3. Define the type of each KnowledgeEntity using the following categories: Idea, Project, Document, Page, TextSnippet.\n4. Establish relationships between entities using types like RelatedTo, RelevantTo, SimilarTo.\n5. Use the `source` key to indicate the originating entity and the `target` key to indicate the related entity\"\n6. You will be presented with a few existing KnowledgeEntities that are similar to the current ones. They will have an existing UUID. When creating relationships to these entities, use their UUID.\n7. Only create relationships between existing KnowledgeEntities.\n8. Entities that exist already in the database should NOT be created again. If there is only a minor overlap, skip creating a new entity.\n9. A new relationship MUST include a newly created KnowledgeEntity."

common/migrations/20250514_142342_add_full_text_search_text_content.surql

@@ -1,27 +1,2 @@
-DEFINE ANALYZER IF NOT EXISTS app_default_fts_analyzer
-    TOKENIZERS class
-    FILTERS lowercase, ascii; 
-
-DEFINE INDEX IF NOT EXISTS text_content_fts_text_idx ON TABLE text_content
-    FIELDS text
-    SEARCH ANALYZER app_default_fts_analyzer BM25 HIGHLIGHTS;
-
-DEFINE INDEX IF NOT EXISTS text_content_fts_category_idx ON TABLE text_content
-    FIELDS category
-    SEARCH ANALYZER app_default_fts_analyzer BM25 HIGHLIGHTS;
-
-DEFINE INDEX IF NOT EXISTS text_content_fts_context_idx ON TABLE text_content
-    FIELDS context
-    SEARCH ANALYZER app_default_fts_analyzer BM25 HIGHLIGHTS;
-
-DEFINE INDEX IF NOT EXISTS text_content_fts_file_name_idx ON TABLE text_content
-    FIELDS file_info.file_name
-    SEARCH ANALYZER app_default_fts_analyzer BM25 HIGHLIGHTS;
-
-DEFINE INDEX IF NOT EXISTS text_content_fts_url_idx ON TABLE text_content
-    FIELDS url_info.url
-    SEARCH ANALYZER app_default_fts_analyzer BM25 HIGHLIGHTS;
-
-DEFINE INDEX IF NOT EXISTS text_content_fts_url_title_idx ON TABLE text_content
-    FIELDS url_info.title
-    SEARCH ANALYZER app_default_fts_analyzer BM25 HIGHLIGHTS;
+-- Runtime-managed: text_content FTS indexes now created at startup via the shared Surreal helper.
+-- This migration is intentionally left as a no-op to avoid heavy index builds during migration.

common/migrations/20250627_231035_remove_job_table.surql

@@ -1 +1 @@
-REMOVE TABLE job;
+-- No-op: legacy `job` table was superseded by `ingestion_task`; kept for migration order compatibility.

common/migrations/20251003_000001_add_fts_for_entities_and_chunks.surql

@@ -1,17 +1 @@
--- Add FTS indexes for searching name and description on entities
-
-DEFINE ANALYZER IF NOT EXISTS app_en_fts_analyzer
-    TOKENIZERS class
-    FILTERS lowercase, ascii, snowball(english);
-
-DEFINE INDEX IF NOT EXISTS knowledge_entity_fts_name_idx ON TABLE knowledge_entity
-    FIELDS name
-    SEARCH ANALYZER app_en_fts_analyzer BM25;
-
-DEFINE INDEX IF NOT EXISTS knowledge_entity_fts_description_idx ON TABLE knowledge_entity
-    FIELDS description
-    SEARCH ANALYZER app_en_fts_analyzer BM25;
-
-DEFINE INDEX IF NOT EXISTS text_chunk_fts_chunk_idx ON TABLE text_chunk
-    FIELDS chunk
-    SEARCH ANALYZER app_en_fts_analyzer BM25;
+-- Runtime-managed: FTS indexes now built at startup; migration retained as a no-op.

common/migrations/20251120_000001_remove_legacy_hnsw_indexes.surql

@@ -0,0 +1,18 @@
+-- Remove HNSW indexes from base tables (now created at runtime on *_embedding tables)
+REMOVE INDEX IF EXISTS idx_embedding_entities ON knowledge_entity;
+REMOVE INDEX IF EXISTS idx_embedding_chunks ON text_chunk;
+
+-- Remove FTS indexes (now created at runtime via indexes.rs)
+REMOVE INDEX IF EXISTS text_content_fts_text_idx ON text_content;
+REMOVE INDEX IF EXISTS text_content_fts_category_idx ON text_content;
+REMOVE INDEX IF EXISTS text_content_fts_context_idx ON text_content;
+REMOVE INDEX IF EXISTS text_content_fts_file_name_idx ON text_content;
+REMOVE INDEX IF EXISTS text_content_fts_url_idx ON text_content;
+REMOVE INDEX IF EXISTS text_content_fts_url_title_idx ON text_content;
+REMOVE INDEX IF EXISTS knowledge_entity_fts_name_idx ON knowledge_entity;
+REMOVE INDEX IF EXISTS knowledge_entity_fts_description_idx ON knowledge_entity;
+REMOVE INDEX IF EXISTS text_chunk_fts_chunk_idx ON text_chunk;
+
+-- Remove legacy analyzers (recreated at runtime with updated configuration)
+REMOVE ANALYZER IF EXISTS app_default_fts_analyzer;
+REMOVE ANALYZER IF EXISTS app_en_fts_analyzer;

common/migrations/20251121_113121_separate_embeddings_to_own_table.surql

@@ -0,0 +1,23 @@
+-- Move chunk/entity embeddings to dedicated tables for index efficiency.
+
+-- Text chunk embeddings table
+DEFINE TABLE IF NOT EXISTS text_chunk_embedding SCHEMAFULL;
+DEFINE FIELD IF NOT EXISTS created_at ON text_chunk_embedding TYPE datetime;
+DEFINE FIELD IF NOT EXISTS updated_at ON text_chunk_embedding TYPE datetime;
+DEFINE FIELD IF NOT EXISTS user_id ON text_chunk_embedding TYPE string;
+DEFINE FIELD IF NOT EXISTS source_id ON text_chunk_embedding TYPE string;
+DEFINE FIELD IF NOT EXISTS chunk_id ON text_chunk_embedding TYPE record<text_chunk>;
+DEFINE FIELD IF NOT EXISTS embedding ON text_chunk_embedding TYPE array<float>;
+DEFINE INDEX IF NOT EXISTS text_chunk_embedding_chunk_id_idx ON text_chunk_embedding FIELDS chunk_id;
+DEFINE INDEX IF NOT EXISTS text_chunk_embedding_user_id_idx ON text_chunk_embedding FIELDS user_id;
+DEFINE INDEX IF NOT EXISTS text_chunk_embedding_source_id_idx ON text_chunk_embedding FIELDS source_id;
+
+-- Knowledge entity embeddings table
+DEFINE TABLE IF NOT EXISTS knowledge_entity_embedding SCHEMAFULL;
+DEFINE FIELD IF NOT EXISTS created_at ON knowledge_entity_embedding TYPE datetime;
+DEFINE FIELD IF NOT EXISTS updated_at ON knowledge_entity_embedding TYPE datetime;
+DEFINE FIELD IF NOT EXISTS user_id ON knowledge_entity_embedding TYPE string;
+DEFINE FIELD IF NOT EXISTS entity_id ON knowledge_entity_embedding TYPE record<knowledge_entity>;
+DEFINE FIELD IF NOT EXISTS embedding ON knowledge_entity_embedding TYPE array<float>;
+DEFINE INDEX IF NOT EXISTS knowledge_entity_embedding_entity_id_idx ON knowledge_entity_embedding FIELDS entity_id;
+DEFINE INDEX IF NOT EXISTS knowledge_entity_embedding_user_id_idx ON knowledge_entity_embedding FIELDS user_id;

common/migrations/20251121_113122_migrate_embedding_data.surql

@@ -0,0 +1,23 @@
+-- Copy embeddings from base tables to dedicated tables
+-- This runs BEFORE the field removal migration
+
+FOR $chunk IN (SELECT * FROM text_chunk WHERE embedding != NONE AND array::len(embedding) > 0) {
+    CREATE text_chunk_embedding CONTENT {
+        chunk_id: $chunk.id,
+        embedding: $chunk.embedding,
+        user_id: $chunk.user_id,
+        source_id: $chunk.source_id,
+        created_at: $chunk.created_at,
+        updated_at: $chunk.updated_at
+    };
+};
+
+FOR $entity IN (SELECT * FROM knowledge_entity WHERE embedding != NONE AND array::len(embedding) > 0) {
+    CREATE knowledge_entity_embedding CONTENT {
+        entity_id: $entity.id,
+        embedding: $entity.embedding,
+        user_id: $entity.user_id,
+        created_at: $entity.created_at,
+        updated_at: $entity.updated_at
+    };
+};

common/migrations/20251122_151002_remove_legacy_embedding_fields.surql

@@ -0,0 +1,3 @@
+-- Drop legacy embedding fields from base tables; embeddings now live in *_embedding tables.
+REMOVE FIELD IF EXISTS embedding ON TABLE text_chunk;
+REMOVE FIELD IF EXISTS embedding ON TABLE knowledge_entity;

common/migrations/20251210_add_embedding_backend_to_system_settings.surql

@@ -0,0 +1,8 @@
+-- Add embedding_backend field to system_settings for visibility of active backend
+
+DEFINE FIELD IF NOT EXISTS embedding_backend ON system_settings TYPE option<string>;
+
+-- Set default to 'openai' for existing installs to preserve backward compatibility
+UPDATE system_settings:current SET
+    embedding_backend = 'openai'
+WHERE embedding_backend == NONE;

common/migrations/20251231_enforce_schemafull.surql

@@ -0,0 +1,97 @@
+-- Enforce SCHEMAFULL on all tables and define missing fields
+
+-- 1. Define missing fields for ingestion_task (formerly job, but now ingestion_task)
+DEFINE TABLE OVERWRITE ingestion_task SCHEMAFULL;
+
+-- Core Fields
+DEFINE FIELD IF NOT EXISTS id ON ingestion_task TYPE record<ingestion_task>;
+DEFINE FIELD IF NOT EXISTS created_at ON ingestion_task TYPE datetime DEFAULT time::now();
+DEFINE FIELD IF NOT EXISTS updated_at ON ingestion_task TYPE datetime DEFAULT time::now();
+DEFINE FIELD IF NOT EXISTS user_id ON ingestion_task TYPE string;
+
+-- State Machine Fields
+DEFINE FIELD IF NOT EXISTS state ON ingestion_task TYPE string ASSERT $value IN ['Pending', 'Reserved', 'Processing', 'Succeeded', 'Failed', 'Cancelled', 'DeadLetter'];
+DEFINE FIELD IF NOT EXISTS attempts ON ingestion_task TYPE int DEFAULT 0;
+DEFINE FIELD IF NOT EXISTS max_attempts ON ingestion_task TYPE int DEFAULT 3;
+DEFINE FIELD IF NOT EXISTS scheduled_at ON ingestion_task TYPE datetime DEFAULT time::now();
+DEFINE FIELD IF NOT EXISTS locked_at ON ingestion_task TYPE option<datetime>;
+DEFINE FIELD IF NOT EXISTS lease_duration_secs ON ingestion_task TYPE int DEFAULT 300;
+DEFINE FIELD IF NOT EXISTS worker_id ON ingestion_task TYPE option<string>;
+DEFINE FIELD IF NOT EXISTS error_code ON ingestion_task TYPE option<string>;
+DEFINE FIELD IF NOT EXISTS error_message ON ingestion_task TYPE option<string>;
+DEFINE FIELD IF NOT EXISTS last_error_at ON ingestion_task TYPE option<datetime>;
+DEFINE FIELD IF NOT EXISTS priority ON ingestion_task TYPE int DEFAULT 0;
+
+-- Content Payload (IngestionPayload Enum)
+DEFINE FIELD IF NOT EXISTS content ON ingestion_task TYPE object;
+DEFINE FIELD IF NOT EXISTS content.Url ON ingestion_task TYPE option<object>;
+DEFINE FIELD IF NOT EXISTS content.Text ON ingestion_task TYPE option<object>;
+DEFINE FIELD IF NOT EXISTS content.File ON ingestion_task TYPE option<object>;
+
+-- Content: Url Variant
+DEFINE FIELD IF NOT EXISTS content.Url.url ON ingestion_task TYPE string;
+DEFINE FIELD IF NOT EXISTS content.Url.context ON ingestion_task TYPE string;
+DEFINE FIELD IF NOT EXISTS content.Url.category ON ingestion_task TYPE string;
+DEFINE FIELD IF NOT EXISTS content.Url.user_id ON ingestion_task TYPE string;
+
+-- Content: Text Variant
+DEFINE FIELD IF NOT EXISTS content.Text.text ON ingestion_task TYPE string;
+DEFINE FIELD IF NOT EXISTS content.Text.context ON ingestion_task TYPE string;
+DEFINE FIELD IF NOT EXISTS content.Text.category ON ingestion_task TYPE string;
+DEFINE FIELD IF NOT EXISTS content.Text.user_id ON ingestion_task TYPE string;
+
+-- Content: File Variant
+DEFINE FIELD IF NOT EXISTS content.File.context ON ingestion_task TYPE string;
+DEFINE FIELD IF NOT EXISTS content.File.category ON ingestion_task TYPE string;
+DEFINE FIELD IF NOT EXISTS content.File.user_id ON ingestion_task TYPE string;
+DEFINE FIELD IF NOT EXISTS content.File.file_info ON ingestion_task TYPE object;
+
+-- Content: File.file_info (FileInfo Struct)
+DEFINE FIELD IF NOT EXISTS content.File.file_info.id ON ingestion_task TYPE string;
+DEFINE FIELD IF NOT EXISTS content.File.file_info.created_at ON ingestion_task TYPE datetime;
+DEFINE FIELD IF NOT EXISTS content.File.file_info.updated_at ON ingestion_task TYPE datetime;
+DEFINE FIELD IF NOT EXISTS content.File.file_info.sha256 ON ingestion_task TYPE string;
+DEFINE FIELD IF NOT EXISTS content.File.file_info.path ON ingestion_task TYPE string;
+DEFINE FIELD IF NOT EXISTS content.File.file_info.file_name ON ingestion_task TYPE string;
+DEFINE FIELD IF NOT EXISTS content.File.file_info.mime_type ON ingestion_task TYPE string;
+DEFINE FIELD IF NOT EXISTS content.File.file_info.user_id ON ingestion_task TYPE string;
+
+-- 2. Enforce SCHEMAFULL on all other tables
+DEFINE TABLE OVERWRITE analytics SCHEMAFULL;
+DEFINE TABLE OVERWRITE conversation SCHEMAFULL;
+DEFINE TABLE OVERWRITE file SCHEMAFULL;
+DEFINE TABLE OVERWRITE knowledge_entity SCHEMAFULL;
+DEFINE TABLE OVERWRITE message SCHEMAFULL;
+DEFINE TABLE OVERWRITE relates_to SCHEMAFULL TYPE RELATION;
+DEFINE FIELD IF NOT EXISTS in ON relates_to TYPE record<knowledge_entity>;
+DEFINE FIELD IF NOT EXISTS out ON relates_to TYPE record<knowledge_entity>;
+DEFINE FIELD IF NOT EXISTS metadata ON relates_to TYPE object;
+DEFINE FIELD IF NOT EXISTS metadata.user_id ON relates_to TYPE string;
+DEFINE FIELD IF NOT EXISTS metadata.source_id ON relates_to TYPE string;
+DEFINE FIELD IF NOT EXISTS metadata.relationship_type ON relates_to TYPE string;
+DEFINE TABLE OVERWRITE scratchpad SCHEMAFULL;
+DEFINE TABLE OVERWRITE system_settings SCHEMAFULL;
+DEFINE TABLE OVERWRITE text_chunk SCHEMAFULL;
+-- text_content must have fields defined before enforcing SCHEMAFULL
+DEFINE TABLE OVERWRITE text_content SCHEMAFULL;
+DEFINE FIELD IF NOT EXISTS created_at ON text_content TYPE datetime;
+DEFINE FIELD IF NOT EXISTS updated_at ON text_content TYPE datetime;
+DEFINE FIELD IF NOT EXISTS text ON text_content TYPE string;
+DEFINE FIELD IF NOT EXISTS file_info ON text_content TYPE option<object>;
+DEFINE FIELD IF NOT EXISTS url_info ON text_content TYPE option<object>;
+DEFINE FIELD IF NOT EXISTS url_info.url ON text_content TYPE string;
+DEFINE FIELD IF NOT EXISTS url_info.title ON text_content TYPE string;
+DEFINE FIELD IF NOT EXISTS url_info.image_id ON text_content TYPE string;
+DEFINE FIELD IF NOT EXISTS context ON text_content TYPE option<string>;
+DEFINE FIELD IF NOT EXISTS category ON text_content TYPE string;
+DEFINE FIELD IF NOT EXISTS user_id ON text_content TYPE string;
+DEFINE FIELD IF NOT EXISTS file_info.id ON text_content TYPE string;
+DEFINE FIELD IF NOT EXISTS file_info.created_at ON text_content TYPE datetime;
+DEFINE FIELD IF NOT EXISTS file_info.updated_at ON text_content TYPE datetime;
+DEFINE FIELD IF NOT EXISTS file_info.sha256 ON text_content TYPE string;
+DEFINE FIELD IF NOT EXISTS file_info.path ON text_content TYPE string;
+DEFINE FIELD IF NOT EXISTS file_info.file_name ON text_content TYPE string;
+DEFINE FIELD IF NOT EXISTS file_info.mime_type ON text_content TYPE string;
+DEFINE FIELD IF NOT EXISTS file_info.user_id ON text_content TYPE string;
+
+DEFINE TABLE OVERWRITE user SCHEMAFULL;

common/migrations/20260116_000000_add_user_theme_preference.surql

@@ -0,0 +1 @@
+DEFINE FIELD IF NOT EXISTS theme ON user TYPE string DEFAULT "system";

common/migrations/definitions/20250514_142342_add_full_text_search_text_content.json

@@ -1 +0,0 @@
-{"schemas":"--- original\n+++ modified\n@@ -98,7 +98,7 @@\n DEFINE INDEX IF NOT EXISTS knowledge_entity_user_id_idx ON knowledge_entity FIELDS user_id;\n DEFINE INDEX IF NOT EXISTS knowledge_entity_source_id_idx ON knowledge_entity FIELDS source_id;\n DEFINE INDEX IF NOT EXISTS knowledge_entity_entity_type_idx ON knowledge_entity FIELDS entity_type;\n-DEFINE INDEX IF NOT EXISTS knowledge_entity_created_at_idx ON knowledge_entity FIELDS created_at; # For get_latest_knowledge_entities\n+DEFINE INDEX IF NOT EXISTS knowledge_entity_created_at_idx ON knowledge_entity FIELDS created_at;\n\n # Defines the schema for the 'message' table.\n\n@@ -157,6 +157,8 @@\n DEFINE FIELD IF NOT EXISTS require_email_verification ON system_settings TYPE bool;\n DEFINE FIELD IF NOT EXISTS query_model ON system_settings TYPE string;\n DEFINE FIELD IF NOT EXISTS processing_model ON system_settings TYPE string;\n+DEFINE FIELD IF NOT EXISTS embedding_model ON system_settings TYPE string;\n+DEFINE FIELD IF NOT EXISTS embedding_dimensions ON system_settings TYPE int;\n DEFINE FIELD IF NOT EXISTS query_system_prompt ON system_settings TYPE string;\n DEFINE FIELD IF NOT EXISTS ingestion_system_prompt ON system_settings TYPE string;\n\n","events":null}

common/migrations/definitions/20250606_234535_add_embedding_model_and_dimensions_to_system_settings.json

@@ -1 +0,0 @@
-{"schemas":"--- original\n+++ modified\n@@ -51,23 +51,23 @@\n\n # Defines the schema for the 'ingestion_task' table (used by IngestionTask).\n\n-DEFINE TABLE IF NOT EXISTS job SCHEMALESS;\n+DEFINE TABLE IF NOT EXISTS ingestion_task SCHEMALESS;\n\n # Standard fields\n-DEFINE FIELD IF NOT EXISTS created_at ON job TYPE string;\n-DEFINE FIELD IF NOT EXISTS updated_at ON job TYPE string;\n+DEFINE FIELD IF NOT EXISTS created_at ON ingestion_task TYPE string;\n+DEFINE FIELD IF NOT EXISTS updated_at ON ingestion_task TYPE string;\n\n # Custom fields from the IngestionTask struct\n # IngestionPayload is complex, store as object\n-DEFINE FIELD IF NOT EXISTS content ON job TYPE object;\n+DEFINE FIELD IF NOT EXISTS content ON ingestion_task TYPE object;\n # IngestionTaskStatus can hold data (InProgress), store as object\n-DEFINE FIELD IF NOT EXISTS status ON job TYPE object;\n-DEFINE FIELD IF NOT EXISTS user_id ON job TYPE string;\n+DEFINE FIELD IF NOT EXISTS status ON ingestion_task TYPE object;\n+DEFINE FIELD IF NOT EXISTS user_id ON ingestion_task TYPE string;\n\n # Indexes explicitly defined in build_indexes and useful for get_unfinished_tasks\n-DEFINE INDEX IF NOT EXISTS idx_job_status ON job FIELDS status;\n-DEFINE INDEX IF NOT EXISTS idx_job_user ON job FIELDS user_id;\n-DEFINE INDEX IF NOT EXISTS idx_job_created ON job FIELDS created_at;\n+DEFINE INDEX IF NOT EXISTS idx_ingestion_task_status ON ingestion_task FIELDS status;\n+DEFINE INDEX IF NOT EXISTS idx_ingestion_task_user ON ingestion_task FIELDS user_id;\n+DEFINE INDEX IF NOT EXISTS idx_ingestion_task_created ON ingestion_task FIELDS created_at;\n\n # Defines the schema for the 'knowledge_entity' table.\n\n","events":null}

common/migrations/definitions/20250627_120000_add_image_processing_settings.json

@@ -1 +0,0 @@
-{"schemas":"--- original\n+++ modified\n@@ -57,10 +57,7 @@\n DEFINE FIELD IF NOT EXISTS created_at ON ingestion_task TYPE string;\n DEFINE FIELD IF NOT EXISTS updated_at ON ingestion_task TYPE string;\n\n-# Custom fields from the IngestionTask struct\n-# IngestionPayload is complex, store as object\n DEFINE FIELD IF NOT EXISTS content ON ingestion_task TYPE object;\n-# IngestionTaskStatus can hold data (InProgress), store as object\n DEFINE FIELD IF NOT EXISTS status ON ingestion_task TYPE object;\n DEFINE FIELD IF NOT EXISTS user_id ON ingestion_task TYPE string;\n\n@@ -157,10 +154,12 @@\n DEFINE FIELD IF NOT EXISTS require_email_verification ON system_settings TYPE bool;\n DEFINE FIELD IF NOT EXISTS query_model ON system_settings TYPE string;\n DEFINE FIELD IF NOT EXISTS processing_model ON system_settings TYPE string;\n+DEFINE FIELD IF NOT EXISTS image_processing_model ON system_settings TYPE string;\n DEFINE FIELD IF NOT EXISTS embedding_model ON system_settings TYPE string;\n DEFINE FIELD IF NOT EXISTS embedding_dimensions ON system_settings TYPE int;\n DEFINE FIELD IF NOT EXISTS query_system_prompt ON system_settings TYPE string;\n DEFINE FIELD IF NOT EXISTS ingestion_system_prompt ON system_settings TYPE string;\n+DEFINE FIELD IF NOT EXISTS image_processing_prompt ON system_settings TYPE string;\n\n # Defines the schema for the 'text_chunk' table.\n\n","events":null}

common/migrations/definitions/20250701_000000_add_voice_processing_model_to_system_settings.json

@@ -1 +0,0 @@
-{"schemas":"--- original\n+++ modified\n@@ -160,6 +160,7 @@\n DEFINE FIELD IF NOT EXISTS query_system_prompt ON system_settings TYPE string;\n DEFINE FIELD IF NOT EXISTS ingestion_system_prompt ON system_settings TYPE string;\n DEFINE FIELD IF NOT EXISTS image_processing_prompt ON system_settings TYPE string;\n+DEFINE FIELD IF NOT EXISTS voice_processing_model ON system_settings TYPE string;\n\n # Defines the schema for the 'text_chunk' table.\n\n","events":null}

common/migrations/definitions/20250921_120004_fix_datetime_fields.json

@@ -1 +0,0 @@
-{"schemas":"--- original\n+++ modified\n@@ -18,8 +18,8 @@\n DEFINE TABLE IF NOT EXISTS conversation SCHEMALESS;\n\n # Standard fields\n-DEFINE FIELD IF NOT EXISTS created_at ON conversation TYPE string;\n-DEFINE FIELD IF NOT EXISTS updated_at ON conversation TYPE string;\n+DEFINE FIELD IF NOT EXISTS created_at ON conversation TYPE datetime;\n+DEFINE FIELD IF NOT EXISTS updated_at ON conversation TYPE datetime;\n\n # Custom fields from the Conversation struct\n DEFINE FIELD IF NOT EXISTS user_id ON conversation TYPE string;\n@@ -34,8 +34,8 @@\n DEFINE TABLE IF NOT EXISTS file SCHEMALESS;\n\n # Standard fields\n-DEFINE FIELD IF NOT EXISTS created_at ON file TYPE string;\n-DEFINE FIELD IF NOT EXISTS updated_at ON file TYPE string;\n+DEFINE FIELD IF NOT EXISTS created_at ON file TYPE datetime;\n+DEFINE FIELD IF NOT EXISTS updated_at ON file TYPE datetime;\n\n # Custom fields from the FileInfo struct\n DEFINE FIELD IF NOT EXISTS sha256 ON file TYPE string;\n@@ -54,8 +54,8 @@\n DEFINE TABLE IF NOT EXISTS ingestion_task SCHEMALESS;\n\n # Standard fields\n-DEFINE FIELD IF NOT EXISTS created_at ON ingestion_task TYPE string;\n-DEFINE FIELD IF NOT EXISTS updated_at ON ingestion_task TYPE string;\n+DEFINE FIELD IF NOT EXISTS created_at ON ingestion_task TYPE datetime;\n+DEFINE FIELD IF NOT EXISTS updated_at ON ingestion_task TYPE datetime;\n\n DEFINE FIELD IF NOT EXISTS content ON ingestion_task TYPE object;\n DEFINE FIELD IF NOT EXISTS status ON ingestion_task TYPE object;\n@@ -71,8 +71,8 @@\n DEFINE TABLE IF NOT EXISTS knowledge_entity SCHEMALESS;\n\n # Standard fields\n-DEFINE FIELD IF NOT EXISTS created_at ON knowledge_entity TYPE string;\n-DEFINE FIELD IF NOT EXISTS updated_at ON knowledge_entity TYPE string;\n+DEFINE FIELD IF NOT EXISTS created_at ON knowledge_entity TYPE datetime;\n+DEFINE FIELD IF NOT EXISTS updated_at ON knowledge_entity TYPE datetime;\n\n # Custom fields from the KnowledgeEntity struct\n DEFINE FIELD IF NOT EXISTS source_id ON knowledge_entity TYPE string;\n@@ -102,8 +102,8 @@\n DEFINE TABLE IF NOT EXISTS message SCHEMALESS;\n\n # Standard fields\n-DEFINE FIELD IF NOT EXISTS created_at ON message TYPE string;\n-DEFINE FIELD IF NOT EXISTS updated_at ON message TYPE string;\n+DEFINE FIELD IF NOT EXISTS created_at ON message TYPE datetime;\n+DEFINE FIELD IF NOT EXISTS updated_at ON message TYPE datetime;\n\n # Custom fields from the Message struct\n DEFINE FIELD IF NOT EXISTS conversation_id ON message TYPE string;\n@@ -167,8 +167,8 @@\n DEFINE TABLE IF NOT EXISTS text_chunk SCHEMALESS;\n\n # Standard fields\n-DEFINE FIELD IF NOT EXISTS created_at ON text_chunk TYPE string;\n-DEFINE FIELD IF NOT EXISTS updated_at ON text_chunk TYPE string;\n+DEFINE FIELD IF NOT EXISTS created_at ON text_chunk TYPE datetime;\n+DEFINE FIELD IF NOT EXISTS updated_at ON text_chunk TYPE datetime;\n\n # Custom fields from the TextChunk struct\n DEFINE FIELD IF NOT EXISTS source_id ON text_chunk TYPE string;\n@@ -191,8 +191,8 @@\n DEFINE TABLE IF NOT EXISTS text_content SCHEMALESS;\n\n # Standard fields\n-DEFINE FIELD IF NOT EXISTS created_at ON text_content TYPE string;\n-DEFINE FIELD IF NOT EXISTS updated_at ON text_content TYPE string;\n+DEFINE FIELD IF NOT EXISTS created_at ON text_content TYPE datetime;\n+DEFINE FIELD IF NOT EXISTS updated_at ON text_content TYPE datetime;\n\n # Custom fields from the TextContent struct\n DEFINE FIELD IF NOT EXISTS text ON text_content TYPE string;\n@@ -215,8 +215,8 @@\n DEFINE TABLE IF NOT EXISTS user SCHEMALESS;\n\n # Standard fields\n-DEFINE FIELD IF NOT EXISTS created_at ON user TYPE string;\n-DEFINE FIELD IF NOT EXISTS updated_at ON user TYPE string;\n+DEFINE FIELD IF NOT EXISTS created_at ON user TYPE datetime;\n+DEFINE FIELD IF NOT EXISTS updated_at ON user TYPE datetime;\n\n # Custom fields from the User struct\n DEFINE FIELD IF NOT EXISTS email ON user TYPE string;\n","events":null}

common/migrations/definitions/20251022_120302_add_scratchpad_table.json

@@ -1 +0,0 @@
-{"schemas":"--- original\n+++ modified\n@@ -137,6 +137,30 @@\n DEFINE INDEX IF NOT EXISTS relates_to_metadata_source_id_idx ON relates_to FIELDS metadata.source_id;\n DEFINE INDEX IF NOT EXISTS relates_to_metadata_user_id_idx ON relates_to FIELDS metadata.user_id;\n\n+# Defines the schema for the 'scratchpad' table.\n+\n+DEFINE TABLE IF NOT EXISTS scratchpad SCHEMALESS;\n+\n+# Standard fields from stored_object! macro\n+DEFINE FIELD IF NOT EXISTS created_at ON scratchpad TYPE datetime;\n+DEFINE FIELD IF NOT EXISTS updated_at ON scratchpad TYPE datetime;\n+\n+# Custom fields from the Scratchpad struct\n+DEFINE FIELD IF NOT EXISTS user_id ON scratchpad TYPE string;\n+DEFINE FIELD IF NOT EXISTS title ON scratchpad TYPE string;\n+DEFINE FIELD IF NOT EXISTS content ON scratchpad TYPE string;\n+DEFINE FIELD IF NOT EXISTS last_saved_at ON scratchpad TYPE datetime;\n+DEFINE FIELD IF NOT EXISTS is_dirty ON scratchpad TYPE bool DEFAULT false;\n+DEFINE FIELD IF NOT EXISTS is_archived ON scratchpad TYPE bool DEFAULT false;\n+DEFINE FIELD IF NOT EXISTS archived_at ON scratchpad TYPE option<datetime>;\n+DEFINE FIELD IF NOT EXISTS ingested_at ON scratchpad TYPE option<datetime>;\n+\n+# Indexes based on query patterns\n+DEFINE INDEX IF NOT EXISTS scratchpad_user_idx ON scratchpad FIELDS user_id;\n+DEFINE INDEX IF NOT EXISTS scratchpad_user_archived_idx ON scratchpad FIELDS user_id, is_archived;\n+DEFINE INDEX IF NOT EXISTS scratchpad_updated_idx ON scratchpad FIELDS updated_at;\n+DEFINE INDEX IF NOT EXISTS scratchpad_archived_idx ON scratchpad FIELDS archived_at;\n+\n DEFINE TABLE OVERWRITE script_migration SCHEMAFULL\n     PERMISSIONS\n         FOR select FULL\n","events":null}

common/migrations/definitions/20251231_enforce_schemafull.json

@@ -0,0 +1 @@
+{"schemas":"--- original\n+++ modified\n@@ -242,7 +242,7 @@\n\n # Defines the schema for the 'text_content' table.\n\n-DEFINE TABLE IF NOT EXISTS text_content SCHEMALESS;\n+DEFINE TABLE IF NOT EXISTS text_content SCHEMAFULL;\n\n # Standard fields\n DEFINE FIELD IF NOT EXISTS created_at ON text_content TYPE datetime;\n@@ -254,10 +254,24 @@\n DEFINE FIELD IF NOT EXISTS file_info ON text_content TYPE option<object>;\n # UrlInfo is a struct, store as object\n DEFINE FIELD IF NOT EXISTS url_info ON text_content TYPE option<object>;\n+DEFINE FIELD IF NOT EXISTS url_info.url ON text_content TYPE string;\n+DEFINE FIELD IF NOT EXISTS url_info.title ON text_content TYPE string;\n+DEFINE FIELD IF NOT EXISTS url_info.image_id ON text_content TYPE string;\n+\n DEFINE FIELD IF NOT EXISTS context ON text_content TYPE option<string>;\n DEFINE FIELD IF NOT EXISTS category ON text_content TYPE string;\n DEFINE FIELD IF NOT EXISTS user_id ON text_content TYPE string;\n\n+# FileInfo fields\n+DEFINE FIELD IF NOT EXISTS file_info.id ON text_content TYPE string;\n+DEFINE FIELD IF NOT EXISTS file_info.created_at ON text_content TYPE datetime;\n+DEFINE FIELD IF NOT EXISTS file_info.updated_at ON text_content TYPE datetime;\n+DEFINE FIELD IF NOT EXISTS file_info.sha256 ON text_content TYPE string;\n+DEFINE FIELD IF NOT EXISTS file_info.path ON text_content TYPE string;\n+DEFINE FIELD IF NOT EXISTS file_info.file_name ON text_content TYPE string;\n+DEFINE FIELD IF NOT EXISTS file_info.mime_type ON text_content TYPE string;\n+DEFINE FIELD IF NOT EXISTS file_info.user_id ON text_content TYPE string;\n+\n # Indexes based on query patterns\n DEFINE INDEX IF NOT EXISTS text_content_user_id_idx ON text_content FIELDS user_id;\n DEFINE INDEX IF NOT EXISTS text_content_created_at_idx ON text_content FIELDS created_at;\n","events":null}

common/migrations/definitions/20260116_000000_add_user_theme_preference.json

@@ -0,0 +1 @@
+{"schemas":"--- original\n+++ modified\n@@ -28,6 +28,7 @@\n # Add indexes based on query patterns (get_complete_conversation ownership check, get_user_conversations)\n DEFINE INDEX IF NOT EXISTS conversation_user_id_idx ON conversation FIELDS user_id;\n DEFINE INDEX IF NOT EXISTS conversation_created_at_idx ON conversation FIELDS created_at; # For get_user_conversations ORDER BY\n+DEFINE INDEX IF NOT EXISTS conversation_user_updated_at_idx ON conversation FIELDS user_id, updated_at; # For sidebar conversation projection ORDER BY\n\n # Defines the schema for the 'file' table (used by FileInfo).\n\n","events":null}

common/schemas/conversation.surql

@@ -13,3 +13,4 @@ DEFINE FIELD IF NOT EXISTS title ON conversation TYPE string;
 # Add indexes based on query patterns (get_complete_conversation ownership check, get_user_conversations)
 DEFINE INDEX IF NOT EXISTS conversation_user_id_idx ON conversation FIELDS user_id;
 DEFINE INDEX IF NOT EXISTS conversation_created_at_idx ON conversation FIELDS created_at; # For get_user_conversations ORDER BY
+DEFINE INDEX IF NOT EXISTS conversation_user_updated_at_idx ON conversation FIELDS user_id, updated_at; # For sidebar conversation projection ORDER BY

common/schemas/knowledge_entity.surql

@@ -15,16 +15,12 @@ DEFINE FIELD IF NOT EXISTS entity_type ON knowledge_entity TYPE string;
 # metadata is Option<serde_json::Value>, store as object
 DEFINE FIELD IF NOT EXISTS metadata ON knowledge_entity TYPE option<object>;
 
-# Define embedding as a standard array of floats for schema definition
-DEFINE FIELD IF NOT EXISTS embedding ON knowledge_entity TYPE array<float>;
-# The specific vector nature is handled by the index definition below
-
 DEFINE FIELD IF NOT EXISTS user_id ON knowledge_entity TYPE string;
 
-# Indexes based on build_indexes and query patterns
-# The INDEX definition correctly specifies the vector properties
-DEFINE INDEX IF NOT EXISTS idx_embedding_entities ON knowledge_entity FIELDS embedding HNSW DIMENSION 1536;
-DEFINE INDEX IF NOT EXISTS knowledge_entity_user_id_idx ON knowledge_entity FIELDS user_id;
+-- Indexes based on build_indexes and query patterns
+-- HNSW index now defined on knowledge_entity_embedding table for better memory usage
+-- DEFINE INDEX IF NOT EXISTS idx_embedding_entities ON knowledge_entity FIELDS embedding HNSW DIMENSION 1536;
 DEFINE INDEX IF NOT EXISTS knowledge_entity_source_id_idx ON knowledge_entity FIELDS source_id;
+DEFINE INDEX IF NOT EXISTS knowledge_entity_user_id_idx ON knowledge_entity FIELDS user_id;
 DEFINE INDEX IF NOT EXISTS knowledge_entity_entity_type_idx ON knowledge_entity FIELDS entity_type;
 DEFINE INDEX IF NOT EXISTS knowledge_entity_created_at_idx ON knowledge_entity FIELDS created_at;

common/schemas/knowledge_entity_embedding.surql

@@ -0,0 +1,18 @@
+-- Defines the schema for the 'knowledge_entity_embedding' table.
+-- Separate table to optimize HNSW index creation memory usage
+
+DEFINE TABLE IF NOT EXISTS knowledge_entity_embedding SCHEMAFULL;
+
+-- Standard fields
+DEFINE FIELD IF NOT EXISTS created_at ON knowledge_entity_embedding TYPE datetime;
+DEFINE FIELD IF NOT EXISTS updated_at ON knowledge_entity_embedding TYPE datetime;
+DEFINE FIELD IF NOT EXISTS user_id ON knowledge_entity_embedding TYPE string;
+
+-- Custom fields
+DEFINE FIELD IF NOT EXISTS entity_id ON knowledge_entity_embedding TYPE record<knowledge_entity>;
+DEFINE FIELD IF NOT EXISTS embedding ON knowledge_entity_embedding TYPE array<float>;
+
+-- Indexes
+-- DEFINE INDEX IF NOT EXISTS idx_embedding_knowledge_entity_embedding ON knowledge_entity_embedding FIELDS embedding HNSW DIMENSION 1536;
+DEFINE INDEX IF NOT EXISTS knowledge_entity_embedding_entity_id_idx ON knowledge_entity_embedding FIELDS entity_id;
+DEFINE INDEX IF NOT EXISTS knowledge_entity_embedding_user_id_idx ON knowledge_entity_embedding FIELDS user_id;

common/schemas/text_chunk.surql

@@ -10,14 +10,8 @@ DEFINE FIELD IF NOT EXISTS updated_at ON text_chunk TYPE datetime;
 DEFINE FIELD IF NOT EXISTS source_id ON text_chunk TYPE string;
 DEFINE FIELD IF NOT EXISTS chunk ON text_chunk TYPE string;
 
-# Define embedding as a standard array of floats for schema definition
-DEFINE FIELD IF NOT EXISTS embedding ON text_chunk TYPE array<float>;
-# The specific vector nature is handled by the index definition below
-
 DEFINE FIELD IF NOT EXISTS user_id ON text_chunk TYPE string;
 
 # Indexes based on build_indexes and query patterns (delete_by_source_id)
-# The INDEX definition correctly specifies the vector properties
-DEFINE INDEX IF NOT EXISTS idx_embedding_chunks ON text_chunk FIELDS embedding HNSW DIMENSION 1536;
 DEFINE INDEX IF NOT EXISTS text_chunk_source_id_idx ON text_chunk FIELDS source_id;
 DEFINE INDEX IF NOT EXISTS text_chunk_user_id_idx ON text_chunk FIELDS user_id;

common/schemas/text_chunk_embedding.surql

@@ -0,0 +1,20 @@
+-- Defines the schema for the 'text_chunk_embedding' table.
+-- Separate table to optimize HNSW index creation memory usage
+
+DEFINE TABLE IF NOT EXISTS text_chunk_embedding SCHEMAFULL;
+
+# Standard fields
+DEFINE FIELD IF NOT EXISTS created_at ON text_chunk_embedding TYPE datetime;
+DEFINE FIELD IF NOT EXISTS updated_at ON text_chunk_embedding TYPE datetime;
+DEFINE FIELD IF NOT EXISTS user_id ON text_chunk_embedding TYPE string;
+DEFINE FIELD IF NOT EXISTS source_id ON text_chunk_embedding TYPE string;
+
+# Custom fields
+DEFINE FIELD IF NOT EXISTS chunk_id ON text_chunk_embedding TYPE record<text_chunk>;
+DEFINE FIELD IF NOT EXISTS embedding ON text_chunk_embedding TYPE array<float>;
+
+-- Indexes
+-- DEFINE INDEX IF NOT EXISTS idx_embedding_text_chunk_embedding ON text_chunk_embedding FIELDS embedding HNSW DIMENSION 1536;
+DEFINE INDEX IF NOT EXISTS text_chunk_embedding_chunk_id_idx ON text_chunk_embedding FIELDS chunk_id;
+DEFINE INDEX IF NOT EXISTS text_chunk_embedding_user_id_idx ON text_chunk_embedding FIELDS user_id;
+DEFINE INDEX IF NOT EXISTS text_chunk_embedding_source_id_idx ON text_chunk_embedding FIELDS source_id;

common/schemas/text_content.surql

@@ -1,6 +1,6 @@
 # Defines the schema for the 'text_content' table.
 
-DEFINE TABLE IF NOT EXISTS text_content SCHEMALESS;
+DEFINE TABLE IF NOT EXISTS text_content SCHEMAFULL;
 
 # Standard fields
 DEFINE FIELD IF NOT EXISTS created_at ON text_content TYPE datetime;
@@ -12,10 +12,24 @@ DEFINE FIELD IF NOT EXISTS text ON text_content TYPE string;
 DEFINE FIELD IF NOT EXISTS file_info ON text_content TYPE option<object>;
 # UrlInfo is a struct, store as object
 DEFINE FIELD IF NOT EXISTS url_info ON text_content TYPE option<object>;
+DEFINE FIELD IF NOT EXISTS url_info.url ON text_content TYPE string;
+DEFINE FIELD IF NOT EXISTS url_info.title ON text_content TYPE string;
+DEFINE FIELD IF NOT EXISTS url_info.image_id ON text_content TYPE string;
+
 DEFINE FIELD IF NOT EXISTS context ON text_content TYPE option<string>;
 DEFINE FIELD IF NOT EXISTS category ON text_content TYPE string;
 DEFINE FIELD IF NOT EXISTS user_id ON text_content TYPE string;
 
+# FileInfo fields
+DEFINE FIELD IF NOT EXISTS file_info.id ON text_content TYPE string;
+DEFINE FIELD IF NOT EXISTS file_info.created_at ON text_content TYPE datetime;
+DEFINE FIELD IF NOT EXISTS file_info.updated_at ON text_content TYPE datetime;
+DEFINE FIELD IF NOT EXISTS file_info.sha256 ON text_content TYPE string;
+DEFINE FIELD IF NOT EXISTS file_info.path ON text_content TYPE string;
+DEFINE FIELD IF NOT EXISTS file_info.file_name ON text_content TYPE string;
+DEFINE FIELD IF NOT EXISTS file_info.mime_type ON text_content TYPE string;
+DEFINE FIELD IF NOT EXISTS file_info.user_id ON text_content TYPE string;
+
 # Indexes based on query patterns
 DEFINE INDEX IF NOT EXISTS text_content_user_id_idx ON text_content FIELDS user_id;
 DEFINE INDEX IF NOT EXISTS text_content_created_at_idx ON text_content FIELDS created_at;

common/src/error.rs

@@ -5,6 +5,7 @@ use tokio::task::JoinError;
 use crate::storage::types::file_info::FileError;
 
 // Core internal errors
+#[allow(clippy::module_name_repetitions)]
 #[derive(Error, Debug)]
 pub enum AppError {
     #[error("Database error: {0}")]

common/src/lib.rs

@@ -1,3 +1,5 @@
+#![allow(clippy::doc_markdown)]
+//! Shared utilities and storage helpers for the workspace crates.
 pub mod error;
 pub mod storage;
 pub mod utils;

common/src/storage/db.rs

@@ -7,18 +7,20 @@ use include_dir::{include_dir, Dir};
 use std::{ops::Deref, sync::Arc};
 use surrealdb::{
     engine::any::{connect, Any},
-    opt::auth::Root,
+    opt::auth::{Namespace, Root},
     Error, Notification, Surreal,
 };
 use surrealdb_migrations::MigrationRunner;
 use tracing::debug;
 
+/// Embedded SurrealDB migration directory packaged with the crate.
 static MIGRATIONS_DIR: Dir<'_> = include_dir!("$CARGO_MANIFEST_DIR/");
 
 #[derive(Clone)]
 pub struct SurrealDbClient {
     pub client: Surreal<Any>,
 }
+#[allow(clippy::module_name_repetitions)]
 pub trait ProvidesDb {
     fn db(&self) -> &Arc<SurrealDbClient>;
 }
@@ -48,6 +50,24 @@ impl SurrealDbClient {
         Ok(SurrealDbClient { client: db })
     }
 
+    pub async fn new_with_namespace_user(
+        address: &str,
+        namespace: &str,
+        username: &str,
+        password: &str,
+        database: &str,
+    ) -> Result<Self, Error> {
+        let db = connect(address).await?;
+        db.signin(Namespace {
+            namespace,
+            username,
+            password,
+        })
+        .await?;
+        db.use_ns(namespace).use_db(database).await?;
+        Ok(SurrealDbClient { client: db })
+    }
+
     pub async fn create_session_store(
         &self,
     ) -> Result<SessionStore<SessionSurrealPool<Any>>, SessionError> {
@@ -77,24 +97,6 @@ impl SurrealDbClient {
         Ok(())
     }
 
-    /// Operation to rebuild indexes
-    pub async fn rebuild_indexes(&self) -> Result<(), Error> {
-        debug!("Rebuilding indexes");
-        let rebuild_sql = r#"
-            BEGIN TRANSACTION;
-            REBUILD INDEX IF EXISTS idx_embedding_chunks ON text_chunk;
-            REBUILD INDEX IF EXISTS idx_embedding_entities ON knowledge_entity;
-            REBUILD INDEX IF EXISTS text_content_fts_idx ON text_content;
-            REBUILD INDEX IF EXISTS knowledge_entity_fts_name_idx ON knowledge_entity;
-            REBUILD INDEX IF EXISTS knowledge_entity_fts_description_idx ON knowledge_entity;
-            REBUILD INDEX IF EXISTS text_chunk_fts_chunk_idx ON text_chunk;
-            COMMIT TRANSACTION;
-        "#;
-
-        self.client.query(rebuild_sql).await?;
-        Ok(())
-    }
-
     /// Operation to store a object in SurrealDB, requires the struct to implement StoredObject
     ///
     /// # Arguments
@@ -112,6 +114,19 @@ impl SurrealDbClient {
             .await
     }
 
+    /// Operation to upsert an object in SurrealDB, replacing any existing record
+    /// with the same ID. Useful for idempotent ingestion flows.
+    pub async fn upsert_item<T>(&self, item: T) -> Result<Option<T>, Error>
+    where
+        T: StoredObject + Send + Sync + 'static,
+    {
+        let id = item.get_id().to_string();
+        self.client
+            .upsert((T::table_name(), id))
+            .content(item)
+            .await
+    }
+
     /// Operation to retrieve all objects from a certain table, requires the struct to implement StoredObject
     ///
     /// # Returns
@@ -250,6 +265,56 @@ mod tests {
         assert!(fetch_post.is_none());
     }
 
+    #[tokio::test]
+    async fn upsert_item_overwrites_existing_records() {
+        let namespace = "test_ns";
+        let database = &Uuid::new_v4().to_string();
+        let db = SurrealDbClient::memory(namespace, database)
+            .await
+            .expect("Failed to start in-memory surrealdb");
+
+        db.apply_migrations()
+            .await
+            .expect("Failed to initialize schema");
+
+        let mut dummy = Dummy {
+            id: "abc".to_string(),
+            name: "first".to_string(),
+            created_at: Utc::now(),
+            updated_at: Utc::now(),
+        };
+
+        db.store_item(dummy.clone())
+            .await
+            .expect("Failed to store initial record");
+
+        dummy.name = "updated".to_string();
+        let upserted = db
+            .upsert_item(dummy.clone())
+            .await
+            .expect("Failed to upsert record");
+        assert!(upserted.is_some());
+
+        let fetched: Option<Dummy> = db.get_item(&dummy.id).await.expect("fetch after upsert");
+        assert_eq!(fetched.unwrap().name, "updated");
+
+        let new_record = Dummy {
+            id: "def".to_string(),
+            name: "brand-new".to_string(),
+            created_at: Utc::now(),
+            updated_at: Utc::now(),
+        };
+        db.upsert_item(new_record.clone())
+            .await
+            .expect("Failed to upsert new record");
+
+        let fetched_new: Option<Dummy> = db
+            .get_item(&new_record.id)
+            .await
+            .expect("fetch inserted via upsert");
+        assert_eq!(fetched_new, Some(new_record));
+    }
+
     #[tokio::test]
     async fn test_applying_migrations() {
         let namespace = "test_ns";

common/src/storage/indexes.rs

@@ -0,0 +1,795 @@
+use std::time::Duration;
+
+use anyhow::{Context, Result};
+use futures::future::try_join_all;
+use serde::Deserialize;
+use serde_json::{Map, Value};
+use tracing::{debug, info, warn};
+
+use crate::{error::AppError, storage::db::SurrealDbClient};
+
+const INDEX_POLL_INTERVAL: Duration = Duration::from_millis(50);
+const FTS_ANALYZER_NAME: &str = "app_en_fts_analyzer";
+
+#[derive(Clone, Copy)]
+struct HnswIndexSpec {
+    index_name: &'static str,
+    table: &'static str,
+    options: &'static str,
+}
+
+const fn hnsw_index_specs() -> [HnswIndexSpec; 2] {
+    [
+        HnswIndexSpec {
+            index_name: "idx_embedding_text_chunk_embedding",
+            table: "text_chunk_embedding",
+            options: "DIST COSINE TYPE F32 EFC 100 M 8 CONCURRENTLY",
+        },
+        HnswIndexSpec {
+            index_name: "idx_embedding_knowledge_entity_embedding",
+            table: "knowledge_entity_embedding",
+            options: "DIST COSINE TYPE F32 EFC 100 M 8 CONCURRENTLY",
+        },
+    ]
+}
+
+const fn fts_index_specs() -> [FtsIndexSpec; 8] {
+    [
+        FtsIndexSpec {
+            index_name: "text_content_fts_idx",
+            table: "text_content",
+            field: "text",
+            analyzer: Some(FTS_ANALYZER_NAME),
+            method: "BM25",
+        },
+        FtsIndexSpec {
+            index_name: "text_content_context_fts_idx",
+            table: "text_content",
+            field: "context",
+            analyzer: Some(FTS_ANALYZER_NAME),
+            method: "BM25",
+        },
+        FtsIndexSpec {
+            index_name: "text_content_file_name_fts_idx",
+            table: "text_content",
+            field: "file_info.file_name",
+            analyzer: Some(FTS_ANALYZER_NAME),
+            method: "BM25",
+        },
+        FtsIndexSpec {
+            index_name: "text_content_url_fts_idx",
+            table: "text_content",
+            field: "url_info.url",
+            analyzer: Some(FTS_ANALYZER_NAME),
+            method: "BM25",
+        },
+        FtsIndexSpec {
+            index_name: "text_content_url_title_fts_idx",
+            table: "text_content",
+            field: "url_info.title",
+            analyzer: Some(FTS_ANALYZER_NAME),
+            method: "BM25",
+        },
+        FtsIndexSpec {
+            index_name: "knowledge_entity_fts_name_idx",
+            table: "knowledge_entity",
+            field: "name",
+            analyzer: Some(FTS_ANALYZER_NAME),
+            method: "BM25",
+        },
+        FtsIndexSpec {
+            index_name: "knowledge_entity_fts_description_idx",
+            table: "knowledge_entity",
+            field: "description",
+            analyzer: Some(FTS_ANALYZER_NAME),
+            method: "BM25",
+        },
+        FtsIndexSpec {
+            index_name: "text_chunk_fts_chunk_idx",
+            table: "text_chunk",
+            field: "chunk",
+            analyzer: Some(FTS_ANALYZER_NAME),
+            method: "BM25",
+        },
+    ]
+}
+
+impl HnswIndexSpec {
+    fn definition_if_not_exists(&self, dimension: usize) -> String {
+        format!(
+            "DEFINE INDEX IF NOT EXISTS {index} ON TABLE {table} \
+             FIELDS embedding HNSW DIMENSION {dimension} {options};",
+            index = self.index_name,
+            table = self.table,
+            dimension = dimension,
+            options = self.options,
+        )
+    }
+
+    fn definition_overwrite(&self, dimension: usize) -> String {
+        format!(
+            "DEFINE INDEX OVERWRITE {index} ON TABLE {table} \
+             FIELDS embedding HNSW DIMENSION {dimension} {options};",
+            index = self.index_name,
+            table = self.table,
+            dimension = dimension,
+            options = self.options,
+        )
+    }
+}
+
+#[derive(Clone, Copy)]
+struct FtsIndexSpec {
+    index_name: &'static str,
+    table: &'static str,
+    field: &'static str,
+    analyzer: Option<&'static str>,
+    method: &'static str,
+}
+
+impl FtsIndexSpec {
+    fn definition(&self) -> String {
+        let analyzer_clause = self
+            .analyzer
+            .map(|analyzer| format!(" SEARCH ANALYZER {analyzer} {}", self.method))
+            .unwrap_or_default();
+
+        format!(
+            "DEFINE INDEX IF NOT EXISTS {index} ON TABLE {table} FIELDS {field}{analyzer_clause} CONCURRENTLY;",
+            index = self.index_name,
+            table = self.table,
+            field = self.field,
+        )
+    }
+
+    fn overwrite_definition(&self) -> String {
+        let analyzer_clause = self
+            .analyzer
+            .map(|analyzer| format!(" SEARCH ANALYZER {analyzer} {}", self.method))
+            .unwrap_or_default();
+
+        format!(
+            "DEFINE INDEX OVERWRITE {index} ON TABLE {table} FIELDS {field}{analyzer_clause} CONCURRENTLY;",
+            index = self.index_name,
+            table = self.table,
+            field = self.field,
+        )
+    }
+}
+
+/// Build runtime Surreal indexes (FTS + HNSW) using concurrent creation with readiness polling.
+/// Idempotent: safe to call multiple times and will overwrite HNSW definitions when the dimension changes.
+pub async fn ensure_runtime_indexes(
+    db: &SurrealDbClient,
+    embedding_dimension: usize,
+) -> Result<(), AppError> {
+    ensure_runtime_indexes_inner(db, embedding_dimension)
+        .await
+        .map_err(|err| AppError::InternalError(err.to_string()))
+}
+
+/// Rebuild known FTS and HNSW indexes, skipping any that are not yet defined.
+pub async fn rebuild_indexes(db: &SurrealDbClient) -> Result<(), AppError> {
+    rebuild_indexes_inner(db)
+        .await
+        .map_err(|err| AppError::InternalError(err.to_string()))
+}
+
+async fn ensure_runtime_indexes_inner(
+    db: &SurrealDbClient,
+    embedding_dimension: usize,
+) -> Result<()> {
+    create_fts_analyzer(db).await?;
+
+    for spec in fts_index_specs() {
+        if index_exists(db, spec.table, spec.index_name).await? {
+            continue;
+        }
+        // We need to create these sequentially otherwise SurrealDB errors with read/write clash
+        create_index_with_polling(
+            db,
+            spec.definition(),
+            spec.index_name,
+            spec.table,
+            Some(spec.table),
+        )
+        .await?;
+    }
+
+    let hnsw_tasks = hnsw_index_specs().into_iter().map(|spec| async move {
+        match hnsw_index_state(db, &spec, embedding_dimension).await? {
+            HnswIndexState::Missing => {
+                create_index_with_polling(
+                    db,
+                    spec.definition_if_not_exists(embedding_dimension),
+                    spec.index_name,
+                    spec.table,
+                    Some(spec.table),
+                )
+                .await
+            }
+            HnswIndexState::Matches => {
+                let status = get_index_status(db, spec.index_name, spec.table).await?;
+                if status.eq_ignore_ascii_case("error") {
+                    warn!(
+                        index = spec.index_name,
+                        table = spec.table,
+                        "HNSW index found in error state; triggering rebuild"
+                    );
+                    create_index_with_polling(
+                        db,
+                        spec.definition_overwrite(embedding_dimension),
+                        spec.index_name,
+                        spec.table,
+                        Some(spec.table),
+                    )
+                    .await
+                } else {
+                    Ok(())
+                }
+            }
+            HnswIndexState::Different(existing) => {
+                info!(
+                    index = spec.index_name,
+                    table = spec.table,
+                    existing_dimension = existing,
+                    target_dimension = embedding_dimension,
+                    "Overwriting HNSW index to match new embedding dimension"
+                );
+                create_index_with_polling(
+                    db,
+                    spec.definition_overwrite(embedding_dimension),
+                    spec.index_name,
+                    spec.table,
+                    Some(spec.table),
+                )
+                .await
+            }
+        }
+    });
+
+    try_join_all(hnsw_tasks).await.map(|_| ())?;
+
+    Ok(())
+}
+
+async fn get_index_status(db: &SurrealDbClient, index_name: &str, table: &str) -> Result<String> {
+    let info_query = format!("INFO FOR INDEX {index_name} ON TABLE {table};");
+    let mut info_res = db
+        .client
+        .query(info_query)
+        .await
+        .context("checking index status")?;
+    let info: Option<Value> = info_res.take(0).context("failed to take info result")?;
+
+    let info = match info {
+        Some(i) => i,
+        None => return Ok("unknown".to_string()),
+    };
+
+    let building = info.get("building");
+    let status = building
+        .and_then(|b| b.get("status"))
+        .and_then(|s| s.as_str())
+        .unwrap_or("ready")
+        .to_string();
+
+    Ok(status)
+}
+
+async fn rebuild_indexes_inner(db: &SurrealDbClient) -> Result<()> {
+    debug!("Rebuilding indexes with concurrent definitions");
+    create_fts_analyzer(db).await?;
+
+    for spec in fts_index_specs() {
+        if !index_exists(db, spec.table, spec.index_name).await? {
+            debug!(
+                index = spec.index_name,
+                table = spec.table,
+                "Skipping FTS rebuild because index is missing"
+            );
+            continue;
+        }
+
+        create_index_with_polling(
+            db,
+            spec.overwrite_definition(),
+            spec.index_name,
+            spec.table,
+            Some(spec.table),
+        )
+        .await?;
+    }
+
+    let hnsw_tasks = hnsw_index_specs().into_iter().map(|spec| async move {
+        if !index_exists(db, spec.table, spec.index_name).await? {
+            debug!(
+                index = spec.index_name,
+                table = spec.table,
+                "Skipping HNSW rebuild because index is missing"
+            );
+            return Ok(());
+        }
+
+        let Some(dimension) = existing_hnsw_dimension(db, &spec).await? else {
+            warn!(
+                index = spec.index_name,
+                table = spec.table,
+                "HNSW index missing dimension; skipping rebuild"
+            );
+            return Ok(());
+        };
+
+        create_index_with_polling(
+            db,
+            spec.definition_overwrite(dimension),
+            spec.index_name,
+            spec.table,
+            Some(spec.table),
+        )
+        .await
+    });
+
+    try_join_all(hnsw_tasks).await.map(|_| ())
+}
+
+async fn existing_hnsw_dimension(
+    db: &SurrealDbClient,
+    spec: &HnswIndexSpec,
+) -> Result<Option<usize>> {
+    let Some(indexes) = table_index_definitions(db, spec.table).await? else {
+        return Ok(None);
+    };
+
+    let Some(definition) = indexes
+        .get(spec.index_name)
+        .and_then(|details| details.get("Strand"))
+        .and_then(|v| v.as_str())
+    else {
+        return Ok(None);
+    };
+
+    Ok(extract_dimension(definition).and_then(|d| usize::try_from(d).ok()))
+}
+
+async fn hnsw_index_state(
+    db: &SurrealDbClient,
+    spec: &HnswIndexSpec,
+    expected_dimension: usize,
+) -> Result<HnswIndexState> {
+    match existing_hnsw_dimension(db, spec).await? {
+        None => Ok(HnswIndexState::Missing),
+        Some(current_dimension) if current_dimension == expected_dimension => {
+            Ok(HnswIndexState::Matches)
+        }
+        Some(current_dimension) => Ok(HnswIndexState::Different(current_dimension as u64)),
+    }
+}
+
+enum HnswIndexState {
+    Missing,
+    Matches,
+    Different(u64),
+}
+
+fn extract_dimension(definition: &str) -> Option<u64> {
+    definition
+        .split("DIMENSION")
+        .nth(1)
+        .and_then(|rest| rest.split_whitespace().next())
+        .and_then(|token| token.trim_end_matches(';').parse::<u64>().ok())
+}
+
+async fn create_fts_analyzer(db: &SurrealDbClient) -> Result<()> {
+    // Prefer snowball stemming when supported; fall back to ascii-only when the filter
+    // is unavailable in the running Surreal build. Use IF NOT EXISTS to avoid clobbering
+    // an existing analyzer definition.
+    let snowball_query = format!(
+        "DEFINE ANALYZER IF NOT EXISTS {analyzer}
+            TOKENIZERS class
+            FILTERS lowercase, ascii, snowball(english);",
+        analyzer = FTS_ANALYZER_NAME
+    );
+
+    match db.client.query(snowball_query).await {
+        Ok(res) => {
+            if res.check().is_ok() {
+                return Ok(());
+            }
+            warn!(
+                "Snowball analyzer check failed; attempting ascii fallback definition (analyzer: {})",
+                FTS_ANALYZER_NAME
+            );
+        }
+        Err(err) => {
+            warn!(
+                error = %err,
+                "Snowball analyzer creation errored; attempting ascii fallback definition"
+            );
+        }
+    }
+
+    let fallback_query = format!(
+        "DEFINE ANALYZER IF NOT EXISTS {analyzer}
+            TOKENIZERS class
+            FILTERS lowercase, ascii;",
+        analyzer = FTS_ANALYZER_NAME
+    );
+
+    let res = db
+        .client
+        .query(fallback_query)
+        .await
+        .context("creating fallback FTS analyzer")?;
+
+    if let Err(err) = res.check() {
+        warn!(
+            error = %err,
+            "Fallback analyzer creation failed; FTS will run without snowball/ascii analyzer ({})",
+            FTS_ANALYZER_NAME
+        );
+        return Err(err).context("failed to create fallback FTS analyzer");
+    }
+
+    warn!(
+        "Snowball analyzer unavailable; using fallback analyzer ({}) with lowercase+ascii only",
+        FTS_ANALYZER_NAME
+    );
+
+    Ok(())
+}
+
+async fn create_index_with_polling(
+    db: &SurrealDbClient,
+    definition: String,
+    index_name: &str,
+    table: &str,
+    progress_table: Option<&str>,
+) -> Result<()> {
+    let expected_total = match progress_table {
+        Some(table) => Some(count_table_rows(db, table).await.with_context(|| {
+            format!("counting rows in {table} for index {index_name} progress")
+        })?),
+        None => None,
+    };
+
+    let mut attempts = 0;
+    const MAX_ATTEMPTS: usize = 3;
+    loop {
+        attempts += 1;
+        let res = db
+            .client
+            .query(definition.clone())
+            .await
+            .with_context(|| format!("creating index {index_name} on table {table}"))?;
+        match res.check() {
+            Ok(_) => break,
+            Err(err) => {
+                let msg = err.to_string();
+                let conflict = msg.contains("read or write conflict");
+                warn!(
+                    index = %index_name,
+                    table = %table,
+                    error = ?err,
+                    attempt = attempts,
+                    definition = %definition,
+                    "Index definition failed"
+                );
+                if conflict && attempts < MAX_ATTEMPTS {
+                    tokio::time::sleep(Duration::from_millis(100)).await;
+                    continue;
+                }
+                return Err(err).with_context(|| {
+                    format!("index definition failed for {index_name} on {table}")
+                });
+            }
+        }
+    }
+
+    debug!(
+        index = %index_name,
+        table = %table,
+        expected_rows = ?expected_total,
+        "Index definition submitted; waiting for build to finish"
+    );
+
+    poll_index_build_status(db, index_name, table, expected_total, INDEX_POLL_INTERVAL).await
+}
+
+async fn poll_index_build_status(
+    db: &SurrealDbClient,
+    index_name: &str,
+    table: &str,
+    total_rows: Option<u64>,
+    poll_every: Duration,
+) -> Result<()> {
+    let started_at = std::time::Instant::now();
+
+    loop {
+        tokio::time::sleep(poll_every).await;
+
+        let info_query = format!("INFO FOR INDEX {index_name} ON TABLE {table};");
+        let mut info_res =
+            db.client.query(info_query).await.with_context(|| {
+                format!("checking index build status for {index_name} on {table}")
+            })?;
+
+        let info: Option<Value> = info_res
+            .take(0)
+            .context("failed to deserialize INFO FOR INDEX result")?;
+
+        let Some(snapshot) = parse_index_build_info(info, total_rows) else {
+            warn!(
+                index = %index_name,
+                table = %table,
+                "INFO FOR INDEX returned no data; assuming index definition might be missing"
+            );
+            break;
+        };
+
+        match snapshot.progress_pct {
+            Some(pct) => debug!(
+                index = %index_name,
+                table = %table,
+                status = snapshot.status,
+                initial = snapshot.initial,
+                pending = snapshot.pending,
+                updated = snapshot.updated,
+                processed = snapshot.processed,
+                total = snapshot.total_rows,
+                progress_pct = format_args!("{pct:.1}"),
+                "Index build status"
+            ),
+            None => debug!(
+                index = %index_name,
+                table = %table,
+                status = snapshot.status,
+                initial = snapshot.initial,
+                pending = snapshot.pending,
+                updated = snapshot.updated,
+                processed = snapshot.processed,
+                "Index build status"
+            ),
+        }
+
+        if snapshot.is_ready() {
+            debug!(
+                index = %index_name,
+                table = %table,
+                elapsed = ?started_at.elapsed(),
+                processed = snapshot.processed,
+                total = snapshot.total_rows,
+                "Index is ready"
+            );
+            break;
+        }
+
+        if snapshot.status.eq_ignore_ascii_case("error") {
+            warn!(
+                index = %index_name,
+                table = %table,
+                status = snapshot.status,
+                "Index build reported error status; stopping polling"
+            );
+            break;
+        }
+    }
+
+    Ok(())
+}
+
+#[derive(Debug, PartialEq)]
+struct IndexBuildSnapshot {
+    status: String,
+    initial: u64,
+    pending: u64,
+    updated: u64,
+    processed: u64,
+    total_rows: Option<u64>,
+    progress_pct: Option<f64>,
+}
+
+impl IndexBuildSnapshot {
+    fn is_ready(&self) -> bool {
+        self.status.eq_ignore_ascii_case("ready")
+    }
+}
+
+fn parse_index_build_info(
+    info: Option<Value>,
+    total_rows: Option<u64>,
+) -> Option<IndexBuildSnapshot> {
+    let info = info?;
+    let building = info.get("building");
+
+    let status = building
+        .and_then(|b| b.get("status"))
+        .and_then(|s| s.as_str())
+        // If there's no `building` block at all, treat as "ready" (index not building anymore)
+        .unwrap_or("ready")
+        .to_string();
+
+    let initial = building
+        .and_then(|b| b.get("initial"))
+        .and_then(|v| v.as_u64())
+        .unwrap_or(0);
+
+    let pending = building
+        .and_then(|b| b.get("pending"))
+        .and_then(|v| v.as_u64())
+        .unwrap_or(0);
+
+    let updated = building
+        .and_then(|b| b.get("updated"))
+        .and_then(|v| v.as_u64())
+        .unwrap_or(0);
+
+    // `initial` is the number of rows seen when the build started; `updated` accounts for later writes.
+    let processed = initial.saturating_add(updated);
+
+    let progress_pct = total_rows.map(|total| {
+        if total == 0 {
+            0.0
+        } else {
+            ((processed as f64 / total as f64).min(1.0)) * 100.0
+        }
+    });
+
+    Some(IndexBuildSnapshot {
+        status,
+        initial,
+        pending,
+        updated,
+        processed,
+        total_rows,
+        progress_pct,
+    })
+}
+
+#[derive(Debug, Deserialize)]
+struct CountRow {
+    count: u64,
+}
+
+async fn count_table_rows(db: &SurrealDbClient, table: &str) -> Result<u64> {
+    let query = format!("SELECT count() AS count FROM {table} GROUP ALL;");
+    let mut response = db
+        .client
+        .query(query)
+        .await
+        .with_context(|| format!("counting rows in {table}"))?;
+    let rows: Vec<CountRow> = response
+        .take(0)
+        .context("failed to deserialize count() response")?;
+    Ok(rows.first().map_or(0, |r| r.count))
+}
+
+async fn table_index_definitions(
+    db: &SurrealDbClient,
+    table: &str,
+) -> Result<Option<Map<String, Value>>> {
+    let info_query = format!("INFO FOR TABLE {table};");
+    let mut response = db
+        .client
+        .query(info_query)
+        .await
+        .with_context(|| format!("fetching table info for {}", table))?;
+
+    let info: surrealdb::Value = response
+        .take(0)
+        .context("failed to take table info response")?;
+
+    let info_json: Value =
+        serde_json::to_value(info).context("serializing table info to JSON for parsing")?;
+
+    Ok(info_json
+        .get("Object")
+        .and_then(|o| o.get("indexes"))
+        .and_then(|i| i.get("Object"))
+        .and_then(|i| i.as_object())
+        .cloned())
+}
+
+async fn index_exists(db: &SurrealDbClient, table: &str, index_name: &str) -> Result<bool> {
+    let Some(indexes) = table_index_definitions(db, table).await? else {
+        return Ok(false);
+    };
+
+    Ok(indexes.contains_key(index_name))
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use serde_json::json;
+    use uuid::Uuid;
+
+    #[test]
+    fn parse_index_build_info_reports_progress() {
+        let info = json!({
+            "building": {
+                "initial": 56894,
+                "pending": 0,
+                "status": "indexing",
+                "updated": 0
+            }
+        });
+
+        let snapshot = parse_index_build_info(Some(info), Some(61081)).expect("snapshot");
+        assert_eq!(
+            snapshot,
+            IndexBuildSnapshot {
+                status: "indexing".to_string(),
+                initial: 56894,
+                pending: 0,
+                updated: 0,
+                processed: 56894,
+                total_rows: Some(61081),
+                progress_pct: Some((56894_f64 / 61081_f64) * 100.0),
+            }
+        );
+        assert!(!snapshot.is_ready());
+    }
+
+    #[test]
+    fn parse_index_build_info_defaults_to_ready_when_no_building_block() {
+        // Surreal returns `{}` when the index exists but isn't building.
+        let info = json!({});
+        let snapshot = parse_index_build_info(Some(info), Some(10)).expect("snapshot");
+        assert!(snapshot.is_ready());
+        assert_eq!(snapshot.processed, 0);
+        assert_eq!(snapshot.progress_pct, Some(0.0));
+    }
+
+    #[test]
+    fn extract_dimension_parses_value() {
+        let definition = "DEFINE INDEX idx_embedding_text_chunk_embedding ON TABLE text_chunk_embedding FIELDS embedding HNSW DIMENSION 1536 DIST COSINE TYPE F32 EFC 100 M 8;";
+        assert_eq!(extract_dimension(definition), Some(1536));
+    }
+
+    #[tokio::test]
+    async fn ensure_runtime_indexes_is_idempotent() {
+        let namespace = "indexes_ns";
+        let database = &Uuid::new_v4().to_string();
+        let db = SurrealDbClient::memory(namespace, database)
+            .await
+            .expect("in-memory db");
+
+        db.apply_migrations()
+            .await
+            .expect("migrations should succeed");
+
+        // First run creates everything
+        ensure_runtime_indexes(&db, 1536)
+            .await
+            .expect("initial index creation");
+
+        // Second run should be a no-op and still succeed
+        ensure_runtime_indexes(&db, 1536)
+            .await
+            .expect("second index creation");
+    }
+
+    #[tokio::test]
+    async fn ensure_hnsw_index_overwrites_dimension() {
+        let namespace = "indexes_dim";
+        let database = &Uuid::new_v4().to_string();
+        let db = SurrealDbClient::memory(namespace, database)
+            .await
+            .expect("in-memory db");
+
+        db.apply_migrations()
+            .await
+            .expect("migrations should succeed");
+
+        // Create initial index with default dimension
+        ensure_runtime_indexes(&db, 1536)
+            .await
+            .expect("initial index creation");
+
+        // Change dimension and ensure overwrite path is exercised
+        ensure_runtime_indexes(&db, 128)
+            .await
+            .expect("overwritten index creation");
+    }
+}

common/src/storage/mod.rs

@@ -1,3 +1,4 @@
 pub mod db;
+pub mod indexes;
 pub mod store;
 pub mod types;

common/src/storage/store.rs

@@ -6,6 +6,7 @@ use anyhow::{anyhow, Result as AnyResult};
 use bytes::Bytes;
 use futures::stream::BoxStream;
 use futures::{StreamExt, TryStreamExt};
+use object_store::aws::AmazonS3Builder;
 use object_store::local::LocalFileSystem;
 use object_store::memory::InMemory;
 use object_store::{path::Path as ObjPath, ObjectStore};
@@ -17,8 +18,11 @@ pub type DynStore = Arc<dyn ObjectStore>;
 /// Storage manager with persistent state and proper lifecycle management.
 #[derive(Clone)]
 pub struct StorageManager {
+    // Store from objectstore wrapped as dyn
     store: DynStore,
+    // Simple enum to track which kind
     backend_kind: StorageKind,
+    // Where on disk
     local_base: Option<PathBuf>,
 }
 
@@ -183,7 +187,7 @@ impl StorageManager {
 
         while current.starts_with(base) && current.as_path() != base.as_path() {
             match tokio::fs::remove_dir(&current).await {
-                Ok(_) => {}
+                Ok(()) => {}
                 Err(err) => match err.kind() {
                     ErrorKind::NotFound => {}
                     ErrorKind::DirectoryNotEmpty => break,
@@ -231,6 +235,39 @@ async fn create_storage_backend(
             let store = InMemory::new();
             Ok((Arc::new(store), None))
         }
+        StorageKind::S3 => {
+            let bucket = cfg
+                .s3_bucket
+                .as_ref()
+                .ok_or_else(|| object_store::Error::Generic {
+                    store: "S3",
+                    source: anyhow!("s3_bucket is required for S3 storage").into(),
+                })?;
+
+            let mut builder = AmazonS3Builder::new()
+                .with_bucket_name(bucket)
+                .with_allow_http(true);
+
+            if let (Ok(key), Ok(secret)) = (
+                std::env::var("AWS_ACCESS_KEY_ID"),
+                std::env::var("AWS_SECRET_ACCESS_KEY"),
+            ) {
+                builder = builder
+                    .with_access_key_id(key)
+                    .with_secret_access_key(secret);
+            }
+
+            if let Some(endpoint) = &cfg.s3_endpoint {
+                builder = builder.with_endpoint(endpoint);
+            }
+
+            if let Some(region) = &cfg.s3_region {
+                builder = builder.with_region(region);
+            }
+
+            let store = builder.build()?;
+            Ok((Arc::new(store), None))
+        }
     }
 }
 
@@ -244,6 +281,33 @@ pub mod testing {
     use crate::utils::config::{AppConfig, PdfIngestMode};
     use uuid;
 
+    const DEFAULT_TEST_S3_BUCKET: &str = "minne-tests";
+    const DEFAULT_TEST_S3_ENDPOINT: &str = "http://127.0.0.1:19000";
+
+    fn configured_test_s3_bucket() -> String {
+        std::env::var("MINNE_TEST_S3_BUCKET")
+            .ok()
+            .filter(|value| !value.trim().is_empty())
+            .or_else(|| {
+                std::env::var("S3_BUCKET")
+                    .ok()
+                    .filter(|value| !value.trim().is_empty())
+            })
+            .unwrap_or_else(|| DEFAULT_TEST_S3_BUCKET.to_string())
+    }
+
+    fn configured_test_s3_endpoint() -> String {
+        std::env::var("MINNE_TEST_S3_ENDPOINT")
+            .ok()
+            .filter(|value| !value.trim().is_empty())
+            .or_else(|| {
+                std::env::var("S3_ENDPOINT")
+                    .ok()
+                    .filter(|value| !value.trim().is_empty())
+            })
+            .unwrap_or_else(|| DEFAULT_TEST_S3_ENDPOINT.to_string())
+    }
+
     /// Create a test configuration with memory storage.
     ///
     /// This provides a ready-to-use configuration for testing scenarios
@@ -287,6 +351,30 @@ pub mod testing {
         }
     }
 
+    /// Create a test configuration with S3 storage (MinIO).
+    ///
+    /// Uses `MINNE_TEST_S3_ENDPOINT` / `S3_ENDPOINT` and
+    /// `MINNE_TEST_S3_BUCKET` / `S3_BUCKET` when provided.
+    pub fn test_config_s3() -> AppConfig {
+        AppConfig {
+            openai_api_key: "test".into(),
+            surrealdb_address: "test".into(),
+            surrealdb_username: "test".into(),
+            surrealdb_password: "test".into(),
+            surrealdb_namespace: "test".into(),
+            surrealdb_database: "test".into(),
+            data_dir: "/tmp/unused".into(),
+            http_port: 0,
+            openai_base_url: "..".into(),
+            storage: StorageKind::S3,
+            s3_bucket: Some(configured_test_s3_bucket()),
+            s3_endpoint: Some(configured_test_s3_endpoint()),
+            s3_region: Some("us-east-1".into()),
+            pdf_ingest_mode: PdfIngestMode::LlmFirst,
+            ..Default::default()
+        }
+    }
+
     /// A specialized StorageManager for testing scenarios.
     ///
     /// This provides automatic setup for memory storage with proper isolation
@@ -329,6 +417,30 @@ pub mod testing {
             })
         }
 
+        /// Create a new TestStorageManager with S3 backend (MinIO).
+        ///
+        /// This requires a reachable MinIO endpoint and an existing test bucket.
+        pub async fn new_s3() -> object_store::Result<Self> {
+            // Ensure credentials are set for MinIO
+            // We set these env vars for the process, which AmazonS3Builder will pick up
+            std::env::set_var("AWS_ACCESS_KEY_ID", "minioadmin");
+            std::env::set_var("AWS_SECRET_ACCESS_KEY", "minioadmin");
+            std::env::set_var("AWS_REGION", "us-east-1");
+
+            let cfg = test_config_s3();
+            let storage = StorageManager::new(&cfg).await?;
+
+            // Probe the bucket so tests can cleanly skip when the endpoint is unreachable
+            // or the test bucket is not provisioned.
+            let probe_prefix = format!("__minne_s3_probe__/{}", uuid::Uuid::new_v4());
+            storage.list(Some(&probe_prefix)).await?;
+
+            Ok(Self {
+                storage,
+                _temp_dir: None,
+            })
+        }
+
         /// Create a TestStorageManager with custom configuration.
         pub async fn with_config(cfg: &AppConfig) -> object_store::Result<Self> {
             let storage = StorageManager::new(cfg).await?;
@@ -366,6 +478,14 @@ pub mod testing {
             self.storage.get(location).await
         }
 
+        /// Get a streaming handle for test data.
+        pub async fn get_stream(
+            &self,
+            location: &str,
+        ) -> object_store::Result<BoxStream<'static, object_store::Result<Bytes>>> {
+            self.storage.get_stream(location).await
+        }
+
         /// Delete test data below the specified prefix.
         pub async fn delete_prefix(&self, prefix: &str) -> object_store::Result<()> {
             self.storage.delete_prefix(prefix).await
@@ -834,4 +954,117 @@ mod tests {
         // Verify it's using memory backend
         assert_eq!(*test_storage.storage().backend_kind(), StorageKind::Memory);
     }
+
+    // S3 Tests - Require a reachable MinIO endpoint and test bucket.
+    // `TestStorageManager::new_s3()` probes connectivity and these tests auto-skip when unavailable.
+
+    #[tokio::test]
+    async fn test_storage_manager_s3_basic_operations() {
+        // Skip if S3 connection fails (e.g. no MinIO)
+        let Ok(storage) = testing::TestStorageManager::new_s3().await else {
+            eprintln!("Skipping S3 test (setup failed)");
+            return;
+        };
+
+        let prefix = format!("test-basic-{}", Uuid::new_v4());
+        let location = format!("{prefix}/file.txt");
+        let data = b"test data for S3";
+
+        // Test put
+        if let Err(e) = storage.put(&location, data).await {
+            eprintln!("Skipping S3 test (put failed - bucket missing?): {e}");
+            return;
+        }
+
+        // Test get
+        let retrieved = storage.get(&location).await.expect("get");
+        assert_eq!(retrieved.as_ref(), data);
+
+        // Test exists
+        assert!(storage.exists(&location).await.expect("exists"));
+
+        // Test delete
+        storage
+            .delete_prefix(&format!("{prefix}/"))
+            .await
+            .expect("delete");
+        assert!(!storage
+            .exists(&location)
+            .await
+            .expect("exists after delete"));
+    }
+
+    #[tokio::test]
+    async fn test_storage_manager_s3_list_operations() {
+        let Ok(storage) = testing::TestStorageManager::new_s3().await else {
+            return;
+        };
+
+        let prefix = format!("test-list-{}", Uuid::new_v4());
+        let files = vec![
+            (format!("{prefix}/file1.txt"), b"content1"),
+            (format!("{prefix}/file2.txt"), b"content2"),
+            (format!("{prefix}/sub/file3.txt"), b"content3"),
+        ];
+
+        for (loc, data) in &files {
+            if storage.put(loc, *data).await.is_err() {
+                return; // Abort if put fails
+            }
+        }
+
+        // List with prefix
+        let list_prefix = format!("{prefix}/");
+        let items = storage.list(Some(&list_prefix)).await.expect("list");
+        assert_eq!(items.len(), 3);
+
+        // Cleanup
+        storage.delete_prefix(&list_prefix).await.expect("cleanup");
+    }
+
+    #[tokio::test]
+    async fn test_storage_manager_s3_stream_operations() {
+        let Ok(storage) = testing::TestStorageManager::new_s3().await else {
+            return;
+        };
+
+        let prefix = format!("test-stream-{}", Uuid::new_v4());
+        let location = format!("{prefix}/large.bin");
+        let content = vec![42u8; 1024 * 10]; // 10KB
+
+        if storage.put(&location, &content).await.is_err() {
+            return;
+        }
+
+        let mut stream = storage.get_stream(&location).await.expect("get stream");
+        let mut collected = Vec::new();
+        while let Some(chunk) = stream.next().await {
+            collected.extend_from_slice(&chunk.expect("chunk"));
+        }
+        assert_eq!(collected, content);
+
+        storage
+            .delete_prefix(&format!("{prefix}/"))
+            .await
+            .expect("cleanup");
+    }
+
+    #[tokio::test]
+    async fn test_storage_manager_s3_backend_kind() {
+        let Ok(storage) = testing::TestStorageManager::new_s3().await else {
+            return;
+        };
+        assert_eq!(*storage.storage().backend_kind(), StorageKind::S3);
+    }
+
+    #[tokio::test]
+    async fn test_storage_manager_s3_error_handling() {
+        let Ok(storage) = testing::TestStorageManager::new_s3().await else {
+            return;
+        };
+
+        let location = format!("nonexistent-{}/file.txt", Uuid::new_v4());
+        assert!(storage.get(&location).await.is_err());
+        assert!(!storage.exists(&location).await.expect("exists check"));
+    }
 }

common/src/storage/types/analytics.rs

@@ -71,6 +71,7 @@ impl Analytics {
         // We need to use a direct query for COUNT aggregation
         #[derive(Debug, Deserialize)]
         struct CountResult {
+            /// Total user count.
             count: i64,
         }
 
@@ -81,7 +82,7 @@ impl Analytics {
             .await?
             .take(0)?;
 
-        Ok(result.map(|r| r.count).unwrap_or(0))
+        Ok(result.map_or(0, |r| r.count))
     }
 }
 

common/src/storage/types/conversation.rs

@@ -10,6 +10,54 @@ stored_object!(Conversation, "conversation", {
     title: String
 });
 
+#[derive(Debug, Clone, serde::Serialize, serde::Deserialize, PartialEq, Eq)]
+pub struct SidebarConversation {
+    #[serde(deserialize_with = "deserialize_sidebar_id")]
+    pub id: String,
+    pub title: String,
+}
+
+struct SidebarIdVisitor;
+
+impl<'de> serde::de::Visitor<'de> for SidebarIdVisitor {
+    type Value = String;
+
+    fn expecting(&self, formatter: &mut std::fmt::Formatter) -> std::fmt::Result {
+        formatter.write_str("a string id or a SurrealDB Thing")
+    }
+
+    fn visit_str<E>(self, value: &str) -> Result<Self::Value, E>
+    where
+        E: serde::de::Error,
+    {
+        Ok(value.to_string())
+    }
+
+    fn visit_string<E>(self, value: String) -> Result<Self::Value, E>
+    where
+        E: serde::de::Error,
+    {
+        Ok(value)
+    }
+
+    fn visit_map<A>(self, map: A) -> Result<Self::Value, A::Error>
+    where
+        A: serde::de::MapAccess<'de>,
+    {
+        let thing = <surrealdb::sql::Thing as serde::Deserialize>::deserialize(
+            serde::de::value::MapAccessDeserializer::new(map),
+        )?;
+        Ok(thing.id.to_raw())
+    }
+}
+
+fn deserialize_sidebar_id<'de, D>(deserializer: D) -> Result<String, D::Error>
+where
+    D: serde::Deserializer<'de>,
+{
+    deserializer.deserialize_any(SidebarIdVisitor)
+}
+
 impl Conversation {
     pub fn new(user_id: String, title: String) -> Self {
         let now = Utc::now();
@@ -75,6 +123,23 @@ impl Conversation {
 
         Ok(())
     }
+
+    pub async fn get_user_sidebar_conversations(
+        user_id: &str,
+        db: &SurrealDbClient,
+    ) -> Result<Vec<SidebarConversation>, AppError> {
+        let conversations: Vec<SidebarConversation> = db
+            .client
+            .query(
+                "SELECT id, title, updated_at FROM type::table($table_name) WHERE user_id = $user_id ORDER BY updated_at DESC",
+            )
+            .bind(("table_name", Self::table_name()))
+            .bind(("user_id", user_id.to_string()))
+            .await?
+            .take(0)?;
+
+        Ok(conversations)
+    }
 }
 
 #[cfg(test)]
@@ -249,6 +314,96 @@ mod tests {
         }
     }
 
+    #[tokio::test]
+    async fn test_get_user_sidebar_conversations_filters_and_orders_by_updated_at_desc() {
+        let namespace = "test_ns";
+        let database = &Uuid::new_v4().to_string();
+        let db = SurrealDbClient::memory(namespace, database)
+            .await
+            .expect("Failed to start in-memory surrealdb");
+
+        let user_id = "sidebar_user";
+        let other_user_id = "other_user";
+        let base = Utc::now();
+
+        let mut oldest = Conversation::new(user_id.to_string(), "Oldest".to_string());
+        oldest.updated_at = base - chrono::Duration::minutes(30);
+
+        let mut newest = Conversation::new(user_id.to_string(), "Newest".to_string());
+        newest.updated_at = base - chrono::Duration::minutes(5);
+
+        let mut middle = Conversation::new(user_id.to_string(), "Middle".to_string());
+        middle.updated_at = base - chrono::Duration::minutes(15);
+
+        let mut other_user = Conversation::new(other_user_id.to_string(), "Other".to_string());
+        other_user.updated_at = base;
+
+        db.store_item(oldest.clone())
+            .await
+            .expect("Failed to store oldest conversation");
+        db.store_item(newest.clone())
+            .await
+            .expect("Failed to store newest conversation");
+        db.store_item(middle.clone())
+            .await
+            .expect("Failed to store middle conversation");
+        db.store_item(other_user)
+            .await
+            .expect("Failed to store other-user conversation");
+
+        let sidebar_items = Conversation::get_user_sidebar_conversations(user_id, &db)
+            .await
+            .expect("Failed to get sidebar conversations");
+
+        assert_eq!(sidebar_items.len(), 3);
+        assert_eq!(sidebar_items[0].id, newest.id);
+        assert_eq!(sidebar_items[0].title, "Newest");
+        assert_eq!(sidebar_items[1].id, middle.id);
+        assert_eq!(sidebar_items[1].title, "Middle");
+        assert_eq!(sidebar_items[2].id, oldest.id);
+        assert_eq!(sidebar_items[2].title, "Oldest");
+    }
+
+    #[tokio::test]
+    async fn test_sidebar_projection_reflects_patch_title_and_updated_at_reorder() {
+        let namespace = "test_ns";
+        let database = &Uuid::new_v4().to_string();
+        let db = SurrealDbClient::memory(namespace, database)
+            .await
+            .expect("Failed to start in-memory surrealdb");
+
+        let user_id = "sidebar_patch_user";
+        let base = Utc::now();
+
+        let mut first = Conversation::new(user_id.to_string(), "First".to_string());
+        first.updated_at = base - chrono::Duration::minutes(20);
+
+        let mut second = Conversation::new(user_id.to_string(), "Second".to_string());
+        second.updated_at = base - chrono::Duration::minutes(10);
+
+        db.store_item(first.clone())
+            .await
+            .expect("Failed to store first conversation");
+        db.store_item(second.clone())
+            .await
+            .expect("Failed to store second conversation");
+
+        let before_patch = Conversation::get_user_sidebar_conversations(user_id, &db)
+            .await
+            .expect("Failed to get sidebar conversations before patch");
+        assert_eq!(before_patch[0].id, second.id);
+
+        Conversation::patch_title(&first.id, user_id, "First (renamed)", &db)
+            .await
+            .expect("Failed to patch conversation title");
+
+        let after_patch = Conversation::get_user_sidebar_conversations(user_id, &db)
+            .await
+            .expect("Failed to get sidebar conversations after patch");
+        assert_eq!(after_patch[0].id, first.id);
+        assert_eq!(after_patch[0].title, "First (renamed)");
+    }
+
     #[tokio::test]
     async fn test_get_complete_conversation_with_messages() {
         // Setup in-memory database for testing

common/src/storage/types/file_info.rs

@@ -9,6 +9,7 @@ use std::{
 };
 use tempfile::NamedTempFile;
 use thiserror::Error;
+use tokio::task;
 use tracing::info;
 use uuid::Uuid;
 
@@ -71,21 +72,29 @@ impl FileInfo {
     ///
     /// # Returns
     /// * `Result<String, FileError>` - The SHA256 hash as a hex string or an error.
+    #[allow(clippy::indexing_slicing)]
     async fn get_sha(file: &NamedTempFile) -> Result<String, FileError> {
-        let mut reader = BufReader::new(file.as_file());
-        let mut hasher = Sha256::new();
-        let mut buffer = [0u8; 8192]; // 8KB buffer
+        let mut file_clone = file.as_file().try_clone()?;
 
-        loop {
-            let n = reader.read(&mut buffer)?;
-            if n == 0 {
-                break;
+        let digest = task::spawn_blocking(move || -> Result<_, std::io::Error> {
+            let mut reader = BufReader::new(&mut file_clone);
+            let mut hasher = Sha256::new();
+            let mut buffer = [0u8; 8192]; // 8KB buffer
+
+            loop {
+                let n = reader.read(&mut buffer)?;
+                if n == 0 {
+                    break;
+                }
+                hasher.update(&buffer[..n]);
             }
-            hasher.update(&buffer[..n]);
-        }
 
-        let digest = hasher.finalize();
-        Ok(format!("{:x}", digest))
+            Ok::<_, std::io::Error>(hasher.finalize())
+        })
+        .await
+        .map_err(std::io::Error::other)??;
+
+        Ok(format!("{digest:x}"))
     }
 
     /// Sanitizes the file name to prevent security vulnerabilities like directory traversal.
@@ -103,7 +112,7 @@ impl FileInfo {
                     }
                 })
                 .collect();
-            format!("{}{}", sanitized_name, ext)
+            format!("{sanitized_name}{ext}")
         } else {
             // No extension
             file_name
@@ -128,8 +137,12 @@ impl FileInfo {
     /// # Returns
     /// * `Result<Option<FileInfo>, FileError>` - The `FileInfo` or `None` if not found.
     async fn get_by_sha(sha256: &str, db_client: &SurrealDbClient) -> Result<FileInfo, FileError> {
-        let query = format!("SELECT * FROM file WHERE sha256 = '{}'", &sha256);
-        let response: Vec<FileInfo> = db_client.client.query(query).await?.take(0)?;
+        let mut response = db_client
+            .client
+            .query("SELECT * FROM file WHERE sha256 = $sha256 LIMIT 1")
+            .bind(("sha256", sha256.to_owned()))
+            .await?;
+        let response: Vec<FileInfo> = response.take(0)?;
 
         response
             .into_iter()
@@ -292,7 +305,7 @@ impl FileInfo {
         storage: &StorageManager,
     ) -> Result<String, FileError> {
         // Logical object location relative to the store root
-        let location = format!("{}/{}/{}", user_id, uuid, file_name);
+        let location = format!("{user_id}/{uuid}/{file_name}");
         info!("Persisting to object location: {}", location);
 
         let bytes = tokio::fs::read(file.path()).await?;
@@ -656,6 +669,36 @@ mod tests {
         }
     }
 
+    #[tokio::test]
+    async fn test_get_by_sha_resists_query_injection() {
+        let namespace = "test_ns";
+        let database = &Uuid::new_v4().to_string();
+        let db = SurrealDbClient::memory(namespace, database)
+            .await
+            .expect("Failed to start in-memory surrealdb");
+
+        let now = Utc::now();
+        let file_info = FileInfo {
+            id: Uuid::new_v4().to_string(),
+            created_at: now,
+            updated_at: now,
+            user_id: "user123".to_string(),
+            sha256: "known_sha_value".to_string(),
+            path: "/path/to/file.txt".to_string(),
+            file_name: "file.txt".to_string(),
+            mime_type: "text/plain".to_string(),
+        };
+
+        db.store_item(file_info)
+            .await
+            .expect("Failed to store test file info");
+
+        let malicious_sha = "known_sha_value' OR true --";
+        let result = FileInfo::get_by_sha(malicious_sha, &db).await;
+
+        assert!(matches!(result, Err(FileError::FileNotFound(_))));
+    }
+
     #[tokio::test]
     async fn test_manual_file_info_creation() {
         let namespace = "test_ns";

common/src/storage/types/ingestion_payload.rs

@@ -1,3 +1,9 @@
+#![allow(
+    clippy::result_large_err,
+    clippy::needless_pass_by_value,
+    clippy::implicit_clone,
+    clippy::semicolon_if_nothing_returned
+)]
 use crate::{error::AppError, storage::types::file_info::FileInfo};
 use serde::{Deserialize, Serialize};
 use tracing::info;
@@ -38,6 +44,7 @@ impl IngestionPayload {
     /// # Returns
     /// * `Result<Vec<IngestionPayload>, AppError>` - On success, returns a vector of ingress objects
     ///   (one per file/content type). On failure, returns an `AppError`.
+    #[allow(clippy::similar_names)]
     pub fn create_ingestion_payload(
         content: Option<String>,
         context: String,

common/src/storage/types/ingestion_task.rs

@@ -1,3 +1,12 @@
+#![allow(
+    clippy::cast_possible_wrap,
+    clippy::items_after_statements,
+    clippy::arithmetic_side_effects,
+    clippy::cast_sign_loss,
+    clippy::missing_docs_in_private_items,
+    clippy::trivially_copy_pass_by_ref,
+    clippy::expect_used
+)]
 use std::time::Duration;
 
 use chrono::Duration as ChronoDuration;

common/src/storage/types/knowledge_entity.rs

@@ -1,7 +1,19 @@
+#![allow(
+    clippy::missing_docs_in_private_items,
+    clippy::module_name_repetitions,
+    clippy::match_same_arms,
+    clippy::format_push_string,
+    clippy::uninlined_format_args,
+    clippy::explicit_iter_loop,
+    clippy::items_after_statements,
+    clippy::get_first,
+    clippy::redundant_closure_for_method_calls
+)]
 use std::collections::HashMap;
 
 use crate::{
-    error::AppError, storage::db::SurrealDbClient, stored_object,
+    error::AppError, storage::db::SurrealDbClient,
+    storage::types::knowledge_entity_embedding::KnowledgeEntityEmbedding, stored_object,
     utils::embedding::generate_embedding,
 };
 use async_openai::{config::OpenAIConfig, Client};
@@ -78,10 +90,16 @@ stored_object!(KnowledgeEntity, "knowledge_entity", {
     description: String,
     entity_type: KnowledgeEntityType,
     metadata: Option<serde_json::Value>,
-    embedding: Vec<f32>,
     user_id: String
 });
 
+/// Vector search result including hydrated entity.
+#[derive(Debug, Deserialize, Serialize, Clone, PartialEq)]
+pub struct KnowledgeEntityVectorResult {
+    pub entity: KnowledgeEntity,
+    pub score: f32,
+}
+
 impl KnowledgeEntity {
     pub fn new(
         source_id: String,
@@ -89,7 +107,6 @@ impl KnowledgeEntity {
         description: String,
         entity_type: KnowledgeEntityType,
         metadata: Option<serde_json::Value>,
-        embedding: Vec<f32>,
         user_id: String,
     ) -> Self {
         let now = Utc::now();
@@ -102,7 +119,6 @@ impl KnowledgeEntity {
             description,
             entity_type,
             metadata,
-            embedding,
             user_id,
         }
     }
@@ -155,16 +171,107 @@ impl KnowledgeEntity {
         source_id: &str,
         db_client: &SurrealDbClient,
     ) -> Result<(), AppError> {
-        let query = format!(
-            "DELETE {} WHERE source_id = '{}'",
-            Self::table_name(),
-            source_id
-        );
-        db_client.query(query).await?;
+        // Delete embeddings first, while we can still look them up via the entity's source_id
+        KnowledgeEntityEmbedding::delete_by_source_id(source_id, db_client).await?;
+
+        db_client
+            .client
+            .query("DELETE FROM type::table($table) WHERE source_id = $source_id")
+            .bind(("table", Self::table_name()))
+            .bind(("source_id", source_id.to_owned()))
+            .await
+            .map_err(AppError::Database)?
+            .check()
+            .map_err(AppError::Database)?;
 
         Ok(())
     }
 
+    /// Atomically store a knowledge entity and its embedding.
+    /// Writes the entity to `knowledge_entity` and the embedding to `knowledge_entity_embedding`.
+    pub async fn store_with_embedding(
+        entity: KnowledgeEntity,
+        embedding: Vec<f32>,
+        db: &SurrealDbClient,
+    ) -> Result<(), AppError> {
+        let emb = KnowledgeEntityEmbedding::new(&entity.id, embedding, entity.user_id.clone());
+
+        let query = format!(
+            "
+            BEGIN TRANSACTION;
+              CREATE type::thing('{entity_table}', $entity_id) CONTENT $entity;
+              CREATE type::thing('{emb_table}', $emb_id) CONTENT $emb;
+            COMMIT TRANSACTION;
+            ",
+            entity_table = Self::table_name(),
+            emb_table = KnowledgeEntityEmbedding::table_name(),
+        );
+
+        db.client
+            .query(query)
+            .bind(("entity_id", entity.id.clone()))
+            .bind(("entity", entity))
+            .bind(("emb_id", emb.id.clone()))
+            .bind(("emb", emb))
+            .await
+            .map_err(AppError::Database)?
+            .check()
+            .map_err(AppError::Database)?;
+
+        Ok(())
+    }
+
+    /// Vector search over knowledge entities using the embedding table, fetching full entity rows and scores.
+    pub async fn vector_search(
+        take: usize,
+        query_embedding: Vec<f32>,
+        db: &SurrealDbClient,
+        user_id: &str,
+    ) -> Result<Vec<KnowledgeEntityVectorResult>, AppError> {
+        #[derive(Deserialize)]
+        struct Row {
+            entity_id: Option<KnowledgeEntity>,
+            score: f32,
+        }
+
+        let sql = format!(
+            r#"
+            SELECT
+                entity_id,
+                vector::similarity::cosine(embedding, $embedding) AS score
+            FROM {emb_table}
+            WHERE user_id = $user_id
+              AND embedding <|{take},100|> $embedding
+            ORDER BY score DESC
+            LIMIT {take}
+            FETCH entity_id;
+            "#,
+            emb_table = KnowledgeEntityEmbedding::table_name(),
+            take = take
+        );
+
+        let mut response = db
+            .query(&sql)
+            .bind(("embedding", query_embedding))
+            .bind(("user_id", user_id.to_string()))
+            .await
+            .map_err(|e| AppError::InternalError(format!("Surreal query failed: {e}")))?;
+
+        response = response.check().map_err(AppError::Database)?;
+
+        let rows: Vec<Row> = response.take::<Vec<Row>>(0).map_err(AppError::Database)?;
+
+        Ok(rows
+            .into_iter()
+            .filter_map(|r| {
+                r.entity_id.map(|entity| KnowledgeEntityVectorResult {
+                    entity,
+                    score: r.score,
+                })
+            })
+            .collect())
+    }
+
     pub async fn patch(
         id: &str,
         name: &str,
@@ -178,32 +285,55 @@ impl KnowledgeEntity {
             name, description, entity_type
         );
         let embedding = generate_embedding(ai_client, &embedding_input, db_client).await?;
+        let user_id = Self::get_user_id_by_id(id, db_client).await?;
+        let emb = KnowledgeEntityEmbedding::new(id, embedding, user_id);
 
         let now = Utc::now();
 
         db_client
             .client
             .query(
-                "UPDATE type::thing($table, $id)
-                SET name = $name,
-                    description = $description,
-                    updated_at = $updated_at,
-                    entity_type = $entity_type,
-                    embedding = $embedding
-                RETURN AFTER",
+                "BEGIN TRANSACTION;
+                 UPDATE type::thing($table, $id)
+                 SET name = $name,
+                     description = $description,
+                     updated_at = $updated_at,
+                     entity_type = $entity_type;
+                 UPSERT type::thing($emb_table, $emb_id) CONTENT $emb;
+                 COMMIT TRANSACTION;",
             )
             .bind(("table", Self::table_name()))
+            .bind(("emb_table", KnowledgeEntityEmbedding::table_name()))
             .bind(("id", id.to_string()))
             .bind(("name", name.to_string()))
             .bind(("updated_at", surrealdb::Datetime::from(now)))
             .bind(("entity_type", entity_type.to_owned()))
-            .bind(("embedding", embedding))
+            .bind(("emb_id", emb.id.clone()))
+            .bind(("emb", emb))
             .bind(("description", description.to_string()))
             .await?;
 
         Ok(())
     }
 
+    async fn get_user_id_by_id(id: &str, db_client: &SurrealDbClient) -> Result<String, AppError> {
+        let mut response = db_client
+            .client
+            .query("SELECT user_id FROM type::thing($table, $id) LIMIT 1")
+            .bind(("table", Self::table_name()))
+            .bind(("id", id.to_string()))
+            .await
+            .map_err(AppError::Database)?;
+        #[derive(Deserialize)]
+        struct Row {
+            user_id: String,
+        }
+        let rows: Vec<Row> = response.take(0).map_err(AppError::Database)?;
+        rows.get(0)
+            .map(|r| r.user_id.clone())
+            .ok_or_else(|| AppError::InternalError("user not found for entity".to_string()))
+    }
+
     /// Re-creates embeddings for all knowledge entities in the database.
     ///
     /// This is a costly operation that should be run in the background. It follows the same
@@ -228,22 +358,13 @@ impl KnowledgeEntity {
         if total_entities == 0 {
             info!("No knowledge entities to update. Just updating the idx");
 
-            let mut transaction_query = String::from("BEGIN TRANSACTION;");
-            transaction_query
-                .push_str("REMOVE INDEX idx_embedding_entities ON TABLE knowledge_entity;");
-            transaction_query.push_str(&format!(
-                "DEFINE INDEX idx_embedding_entities ON TABLE knowledge_entity FIELDS embedding HNSW DIMENSION {};",
-                new_dimensions
-            ));
-            transaction_query.push_str("COMMIT TRANSACTION;");
-
-            db.query(transaction_query).await?;
+            KnowledgeEntityEmbedding::redefine_hnsw_index(db, new_dimensions as usize).await?;
             return Ok(());
         }
         info!("Found {} entities to process.", total_entities);
 
         // Generate all new embeddings in memory
-        let mut new_embeddings: HashMap<String, Vec<f32>> = HashMap::new();
+        let mut new_embeddings: HashMap<String, (Vec<f32>, String)> = HashMap::new();
         info!("Generating new embeddings for all entities...");
         for entity in all_entities.iter() {
             let embedding_input = format!(
@@ -271,17 +392,16 @@ impl KnowledgeEntity {
                 error!("{}", err_msg);
                 return Err(AppError::InternalError(err_msg));
             }
-            new_embeddings.insert(entity.id.clone(), embedding);
+            new_embeddings.insert(entity.id.clone(), (embedding, entity.user_id.clone()));
         }
         info!("Successfully generated all new embeddings.");
 
         // Perform DB updates in a single transaction
-        info!("Applying schema and data changes in a transaction...");
+        info!("Applying embedding updates in a transaction...");
         let mut transaction_query = String::from("BEGIN TRANSACTION;");
 
-        // Add all update statements
-        for (id, embedding) in new_embeddings {
-            // We must properly serialize the vector for the SurrealQL query string
+        // Add all update statements to the embedding table
+        for (id, (embedding, user_id)) in new_embeddings {
             let embedding_str = format!(
                 "[{}]",
                 embedding
@@ -291,18 +411,22 @@ impl KnowledgeEntity {
                     .join(",")
             );
             transaction_query.push_str(&format!(
-            "UPDATE type::thing('knowledge_entity', '{}') SET embedding = {}, updated_at = time::now();",
-            id, embedding_str
-        ));
+                "UPSERT type::thing('knowledge_entity_embedding', '{id}') SET \
+                    entity_id = type::thing('knowledge_entity', '{id}'), \
+                    embedding = {embedding}, \
+                    user_id = '{user_id}', \
+                    created_at = IF created_at != NONE THEN created_at ELSE time::now() END, \
+                    updated_at = time::now();",
+                id = id,
+                embedding = embedding_str,
+                user_id = user_id
+            ));
         }
 
-        // Re-create the index after updating the data that it will index
-        transaction_query
-            .push_str("REMOVE INDEX idx_embedding_entities ON TABLE knowledge_entity;");
         transaction_query.push_str(&format!(
-        "DEFINE INDEX idx_embedding_entities ON TABLE knowledge_entity FIELDS embedding HNSW DIMENSION {};",
-        new_dimensions
-    ));
+            "DEFINE INDEX OVERWRITE idx_embedding_knowledge_entity_embedding ON TABLE knowledge_entity_embedding FIELDS embedding HNSW DIMENSION {};",
+            new_dimensions
+        ));
 
         transaction_query.push_str("COMMIT TRANSACTION;");
 
@@ -312,12 +436,146 @@ impl KnowledgeEntity {
         info!("Re-embedding process for knowledge entities completed successfully.");
         Ok(())
     }
+
+    /// Re-creates embeddings for all knowledge entities using an `EmbeddingProvider`.
+    ///
+    /// This variant uses the application's configured embedding provider (FastEmbed, OpenAI, etc.)
+    /// instead of directly calling OpenAI. Used during startup when embedding configuration changes.
+    pub async fn update_all_embeddings_with_provider(
+        db: &SurrealDbClient,
+        provider: &crate::utils::embedding::EmbeddingProvider,
+    ) -> Result<(), AppError> {
+        let new_dimensions = provider.dimension();
+        info!(
+            dimensions = new_dimensions,
+            backend = provider.backend_label(),
+            "Starting re-embedding process for all knowledge entities"
+        );
+
+        // Fetch all entities first
+        let all_entities: Vec<KnowledgeEntity> = db.select(Self::table_name()).await?;
+        let total_entities = all_entities.len();
+        if total_entities == 0 {
+            info!("No knowledge entities to update. Just updating the index.");
+            KnowledgeEntityEmbedding::redefine_hnsw_index(db, new_dimensions).await?;
+            return Ok(());
+        }
+        info!(entities = total_entities, "Found entities to process");
+
+        // Generate all new embeddings in memory
+        let mut new_embeddings: HashMap<String, (Vec<f32>, String)> = HashMap::new();
+        info!("Generating new embeddings for all entities...");
+
+        for (i, entity) in all_entities.iter().enumerate() {
+            if i > 0 && i % 100 == 0 {
+                info!(
+                    progress = i,
+                    total = total_entities,
+                    "Re-embedding progress"
+                );
+            }
+
+            let embedding_input = format!(
+                "name: {}, description: {}, type: {:?}",
+                entity.name, entity.description, entity.entity_type
+            );
+
+            let embedding = provider
+                .embed(&embedding_input)
+                .await
+                .map_err(|e| AppError::InternalError(format!("Embedding failed: {e}")))?;
+
+            // Safety check: ensure the generated embedding has the correct dimension.
+            if embedding.len() != new_dimensions {
+                let err_msg = format!(
+                    "CRITICAL: Generated embedding for entity {} has incorrect dimension ({}). Expected {}. Aborting.",
+                    entity.id, embedding.len(), new_dimensions
+                );
+                error!("{}", err_msg);
+                return Err(AppError::InternalError(err_msg));
+            }
+            new_embeddings.insert(entity.id.clone(), (embedding, entity.user_id.clone()));
+        }
+        info!("Successfully generated all new embeddings.");
+        info!("Successfully generated all new embeddings.");
+
+        // Clear existing embeddings and index first to prevent SurrealDB panics and dimension conflicts.
+        info!("Removing old index and clearing embeddings...");
+
+        // Explicitly remove the index first. This prevents background HNSW maintenance from crashing
+        // when we delete/replace data, dealing with a known SurrealDB panic.
+        db.client
+            .query(format!(
+                "REMOVE INDEX idx_embedding_knowledge_entity_embedding ON TABLE {};",
+                KnowledgeEntityEmbedding::table_name()
+            ))
+            .await
+            .map_err(AppError::Database)?
+            .check()
+            .map_err(AppError::Database)?;
+
+        db.client
+            .query(format!(
+                "DELETE FROM {};",
+                KnowledgeEntityEmbedding::table_name()
+            ))
+            .await
+            .map_err(AppError::Database)?
+            .check()
+            .map_err(AppError::Database)?;
+
+        // Perform DB updates in a single transaction
+        info!("Applying embedding updates in a transaction...");
+        let mut transaction_query = String::from("BEGIN TRANSACTION;");
+
+        for (id, (embedding, user_id)) in new_embeddings {
+            let embedding_str = format!(
+                "[{}]",
+                embedding
+                    .iter()
+                    .map(|f| f.to_string())
+                    .collect::<Vec<_>>()
+                    .join(",")
+            );
+            transaction_query.push_str(&format!(
+                "CREATE type::thing('knowledge_entity_embedding', '{id}') SET \
+                    entity_id = type::thing('knowledge_entity', '{id}'), \
+                    embedding = {embedding}, \
+                    user_id = '{user_id}', \
+                    created_at = time::now(), \
+                    updated_at = time::now();",
+                id = id,
+                embedding = embedding_str,
+                user_id = user_id
+            ));
+        }
+
+        transaction_query.push_str(&format!(
+            "DEFINE INDEX OVERWRITE idx_embedding_knowledge_entity_embedding ON TABLE knowledge_entity_embedding FIELDS embedding HNSW DIMENSION {};",
+            new_dimensions
+        ));
+
+        transaction_query.push_str("COMMIT TRANSACTION;");
+
+        // Execute the entire atomic operation
+        db.client
+            .query(transaction_query)
+            .await
+            .map_err(AppError::Database)?
+            .check()
+            .map_err(AppError::Database)?;
+
+        info!("Re-embedding process for knowledge entities completed successfully.");
+        Ok(())
+    }
 }
 
 #[cfg(test)]
 mod tests {
     use super::*;
+    use crate::storage::types::knowledge_entity_embedding::KnowledgeEntityEmbedding;
     use serde_json::json;
+    use uuid::Uuid;
 
     #[tokio::test]
     async fn test_knowledge_entity_creation() {
@@ -327,7 +585,6 @@ mod tests {
         let description = "Test Description".to_string();
         let entity_type = KnowledgeEntityType::Document;
         let metadata = Some(json!({"key": "value"}));
-        let embedding = vec![0.1, 0.2, 0.3, 0.4, 0.5];
         let user_id = "user123".to_string();
 
         let entity = KnowledgeEntity::new(
@@ -336,7 +593,6 @@ mod tests {
             description.clone(),
             entity_type.clone(),
             metadata.clone(),
-            embedding.clone(),
             user_id.clone(),
         );
 
@@ -346,7 +602,6 @@ mod tests {
         assert_eq!(entity.description, description);
         assert_eq!(entity.entity_type, entity_type);
         assert_eq!(entity.metadata, metadata);
-        assert_eq!(entity.embedding, embedding);
         assert_eq!(entity.user_id, user_id);
         assert!(!entity.id.is_empty());
     }
@@ -410,20 +665,25 @@ mod tests {
         let db = SurrealDbClient::memory(namespace, database)
             .await
             .expect("Failed to start in-memory surrealdb");
+        db.apply_migrations()
+            .await
+            .expect("Failed to apply migrations");
 
         // Create two entities with the same source_id
         let source_id = "source123".to_string();
         let entity_type = KnowledgeEntityType::Document;
-        let embedding = vec![0.1, 0.2, 0.3, 0.4, 0.5];
         let user_id = "user123".to_string();
 
+        KnowledgeEntityEmbedding::redefine_hnsw_index(&db, 5)
+            .await
+            .expect("Failed to redefine index length");
+
         let entity1 = KnowledgeEntity::new(
             source_id.clone(),
             "Entity 1".to_string(),
             "Description 1".to_string(),
             entity_type.clone(),
             None,
-            embedding.clone(),
             user_id.clone(),
         );
 
@@ -433,7 +693,6 @@ mod tests {
             "Description 2".to_string(),
             entity_type.clone(),
             None,
-            embedding.clone(),
             user_id.clone(),
         );
 
@@ -445,18 +704,18 @@ mod tests {
             "Different Description".to_string(),
             entity_type.clone(),
             None,
-            embedding.clone(),
             user_id.clone(),
         );
 
+        let emb = vec![0.1, 0.2, 0.3, 0.4, 0.5];
         // Store the entities
-        db.store_item(entity1)
+        KnowledgeEntity::store_with_embedding(entity1.clone(), emb.clone(), &db)
             .await
             .expect("Failed to store entity 1");
-        db.store_item(entity2)
+        KnowledgeEntity::store_with_embedding(entity2.clone(), emb.clone(), &db)
             .await
             .expect("Failed to store entity 2");
-        db.store_item(different_entity.clone())
+        KnowledgeEntity::store_with_embedding(different_entity.clone(), emb.clone(), &db)
             .await
             .expect("Failed to store different entity");
 
@@ -505,6 +764,271 @@ mod tests {
         assert_eq!(different_remaining[0].id, different_entity.id);
     }
 
-    // Note: We can't easily test the patch method without mocking the OpenAI client
-    // and the generate_embedding function. This would require more complex setup.
+    #[tokio::test]
+    async fn test_delete_by_source_id_resists_query_injection() {
+        let namespace = "test_ns";
+        let database = &Uuid::new_v4().to_string();
+        let db = SurrealDbClient::memory(namespace, database)
+            .await
+            .expect("Failed to start in-memory surrealdb");
+        db.apply_migrations()
+            .await
+            .expect("Failed to apply migrations");
+
+        KnowledgeEntityEmbedding::redefine_hnsw_index(&db, 3)
+            .await
+            .expect("Failed to redefine index length");
+
+        let user_id = "user123".to_string();
+
+        let entity1 = KnowledgeEntity::new(
+            "safe_source".to_string(),
+            "Entity 1".to_string(),
+            "Description 1".to_string(),
+            KnowledgeEntityType::Document,
+            None,
+            user_id.clone(),
+        );
+
+        let entity2 = KnowledgeEntity::new(
+            "other_source".to_string(),
+            "Entity 2".to_string(),
+            "Description 2".to_string(),
+            KnowledgeEntityType::Document,
+            None,
+            user_id,
+        );
+
+        KnowledgeEntity::store_with_embedding(entity1, vec![0.1, 0.2, 0.3], &db)
+            .await
+            .expect("store entity1");
+        KnowledgeEntity::store_with_embedding(entity2, vec![0.3, 0.2, 0.1], &db)
+            .await
+            .expect("store entity2");
+
+        let malicious_source = "safe_source' OR 1=1 --";
+        KnowledgeEntity::delete_by_source_id(malicious_source, &db)
+            .await
+            .expect("delete call should succeed");
+
+        let remaining: Vec<KnowledgeEntity> = db
+            .client
+            .query("SELECT * FROM type::table($table)")
+            .bind(("table", KnowledgeEntity::table_name()))
+            .await
+            .expect("query failed")
+            .take(0)
+            .expect("take failed");
+
+        assert_eq!(
+            remaining.len(),
+            2,
+            "malicious input must not delete unrelated entities"
+        );
+    }
+
+    #[tokio::test]
+    async fn test_vector_search_returns_empty_when_no_embeddings() {
+        let namespace = "test_ns";
+        let database = &Uuid::new_v4().to_string();
+        let db = SurrealDbClient::memory(namespace, database)
+            .await
+            .expect("Failed to start in-memory surrealdb");
+        db.apply_migrations()
+            .await
+            .expect("Failed to apply migrations");
+
+        KnowledgeEntityEmbedding::redefine_hnsw_index(&db, 3)
+            .await
+            .expect("Failed to redefine index length");
+
+        let results = KnowledgeEntity::vector_search(5, vec![0.1, 0.2, 0.3], &db, "user")
+            .await
+            .expect("vector search");
+        assert!(results.is_empty());
+    }
+
+    #[tokio::test]
+    async fn test_vector_search_single_result() {
+        let namespace = "test_ns";
+        let database = &Uuid::new_v4().to_string();
+        let db = SurrealDbClient::memory(namespace, database)
+            .await
+            .expect("Failed to start in-memory surrealdb");
+        db.apply_migrations()
+            .await
+            .expect("Failed to apply migrations");
+
+        KnowledgeEntityEmbedding::redefine_hnsw_index(&db, 3)
+            .await
+            .expect("Failed to redefine index length");
+
+        let user_id = "user".to_string();
+        let source_id = "src".to_string();
+        let entity = KnowledgeEntity::new(
+            source_id.clone(),
+            "hello".to_string(),
+            "world".to_string(),
+            KnowledgeEntityType::Document,
+            None,
+            user_id.clone(),
+        );
+
+        KnowledgeEntity::store_with_embedding(entity.clone(), vec![0.1, 0.2, 0.3], &db)
+            .await
+            .expect("store entity with embedding");
+
+        let stored_entity: Option<KnowledgeEntity> = db.get_item(&entity.id).await.unwrap();
+        assert!(stored_entity.is_some());
+
+        let stored_embeddings: Vec<KnowledgeEntityEmbedding> = db
+            .client
+            .query(format!(
+                "SELECT * FROM {}",
+                KnowledgeEntityEmbedding::table_name()
+            ))
+            .await
+            .expect("query embeddings")
+            .take(0)
+            .expect("take embeddings");
+        assert_eq!(stored_embeddings.len(), 1);
+
+        let rid = surrealdb::RecordId::from_table_key(KnowledgeEntity::table_name(), &entity.id);
+        let fetched_emb = KnowledgeEntityEmbedding::get_by_entity_id(&rid, &db)
+            .await
+            .expect("fetch embedding");
+        assert!(fetched_emb.is_some());
+
+        let results = KnowledgeEntity::vector_search(3, vec![0.1, 0.2, 0.3], &db, &user_id)
+            .await
+            .expect("vector search");
+
+        assert_eq!(results.len(), 1);
+        let res = &results[0];
+        assert_eq!(res.entity.id, entity.id);
+        assert_eq!(res.entity.source_id, source_id);
+        assert_eq!(res.entity.name, "hello");
+    }
+
+    #[tokio::test]
+    async fn test_vector_search_orders_by_similarity() {
+        let namespace = "test_ns";
+        let database = &Uuid::new_v4().to_string();
+        let db = SurrealDbClient::memory(namespace, database)
+            .await
+            .expect("Failed to start in-memory surrealdb");
+        db.apply_migrations()
+            .await
+            .expect("Failed to apply migrations");
+
+        KnowledgeEntityEmbedding::redefine_hnsw_index(&db, 3)
+            .await
+            .expect("Failed to redefine index length");
+
+        let user_id = "user".to_string();
+        let e1 = KnowledgeEntity::new(
+            "s1".to_string(),
+            "entity one".to_string(),
+            "desc".to_string(),
+            KnowledgeEntityType::Document,
+            None,
+            user_id.clone(),
+        );
+        let e2 = KnowledgeEntity::new(
+            "s2".to_string(),
+            "entity two".to_string(),
+            "desc".to_string(),
+            KnowledgeEntityType::Document,
+            None,
+            user_id.clone(),
+        );
+
+        KnowledgeEntity::store_with_embedding(e1.clone(), vec![1.0, 0.0, 0.0], &db)
+            .await
+            .expect("store e1");
+        KnowledgeEntity::store_with_embedding(e2.clone(), vec![0.0, 1.0, 0.0], &db)
+            .await
+            .expect("store e2");
+
+        let stored_e1: Option<KnowledgeEntity> = db.get_item(&e1.id).await.unwrap();
+        let stored_e2: Option<KnowledgeEntity> = db.get_item(&e2.id).await.unwrap();
+        assert!(stored_e1.is_some() && stored_e2.is_some());
+
+        let stored_embeddings: Vec<KnowledgeEntityEmbedding> = db
+            .client
+            .query(format!(
+                "SELECT * FROM {}",
+                KnowledgeEntityEmbedding::table_name()
+            ))
+            .await
+            .expect("query embeddings")
+            .take(0)
+            .expect("take embeddings");
+        assert_eq!(stored_embeddings.len(), 2);
+
+        let rid_e1 = surrealdb::RecordId::from_table_key(KnowledgeEntity::table_name(), &e1.id);
+        let rid_e2 = surrealdb::RecordId::from_table_key(KnowledgeEntity::table_name(), &e2.id);
+        assert!(KnowledgeEntityEmbedding::get_by_entity_id(&rid_e1, &db)
+            .await
+            .unwrap()
+            .is_some());
+        assert!(KnowledgeEntityEmbedding::get_by_entity_id(&rid_e2, &db)
+            .await
+            .unwrap()
+            .is_some());
+
+        let results = KnowledgeEntity::vector_search(2, vec![0.0, 1.0, 0.0], &db, &user_id)
+            .await
+            .expect("vector search");
+
+        assert_eq!(results.len(), 2);
+        assert_eq!(results[0].entity.id, e2.id);
+        assert_eq!(results[1].entity.id, e1.id);
+    }
+
+    #[tokio::test]
+    async fn test_vector_search_with_orphaned_embedding() {
+        let namespace = "test_ns_orphan";
+        let database = &Uuid::new_v4().to_string();
+        let db = SurrealDbClient::memory(namespace, database)
+            .await
+            .expect("Failed to start in-memory surrealdb");
+        db.apply_migrations()
+            .await
+            .expect("Failed to apply migrations");
+
+        KnowledgeEntityEmbedding::redefine_hnsw_index(&db, 3)
+            .await
+            .expect("Failed to redefine index length");
+
+        let user_id = "user".to_string();
+        let source_id = "src".to_string();
+        let entity = KnowledgeEntity::new(
+            source_id.clone(),
+            "orphan".to_string(),
+            "orphan desc".to_string(),
+            KnowledgeEntityType::Document,
+            None,
+            user_id.clone(),
+        );
+
+        KnowledgeEntity::store_with_embedding(entity.clone(), vec![0.1, 0.2, 0.3], &db)
+            .await
+            .expect("store entity with embedding");
+
+        // Manually delete the entity to create an orphan
+        let query = format!("DELETE type::thing('knowledge_entity', '{}')", entity.id);
+        db.client.query(query).await.expect("delete entity");
+
+        // Now search
+        let results = KnowledgeEntity::vector_search(3, vec![0.1, 0.2, 0.3], &db, &user_id)
+            .await
+            .expect("search should succeed even with orphans");
+
+        assert!(
+            results.is_empty(),
+            "Should return empty result for orphan, got: {:?}",
+            results
+        );
+    }
 }

common/src/storage/types/knowledge_entity_embedding.rs

@@ -0,0 +1,387 @@
+use std::collections::HashMap;
+
+use surrealdb::RecordId;
+
+use crate::{error::AppError, storage::db::SurrealDbClient, stored_object};
+
+stored_object!(KnowledgeEntityEmbedding, "knowledge_entity_embedding", {
+    entity_id: RecordId,
+    embedding: Vec<f32>,
+    /// Denormalized user id for query scoping
+    user_id: String
+});
+
+impl KnowledgeEntityEmbedding {
+    /// Recreate the HNSW index with a new embedding dimension.
+    pub async fn redefine_hnsw_index(
+        db: &SurrealDbClient,
+        dimension: usize,
+    ) -> Result<(), AppError> {
+        let query = format!(
+            "BEGIN TRANSACTION;
+             REMOVE INDEX IF EXISTS idx_embedding_knowledge_entity_embedding ON TABLE {table};
+             DEFINE INDEX idx_embedding_knowledge_entity_embedding ON TABLE {table} FIELDS embedding HNSW DIMENSION {dimension};
+             COMMIT TRANSACTION;",
+            table = Self::table_name(),
+        );
+
+        let res = db.client.query(query).await.map_err(AppError::Database)?;
+        res.check().map_err(AppError::Database)?;
+
+        Ok(())
+    }
+
+    /// Create a new knowledge entity embedding
+    pub fn new(entity_id: &str, embedding: Vec<f32>, user_id: String) -> Self {
+        let now = Utc::now();
+        Self {
+            id: uuid::Uuid::new_v4().to_string(),
+            created_at: now,
+            updated_at: now,
+            entity_id: RecordId::from_table_key("knowledge_entity", entity_id),
+            embedding,
+            user_id,
+        }
+    }
+
+    /// Get embedding by entity ID
+    pub async fn get_by_entity_id(
+        entity_id: &RecordId,
+        db: &SurrealDbClient,
+    ) -> Result<Option<Self>, AppError> {
+        let query = format!(
+            "SELECT * FROM {} WHERE entity_id = $entity_id LIMIT 1",
+            Self::table_name()
+        );
+        let mut result = db
+            .client
+            .query(query)
+            .bind(("entity_id", entity_id.clone()))
+            .await
+            .map_err(AppError::Database)?;
+        let embeddings: Vec<Self> = result.take(0).map_err(AppError::Database)?;
+        Ok(embeddings.into_iter().next())
+    }
+
+    /// Get embeddings for multiple entities in batch
+    pub async fn get_by_entity_ids(
+        entity_ids: &[RecordId],
+        db: &SurrealDbClient,
+    ) -> Result<HashMap<String, Vec<f32>>, AppError> {
+        if entity_ids.is_empty() {
+            return Ok(HashMap::new());
+        }
+
+        let ids_list: Vec<RecordId> = entity_ids.to_vec();
+
+        let query = format!(
+            "SELECT * FROM {} WHERE entity_id INSIDE $entity_ids",
+            Self::table_name()
+        );
+        let mut result = db
+            .client
+            .query(query)
+            .bind(("entity_ids", ids_list))
+            .await
+            .map_err(AppError::Database)?;
+        let embeddings: Vec<Self> = result.take(0).map_err(AppError::Database)?;
+
+        Ok(embeddings
+            .into_iter()
+            .map(|e| (e.entity_id.key().to_string(), e.embedding))
+            .collect())
+    }
+
+    /// Delete embedding by entity ID
+    pub async fn delete_by_entity_id(
+        entity_id: &RecordId,
+        db: &SurrealDbClient,
+    ) -> Result<(), AppError> {
+        let query = format!(
+            "DELETE FROM {} WHERE entity_id = $entity_id",
+            Self::table_name()
+        );
+        db.client
+            .query(query)
+            .bind(("entity_id", entity_id.clone()))
+            .await
+            .map_err(AppError::Database)?;
+        Ok(())
+    }
+
+    /// Delete embeddings by source_id (via joining to knowledge_entity table)
+    #[allow(clippy::items_after_statements)]
+    pub async fn delete_by_source_id(
+        source_id: &str,
+        db: &SurrealDbClient,
+    ) -> Result<(), AppError> {
+        let query = "SELECT id FROM knowledge_entity WHERE source_id = $source_id";
+        let mut res = db
+            .client
+            .query(query)
+            .bind(("source_id", source_id.to_owned()))
+            .await
+            .map_err(AppError::Database)?;
+        #[allow(clippy::missing_docs_in_private_items)]
+        #[derive(Deserialize)]
+        struct IdRow {
+            id: RecordId,
+        }
+        let ids: Vec<IdRow> = res.take(0).map_err(AppError::Database)?;
+
+        for row in ids {
+            Self::delete_by_entity_id(&row.id, db).await?;
+        }
+        Ok(())
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::storage::db::SurrealDbClient;
+    use crate::storage::types::knowledge_entity::{KnowledgeEntity, KnowledgeEntityType};
+    use chrono::Utc;
+    use surrealdb::Value as SurrealValue;
+    use uuid::Uuid;
+
+    async fn setup_test_db() -> SurrealDbClient {
+        let namespace = "test_ns";
+        let database = Uuid::new_v4().to_string();
+        let db = SurrealDbClient::memory(namespace, &database)
+            .await
+            .expect("Failed to start in-memory surrealdb");
+
+        db.apply_migrations()
+            .await
+            .expect("Failed to apply migrations");
+
+        db
+    }
+
+    fn build_knowledge_entity_with_id(
+        key: &str,
+        source_id: &str,
+        user_id: &str,
+    ) -> KnowledgeEntity {
+        KnowledgeEntity {
+            id: key.to_owned(),
+            created_at: Utc::now(),
+            updated_at: Utc::now(),
+            source_id: source_id.to_owned(),
+            name: "Test entity".to_owned(),
+            description: "Desc".to_owned(),
+            entity_type: KnowledgeEntityType::Document,
+            metadata: None,
+            user_id: user_id.to_owned(),
+        }
+    }
+
+    #[tokio::test]
+    async fn test_create_and_get_by_entity_id() {
+        let db = setup_test_db().await;
+        KnowledgeEntityEmbedding::redefine_hnsw_index(&db, 3)
+            .await
+            .expect("set test index dimension");
+        let user_id = "user_ke";
+        let entity_key = "entity-1";
+        let source_id = "source-ke";
+
+        let embedding_vec = vec![0.11_f32, 0.22, 0.33];
+        let entity = build_knowledge_entity_with_id(entity_key, source_id, user_id);
+
+        KnowledgeEntity::store_with_embedding(entity.clone(), embedding_vec.clone(), &db)
+            .await
+            .expect("Failed to store entity with embedding");
+
+        let entity_rid = RecordId::from_table_key(KnowledgeEntity::table_name(), &entity.id);
+
+        let fetched = KnowledgeEntityEmbedding::get_by_entity_id(&entity_rid, &db)
+            .await
+            .expect("Failed to get embedding by entity_id")
+            .expect("Expected embedding to exist");
+
+        assert_eq!(fetched.user_id, user_id);
+        assert_eq!(fetched.entity_id, entity_rid);
+        assert_eq!(fetched.embedding, embedding_vec);
+    }
+
+    #[tokio::test]
+    async fn test_delete_by_entity_id() {
+        let db = setup_test_db().await;
+        KnowledgeEntityEmbedding::redefine_hnsw_index(&db, 3)
+            .await
+            .expect("set test index dimension");
+        let user_id = "user_ke";
+        let entity_key = "entity-delete";
+        let source_id = "source-del";
+
+        let entity = build_knowledge_entity_with_id(entity_key, source_id, user_id);
+
+        KnowledgeEntity::store_with_embedding(entity.clone(), vec![0.5_f32, 0.6, 0.7], &db)
+            .await
+            .expect("Failed to store entity with embedding");
+
+        let entity_rid = RecordId::from_table_key(KnowledgeEntity::table_name(), &entity.id);
+
+        let existing = KnowledgeEntityEmbedding::get_by_entity_id(&entity_rid, &db)
+            .await
+            .expect("Failed to get embedding before delete");
+        assert!(existing.is_some());
+
+        KnowledgeEntityEmbedding::delete_by_entity_id(&entity_rid, &db)
+            .await
+            .expect("Failed to delete by entity_id");
+
+        let after = KnowledgeEntityEmbedding::get_by_entity_id(&entity_rid, &db)
+            .await
+            .expect("Failed to get embedding after delete");
+        assert!(after.is_none());
+    }
+
+    #[tokio::test]
+    async fn test_store_with_embedding_creates_entity_and_embedding() {
+        let db = setup_test_db().await;
+        let user_id = "user_store";
+        let source_id = "source_store";
+        let embedding = vec![0.2_f32, 0.3, 0.4];
+
+        KnowledgeEntityEmbedding::redefine_hnsw_index(&db, embedding.len())
+            .await
+            .expect("set test index dimension");
+
+        let entity = build_knowledge_entity_with_id("entity-store", source_id, user_id);
+
+        KnowledgeEntity::store_with_embedding(entity.clone(), embedding.clone(), &db)
+            .await
+            .expect("Failed to store entity with embedding");
+
+        let stored_entity: Option<KnowledgeEntity> = db.get_item(&entity.id).await.unwrap();
+        assert!(stored_entity.is_some());
+
+        let entity_rid = RecordId::from_table_key(KnowledgeEntity::table_name(), &entity.id);
+        let stored_embedding = KnowledgeEntityEmbedding::get_by_entity_id(&entity_rid, &db)
+            .await
+            .expect("Failed to fetch embedding");
+        assert!(stored_embedding.is_some());
+        let stored_embedding = stored_embedding.unwrap();
+        assert_eq!(stored_embedding.user_id, user_id);
+        assert_eq!(stored_embedding.entity_id, entity_rid);
+    }
+
+    #[tokio::test]
+    async fn test_delete_by_source_id() {
+        let db = setup_test_db().await;
+        KnowledgeEntityEmbedding::redefine_hnsw_index(&db, 3)
+            .await
+            .expect("set test index dimension");
+        let user_id = "user_ke";
+        let source_id = "shared-ke";
+        let other_source = "other-ke";
+
+        let entity1 = build_knowledge_entity_with_id("entity-s1", source_id, user_id);
+        let entity2 = build_knowledge_entity_with_id("entity-s2", source_id, user_id);
+        let entity_other = build_knowledge_entity_with_id("entity-other", other_source, user_id);
+
+        KnowledgeEntity::store_with_embedding(entity1.clone(), vec![1.0_f32, 1.1, 1.2], &db)
+            .await
+            .expect("Failed to store entity with embedding");
+        KnowledgeEntity::store_with_embedding(entity2.clone(), vec![2.0_f32, 2.1, 2.2], &db)
+            .await
+            .expect("Failed to store entity with embedding");
+        KnowledgeEntity::store_with_embedding(entity_other.clone(), vec![3.0_f32, 3.1, 3.2], &db)
+            .await
+            .expect("Failed to store entity with embedding");
+
+        let entity1_rid = RecordId::from_table_key(KnowledgeEntity::table_name(), &entity1.id);
+        let entity2_rid = RecordId::from_table_key(KnowledgeEntity::table_name(), &entity2.id);
+        let other_rid = RecordId::from_table_key(KnowledgeEntity::table_name(), &entity_other.id);
+
+        KnowledgeEntityEmbedding::delete_by_source_id(source_id, &db)
+            .await
+            .expect("Failed to delete by source_id");
+
+        assert!(
+            KnowledgeEntityEmbedding::get_by_entity_id(&entity1_rid, &db)
+                .await
+                .unwrap()
+                .is_none()
+        );
+        assert!(
+            KnowledgeEntityEmbedding::get_by_entity_id(&entity2_rid, &db)
+                .await
+                .unwrap()
+                .is_none()
+        );
+        assert!(KnowledgeEntityEmbedding::get_by_entity_id(&other_rid, &db)
+            .await
+            .unwrap()
+            .is_some());
+    }
+
+    #[tokio::test]
+    async fn test_redefine_hnsw_index_updates_dimension() {
+        let db = setup_test_db().await;
+
+        KnowledgeEntityEmbedding::redefine_hnsw_index(&db, 16)
+            .await
+            .expect("failed to redefine index");
+
+        let mut info_res = db
+            .client
+            .query("INFO FOR TABLE knowledge_entity_embedding;")
+            .await
+            .expect("info query failed");
+        let info: SurrealValue = info_res.take(0).expect("failed to take info result");
+        let info_json: serde_json::Value =
+            serde_json::to_value(info).expect("failed to convert info to json");
+        let idx_sql = info_json["Object"]["indexes"]["Object"]
+            ["idx_embedding_knowledge_entity_embedding"]["Strand"]
+            .as_str()
+            .unwrap_or_default();
+
+        assert!(
+            idx_sql.contains("DIMENSION 16"),
+            "expected index definition to contain new dimension, got: {idx_sql}"
+        );
+    }
+
+    #[tokio::test]
+    async fn test_fetch_entity_via_record_id() {
+        let db = setup_test_db().await;
+        KnowledgeEntityEmbedding::redefine_hnsw_index(&db, 3)
+            .await
+            .expect("set test index dimension");
+        let user_id = "user_ke";
+        let entity_key = "entity-fetch";
+        let source_id = "source-fetch";
+
+        let entity = build_knowledge_entity_with_id(entity_key, source_id, user_id);
+        KnowledgeEntity::store_with_embedding(entity.clone(), vec![0.7_f32, 0.8, 0.9], &db)
+            .await
+            .expect("Failed to store entity with embedding");
+
+        let entity_rid = RecordId::from_table_key(KnowledgeEntity::table_name(), &entity.id);
+
+        #[derive(Deserialize)]
+        struct Row {
+            entity_id: KnowledgeEntity,
+        }
+
+        let mut res = db
+            .client
+            .query(
+                "SELECT entity_id FROM knowledge_entity_embedding WHERE entity_id = $id FETCH entity_id;",
+            )
+            .bind(("id", entity_rid.clone()))
+            .await
+            .expect("failed to fetch embedding with FETCH");
+        let rows: Vec<Row> = res.take(0).expect("failed to deserialize fetch rows");
+
+        assert_eq!(rows.len(), 1);
+        let fetched_entity = &rows[0].entity_id;
+        assert_eq!(fetched_entity.id, entity_key);
+        assert_eq!(fetched_entity.name, "Test entity");
+        assert_eq!(fetched_entity.user_id, user_id);
+    }
+}

common/src/storage/types/knowledge_relationship.rs

@@ -40,21 +40,28 @@ impl KnowledgeRelationship {
         }
     }
     pub async fn store_relationship(&self, db_client: &SurrealDbClient) -> Result<(), AppError> {
-        let query = format!(
-            r#"RELATE knowledge_entity:`{}`->relates_to:`{}`->knowledge_entity:`{}`
-            SET
-                metadata.user_id = '{}',
-                metadata.source_id = '{}',
-                metadata.relationship_type = '{}'"#,
-            self.in_,
-            self.id,
-            self.out,
-            self.metadata.user_id,
-            self.metadata.source_id,
-            self.metadata.relationship_type
-        );
-
-        db_client.query(query).await?;
+        db_client
+            .client
+            .query(
+                r#"BEGIN TRANSACTION;
+                LET $in_entity = type::thing('knowledge_entity', $in_id);
+                LET $out_entity = type::thing('knowledge_entity', $out_id);
+                LET $relation = type::thing('relates_to', $rel_id);
+                DELETE type::thing('relates_to', $rel_id);
+                RELATE $in_entity->$relation->$out_entity SET
+                    metadata.user_id = $user_id,
+                    metadata.source_id = $source_id,
+                    metadata.relationship_type = $relationship_type;
+                COMMIT TRANSACTION;"#,
+            )
+            .bind(("rel_id", self.id.clone()))
+            .bind(("in_id", self.in_.clone()))
+            .bind(("out_id", self.out.clone()))
+            .bind(("user_id", self.metadata.user_id.clone()))
+            .bind(("source_id", self.metadata.source_id.clone()))
+            .bind(("relationship_type", self.metadata.relationship_type.clone()))
+            .await?
+            .check()?;
 
         Ok(())
     }
@@ -63,12 +70,12 @@ impl KnowledgeRelationship {
         source_id: &str,
         db_client: &SurrealDbClient,
     ) -> Result<(), AppError> {
-        let query = format!(
-            "DELETE knowledge_entity -> relates_to WHERE metadata.source_id = '{}'",
-            source_id
-        );
-
-        db_client.query(query).await?;
+        db_client
+            .client
+            .query("DELETE FROM relates_to WHERE metadata.source_id = $source_id")
+            .bind(("source_id", source_id.to_owned()))
+            .await?
+            .check()?;
 
         Ok(())
     }
@@ -79,16 +86,20 @@ impl KnowledgeRelationship {
         db_client: &SurrealDbClient,
     ) -> Result<(), AppError> {
         let mut authorized_result = db_client
-            .query(format!(
-                "SELECT * FROM relates_to WHERE id = relates_to:`{}` AND metadata.user_id = '{}'",
-                id, user_id
-            ))
+            .client
+            .query(
+                "SELECT * FROM relates_to WHERE id = type::thing('relates_to', $id) AND metadata.user_id = $user_id",
+            )
+            .bind(("id", id.to_owned()))
+            .bind(("user_id", user_id.to_owned()))
             .await?;
         let authorized: Vec<KnowledgeRelationship> = authorized_result.take(0).unwrap_or_default();
 
         if authorized.is_empty() {
             let mut exists_result = db_client
-                .query(format!("SELECT * FROM relates_to:`{}`", id))
+                .client
+                .query("SELECT * FROM type::thing('relates_to', $id)")
+                .bind(("id", id.to_owned()))
                 .await?;
             let existing: Option<KnowledgeRelationship> = exists_result.take(0)?;
 
@@ -97,12 +108,15 @@ impl KnowledgeRelationship {
                     "Not authorized to delete relationship".into(),
                 ))
             } else {
-                Err(AppError::NotFound(format!("Relationship {} not found", id)))
+                Err(AppError::NotFound(format!("Relationship {id} not found")))
             }
         } else {
             db_client
-                .query(format!("DELETE relates_to:`{}`", id))
-                .await?;
+                .client
+                .query("DELETE type::thing('relates_to', $id)")
+                .bind(("id", id.to_owned()))
+                .await?
+                .check()?;
             Ok(())
         }
     }
@@ -113,12 +127,39 @@ mod tests {
     use super::*;
     use crate::storage::types::knowledge_entity::{KnowledgeEntity, KnowledgeEntityType};
 
+    async fn setup_test_db() -> SurrealDbClient {
+        let namespace = "test_ns";
+        let database = &Uuid::new_v4().to_string();
+        let db = SurrealDbClient::memory(namespace, database)
+            .await
+            .expect("Failed to start in-memory surrealdb");
+
+        db.apply_migrations()
+            .await
+            .expect("Failed to apply migrations");
+
+        db
+    }
+
+    async fn get_relationship_by_id(
+        relationship_id: &str,
+        db_client: &SurrealDbClient,
+    ) -> Option<KnowledgeRelationship> {
+        let mut result = db_client
+            .client
+            .query("SELECT * FROM type::thing('relates_to', $id)")
+            .bind(("id", relationship_id.to_owned()))
+            .await
+            .expect("relationship query by id failed");
+
+        result.take(0).expect("failed to take relationship by id")
+    }
+
     // Helper function to create a test knowledge entity for the relationship tests
     async fn create_test_entity(name: &str, db_client: &SurrealDbClient) -> String {
         let source_id = "source123".to_string();
         let description = format!("Description for {}", name);
         let entity_type = KnowledgeEntityType::Document;
-        let embedding = vec![0.1, 0.2, 0.3];
         let user_id = "user123".to_string();
 
         let entity = KnowledgeEntity::new(
@@ -127,7 +168,6 @@ mod tests {
             description,
             entity_type,
             None,
-            embedding,
             user_id,
         );
 
@@ -164,13 +204,9 @@ mod tests {
     }
 
     #[tokio::test]
-    async fn test_store_relationship() {
+    async fn test_store_and_verify_by_source_id() {
         // Setup in-memory database for testing
-        let namespace = "test_ns";
-        let database = &Uuid::new_v4().to_string();
-        let db = SurrealDbClient::memory(namespace, database)
-            .await
-            .expect("Failed to start in-memory surrealdb");
+        let db = setup_test_db().await;
 
         // Create two entities to relate
         let entity1_id = create_test_entity("Entity 1", &db).await;
@@ -195,30 +231,69 @@ mod tests {
             .await
             .expect("Failed to store relationship");
 
+        let persisted = get_relationship_by_id(&relationship.id, &db)
+            .await
+            .expect("Relationship should be retrievable by id");
+        assert_eq!(persisted.in_, entity1_id);
+        assert_eq!(persisted.out, entity2_id);
+        assert_eq!(persisted.metadata.user_id, user_id);
+        assert_eq!(persisted.metadata.source_id, source_id);
+
         // Query to verify the relationship exists by checking for relationships with our source_id
         // This approach is more reliable than trying to look up by ID
-        let check_query = format!(
-            "SELECT * FROM relates_to WHERE metadata.source_id = '{}'",
-            source_id
-        );
-        let mut check_result = db.query(check_query).await.expect("Check query failed");
+        let mut check_result = db
+            .query("SELECT * FROM relates_to WHERE metadata.source_id = $source_id")
+            .bind(("source_id", source_id.clone()))
+            .await
+            .expect("Check query failed");
         let check_results: Vec<KnowledgeRelationship> = check_result.take(0).unwrap_or_default();
 
-        // Just verify that a relationship was created
-        assert!(
-            !check_results.is_empty(),
-            "Relationship should exist in the database"
+        assert_eq!(
+            check_results.len(),
+            1,
+            "Expected one relationship for source_id"
         );
     }
 
     #[tokio::test]
-    async fn test_delete_relationship_by_id() {
-        // Setup in-memory database for testing
-        let namespace = "test_ns";
-        let database = &Uuid::new_v4().to_string();
-        let db = SurrealDbClient::memory(namespace, database)
+    async fn test_store_relationship_resists_query_injection() {
+        let db = setup_test_db().await;
+
+        let entity1_id = create_test_entity("Entity 1", &db).await;
+        let entity2_id = create_test_entity("Entity 2", &db).await;
+
+        let relationship = KnowledgeRelationship::new(
+            entity1_id,
+            entity2_id,
+            "user'123".to_string(),
+            "source123'; DELETE FROM relates_to; --".to_string(),
+            "references'; UPDATE user SET admin = true; --".to_string(),
+        );
+
+        relationship
+            .store_relationship(&db)
             .await
-            .expect("Failed to start in-memory surrealdb");
+            .expect("store relationship should safely handle quote-containing values");
+
+        let mut res = db
+            .client
+            .query("SELECT * FROM relates_to WHERE id = type::thing('relates_to', $id)")
+            .bind(("id", relationship.id.clone()))
+            .await
+            .expect("query relationship by id failed");
+        let rows: Vec<KnowledgeRelationship> = res.take(0).expect("take rows");
+
+        assert_eq!(rows.len(), 1);
+        assert_eq!(
+            rows[0].metadata.source_id,
+            "source123'; DELETE FROM relates_to; --"
+        );
+    }
+
+    #[tokio::test]
+    async fn test_store_and_delete_relationship() {
+        // Setup in-memory database for testing
+        let db = setup_test_db().await;
 
         // Create two entities to relate
         let entity1_id = create_test_entity("Entity 1", &db).await;
@@ -237,7 +312,7 @@ mod tests {
             relationship_type,
         );
 
-        // Store the relationship
+        // Store relationship
         relationship
             .store_relationship(&db)
             .await
@@ -258,12 +333,12 @@ mod tests {
             "Relationship should exist before deletion"
         );
 
-        // Delete the relationship by ID
+        // Delete relationship by ID
         KnowledgeRelationship::delete_relationship_by_id(&relationship.id, &user_id, &db)
             .await
             .expect("Failed to delete relationship by ID");
 
-        // Query to verify the relationship was deleted
+        // Query to verify relationship was deleted
         let mut result = db
             .query(format!(
                 "SELECT * FROM relates_to WHERE metadata.user_id = '{}' AND metadata.source_id = '{}'",
@@ -273,17 +348,13 @@ mod tests {
             .expect("Query failed");
         let results: Vec<KnowledgeRelationship> = result.take(0).unwrap_or_default();
 
-        // Verify the relationship no longer exists
+        // Verify relationship no longer exists
         assert!(results.is_empty(), "Relationship should be deleted");
     }
 
     #[tokio::test]
     async fn test_delete_relationship_by_id_unauthorized() {
-        let namespace = "test_ns";
-        let database = &Uuid::new_v4().to_string();
-        let db = SurrealDbClient::memory(namespace, database)
-            .await
-            .expect("Failed to start in-memory surrealdb");
+        let db = setup_test_db().await;
 
         let entity1_id = create_test_entity("Entity 1", &db).await;
         let entity2_id = create_test_entity("Entity 2", &db).await;
@@ -345,13 +416,9 @@ mod tests {
     }
 
     #[tokio::test]
-    async fn test_delete_relationships_by_source_id() {
+    async fn test_store_relationship_exists() {
         // Setup in-memory database for testing
-        let namespace = "test_ns";
-        let database = &Uuid::new_v4().to_string();
-        let db = SurrealDbClient::memory(namespace, database)
-            .await
-            .expect("Failed to start in-memory surrealdb");
+        let db = setup_test_db().await;
 
         // Create entities to relate
         let entity1_id = create_test_entity("Entity 1", &db).await;
@@ -403,49 +470,87 @@ mod tests {
             .await
             .expect("Failed to store different relationship");
 
+        // Sanity-check setup: exactly two relationships use source_id and one uses different_source_id.
+        let mut before_delete = db
+            .query("SELECT * FROM relates_to WHERE metadata.source_id = $source_id")
+            .bind(("source_id", source_id.clone()))
+            .await
+            .expect("before delete query failed");
+        let before_delete_rows: Vec<KnowledgeRelationship> =
+            before_delete.take(0).unwrap_or_default();
+        assert_eq!(before_delete_rows.len(), 2);
+
+        let mut before_delete_different = db
+            .query("SELECT * FROM relates_to WHERE metadata.source_id = $source_id")
+            .bind(("source_id", different_source_id.clone()))
+            .await
+            .expect("before delete different query failed");
+        let before_delete_different_rows: Vec<KnowledgeRelationship> =
+            before_delete_different.take(0).unwrap_or_default();
+        assert_eq!(before_delete_different_rows.len(), 1);
+
         // Delete relationships by source_id
         KnowledgeRelationship::delete_relationships_by_source_id(&source_id, &db)
             .await
             .expect("Failed to delete relationships by source_id");
 
-        // Query to verify the relationships with source_id were deleted
-        let query1 = format!("SELECT * FROM relates_to WHERE id = '{}'", relationship1.id);
-        let query2 = format!("SELECT * FROM relates_to WHERE id = '{}'", relationship2.id);
-        let different_query = format!(
-            "SELECT * FROM relates_to WHERE id = '{}'",
-            different_relationship.id
-        );
-
-        let mut result1 = db.query(query1).await.expect("Query 1 failed");
-        let results1: Vec<KnowledgeRelationship> = result1.take(0).unwrap_or_default();
-
-        let mut result2 = db.query(query2).await.expect("Query 2 failed");
-        let results2: Vec<KnowledgeRelationship> = result2.take(0).unwrap_or_default();
-
-        let mut different_result = db
-            .query(different_query)
-            .await
-            .expect("Different query failed");
-        let _different_results: Vec<KnowledgeRelationship> =
-            different_result.take(0).unwrap_or_default();
+        // Query to verify the specific relationships with source_id were deleted.
+        let result1 = get_relationship_by_id(&relationship1.id, &db).await;
+        let result2 = get_relationship_by_id(&relationship2.id, &db).await;
+        let different_result = get_relationship_by_id(&different_relationship.id, &db).await;
 
         // Verify relationships with the source_id are deleted
-        assert!(results1.is_empty(), "Relationship 1 should be deleted");
-        assert!(results2.is_empty(), "Relationship 2 should be deleted");
+        assert!(result1.is_none(), "Relationship 1 should be deleted");
+        assert!(result2.is_none(), "Relationship 2 should be deleted");
+        let remaining =
+            different_result.expect("Relationship with different source_id should remain");
+        assert_eq!(remaining.metadata.source_id, different_source_id);
+    }
 
-        // For the relationship with different source ID, we need to check differently
-        // Let's just verify we have a relationship where the source_id matches different_source_id
-        let check_query = format!(
-            "SELECT * FROM relates_to WHERE metadata.source_id = '{}'",
-            different_source_id
+    #[tokio::test]
+    async fn test_delete_relationships_by_source_id_resists_query_injection() {
+        let db = setup_test_db().await;
+
+        let entity1_id = create_test_entity("Entity 1", &db).await;
+        let entity2_id = create_test_entity("Entity 2", &db).await;
+        let entity3_id = create_test_entity("Entity 3", &db).await;
+
+        let safe_relationship = KnowledgeRelationship::new(
+            entity1_id.clone(),
+            entity2_id.clone(),
+            "user123".to_string(),
+            "safe_source".to_string(),
+            "references".to_string(),
         );
-        let mut check_result = db.query(check_query).await.expect("Check query failed");
-        let check_results: Vec<KnowledgeRelationship> = check_result.take(0).unwrap_or_default();
 
-        // Verify the relationship with a different source_id still exists
+        let other_relationship = KnowledgeRelationship::new(
+            entity2_id,
+            entity3_id,
+            "user123".to_string(),
+            "other_source".to_string(),
+            "contains".to_string(),
+        );
+
+        safe_relationship
+            .store_relationship(&db)
+            .await
+            .expect("store safe relationship");
+        other_relationship
+            .store_relationship(&db)
+            .await
+            .expect("store other relationship");
+
+        KnowledgeRelationship::delete_relationships_by_source_id("safe_source' OR 1=1 --", &db)
+            .await
+            .expect("delete call should succeed");
+
+        let remaining_safe = get_relationship_by_id(&safe_relationship.id, &db).await;
+        let remaining_other = get_relationship_by_id(&other_relationship.id, &db).await;
+
+        assert!(remaining_safe.is_some(), "Safe relationship should remain");
         assert!(
-            !check_results.is_empty(),
-            "Relationship with different source_id should still exist"
+            remaining_other.is_some(),
+            "Other relationship should remain"
         );
     }
 }

common/src/storage/types/message.rs

@@ -1,3 +1,4 @@
+#![allow(clippy::module_name_repetitions)]
 use uuid::Uuid;
 
 use crate::stored_object;
@@ -56,7 +57,7 @@ impl fmt::Display for Message {
 pub fn format_history(history: &[Message]) -> String {
     history
         .iter()
-        .map(|msg| format!("{}", msg))
+        .map(|msg| format!("{msg}"))
         .collect::<Vec<String>>()
         .join("\n")
 }

common/src/storage/types/mod.rs

@@ -1,3 +1,4 @@
+#![allow(clippy::unsafe_derive_deserialize)]
 use serde::{Deserialize, Serialize};
 pub mod analytics;
 pub mod conversation;
@@ -5,12 +6,14 @@ pub mod file_info;
 pub mod ingestion_payload;
 pub mod ingestion_task;
 pub mod knowledge_entity;
+pub mod knowledge_entity_embedding;
 pub mod knowledge_relationship;
 pub mod message;
 pub mod scratchpad;
 pub mod system_prompts;
 pub mod system_settings;
 pub mod text_chunk;
+pub mod text_chunk_embedding;
 pub mod text_content;
 pub mod user;
 
@@ -21,7 +24,7 @@ pub trait StoredObject: Serialize + for<'de> Deserialize<'de> {
 
 #[macro_export]
 macro_rules! stored_object {
-    ($name:ident, $table:expr, {$($(#[$attr:meta])* $field:ident: $ty:ty),*}) => {
+    ($(#[$struct_attr:meta])* $name:ident, $table:expr, {$($(#[$field_attr:meta])* $field:ident: $ty:ty),*}) => {
         use serde::{Deserialize, Deserializer, Serialize};
         use surrealdb::sql::Thing;
         use $crate::storage::types::StoredObject;
@@ -85,6 +88,7 @@ macro_rules! stored_object {
         }
 
         #[allow(dead_code)]
+        #[allow(clippy::ref_option)]
         fn serialize_option_datetime<S>(
             date: &Option<DateTime<Utc>>,
             serializer: S,
@@ -100,6 +104,7 @@ macro_rules! stored_object {
         }
 
         #[allow(dead_code)]
+        #[allow(clippy::ref_option)]
         fn deserialize_option_datetime<'de, D>(
             deserializer: D,
         ) -> Result<Option<DateTime<Utc>>, D::Error>
@@ -111,6 +116,7 @@ macro_rules! stored_object {
         }
 
 
+        $(#[$struct_attr])*
         #[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
         pub struct $name {
             #[serde(deserialize_with = "deserialize_flexible_id")]
@@ -119,7 +125,7 @@ macro_rules! stored_object {
             pub created_at: DateTime<Utc>,
             #[serde(serialize_with = "serialize_datetime", deserialize_with = "deserialize_datetime", default)]
             pub updated_at: DateTime<Utc>,
-            $( $(#[$attr])* pub $field: $ty),*
+            $( $(#[$field_attr])* pub $field: $ty),*
         }
 
         impl StoredObject for $name {

common/src/storage/types/system_settings.rs

@@ -13,6 +13,9 @@ pub struct SystemSettings {
     pub processing_model: String,
     pub embedding_model: String,
     pub embedding_dimensions: u32,
+    /// Active embedding backend ("openai", "fastembed", "hashed"). Read-only, synced from config.
+    #[serde(default)]
+    pub embedding_backend: Option<String>,
     pub query_system_prompt: String,
     pub ingestion_system_prompt: String,
     pub image_processing_model: String,
@@ -49,10 +52,62 @@ impl SystemSettings {
             "Something went wrong updating the settings".into(),
         ))
     }
+
+    /// Syncs SystemSettings with the active embedding provider's properties.
+    /// Updates embedding_backend, embedding_model, and embedding_dimensions if they differ.
+    /// Returns true if any settings were changed.
+    pub async fn sync_from_embedding_provider(
+        db: &SurrealDbClient,
+        provider: &crate::utils::embedding::EmbeddingProvider,
+    ) -> Result<(Self, bool), AppError> {
+        let mut settings = Self::get_current(db).await?;
+        let mut needs_update = false;
+
+        let backend_label = provider.backend_label().to_string();
+        let provider_dimensions = provider.dimension() as u32;
+        let provider_model = provider.model_code();
+
+        // Sync backend label
+        if settings.embedding_backend.as_deref() != Some(&backend_label) {
+            settings.embedding_backend = Some(backend_label);
+            needs_update = true;
+        }
+
+        // Sync dimensions
+        if settings.embedding_dimensions != provider_dimensions {
+            tracing::info!(
+                old_dimensions = settings.embedding_dimensions,
+                new_dimensions = provider_dimensions,
+                "Embedding dimensions changed, updating SystemSettings"
+            );
+            settings.embedding_dimensions = provider_dimensions;
+            needs_update = true;
+        }
+
+        // Sync model if provider has one
+        if let Some(model) = provider_model {
+            if settings.embedding_model != model {
+                tracing::info!(
+                    old_model = %settings.embedding_model,
+                    new_model = %model,
+                    "Embedding model changed, updating SystemSettings"
+                );
+                settings.embedding_model = model;
+                needs_update = true;
+            }
+        }
+
+        if needs_update {
+            settings = Self::update(db, settings).await?;
+        }
+
+        Ok((settings, needs_update))
+    }
 }
 
 #[cfg(test)]
 mod tests {
+    use crate::storage::indexes::ensure_runtime_indexes;
     use crate::storage::types::{knowledge_entity::KnowledgeEntity, text_chunk::TextChunk};
     use async_openai::Client;
 
@@ -71,25 +126,22 @@ mod tests {
             .await
             .expect("Failed to fetch table info");
 
-        let info: Option<serde_json::Value> = response
+        let info: surrealdb::Value = response
             .take(0)
             .expect("Failed to extract table info response");
 
-        let info = info.expect("Table info result missing");
+        let info_json: serde_json::Value =
+            serde_json::to_value(info).expect("Failed to convert info to json");
 
-        let indexes = info
-            .get("indexes")
-            .or_else(|| {
-                info.get("tables")
-                    .and_then(|tables| tables.get(table_name))
-                    .and_then(|table| table.get("indexes"))
-            })
-            .unwrap_or_else(|| panic!("Indexes collection missing in table info: {info:#?}"));
+        let indexes = info_json["Object"]["indexes"]["Object"]
+            .as_object()
+            .unwrap_or_else(|| panic!("Indexes collection missing in table info: {info_json:#?}"));
 
         let definition = indexes
             .get(index_name)
-            .and_then(|definition| definition.as_str())
-            .unwrap_or_else(|| panic!("Index definition not found in table info: {info:#?}"));
+            .and_then(|definition| definition.get("Strand"))
+            .and_then(|v| v.as_str())
+            .unwrap_or_else(|| panic!("Index definition not found in table info: {info_json:#?}"));
 
         let dimension_part = definition
             .split("DIMENSION")
@@ -261,48 +313,56 @@ mod tests {
         let initial_chunk = TextChunk::new(
             "source1".into(),
             "This chunk has the original dimension".into(),
-            vec![0.1; 1536],
             "user1".into(),
         );
 
-        db.store_item(initial_chunk.clone())
+        TextChunk::store_with_embedding(initial_chunk.clone(), vec![0.1; 1536], &db)
             .await
-            .expect("Failed to store initial chunk");
+            .expect("Failed to store initial chunk with embedding");
 
         async fn simulate_reembedding(
             db: &SurrealDbClient,
             target_dimension: usize,
             initial_chunk: TextChunk,
         ) {
-            db.query("REMOVE INDEX idx_embedding_chunks ON TABLE text_chunk;")
-                .await
-                .unwrap();
+            db.query(
+                "REMOVE INDEX IF EXISTS idx_embedding_text_chunk_embedding ON TABLE text_chunk_embedding;",
+            )
+            .await
+            .unwrap();
             let define_index_query = format!(
-                         "DEFINE INDEX idx_embedding_chunks ON TABLE text_chunk FIELDS embedding HNSW DIMENSION {};",
-                         target_dimension
-                     );
+                "DEFINE INDEX idx_embedding_text_chunk_embedding ON TABLE text_chunk_embedding FIELDS embedding HNSW DIMENSION {};",
+                target_dimension
+            );
             db.query(define_index_query)
                 .await
                 .expect("Re-defining index should succeed");
 
             let new_embedding = vec![0.5; target_dimension];
-            let sql = "UPDATE type::thing('text_chunk', $id) SET embedding = $embedding;";
+            let sql = "UPSERT type::thing('text_chunk_embedding', $id) SET chunk_id = type::thing('text_chunk', $id), embedding = $embedding, user_id = $user_id;";
 
             let update_result = db
                 .client
                 .query(sql)
                 .bind(("id", initial_chunk.id.clone()))
+                .bind(("user_id", initial_chunk.user_id.clone()))
                 .bind(("embedding", new_embedding))
                 .await;
 
             assert!(update_result.is_ok());
         }
 
-        simulate_reembedding(&db, 768, initial_chunk).await;
+        // Re-embed with the existing configured dimension to ensure migrations remain idempotent.
+        let target_dimension = 1536usize;
+        simulate_reembedding(&db, target_dimension, initial_chunk).await;
 
         let migration_result = db.apply_migrations().await;
 
-        assert!(migration_result.is_ok(), "Migrations should not fail");
+        assert!(
+            migration_result.is_ok(),
+            "Migrations should not fail: {:?}",
+            migration_result.err()
+        );
     }
 
     #[tokio::test]
@@ -320,8 +380,17 @@ mod tests {
             .await
             .expect("Failed to load current settings");
 
-        let initial_chunk_dimension =
-            get_hnsw_index_dimension(&db, "text_chunk", "idx_embedding_chunks").await;
+        // Ensure runtime indexes exist with the current embedding dimension so INFO queries succeed.
+        ensure_runtime_indexes(&db, current_settings.embedding_dimensions as usize)
+            .await
+            .expect("failed to build runtime indexes");
+
+        let initial_chunk_dimension = get_hnsw_index_dimension(
+            &db,
+            "text_chunk_embedding",
+            "idx_embedding_text_chunk_embedding",
+        )
+        .await;
 
         assert_eq!(
             initial_chunk_dimension, current_settings.embedding_dimensions,
@@ -352,10 +421,18 @@ mod tests {
             .await
             .expect("KnowledgeEntity re-embedding should succeed on fresh DB");
 
-        let text_chunk_dimension =
-            get_hnsw_index_dimension(&db, "text_chunk", "idx_embedding_chunks").await;
-        let knowledge_dimension =
-            get_hnsw_index_dimension(&db, "knowledge_entity", "idx_embedding_entities").await;
+        let text_chunk_dimension = get_hnsw_index_dimension(
+            &db,
+            "text_chunk_embedding",
+            "idx_embedding_text_chunk_embedding",
+        )
+        .await;
+        let knowledge_dimension = get_hnsw_index_dimension(
+            &db,
+            "knowledge_entity_embedding",
+            "idx_embedding_knowledge_entity_embedding",
+        )
+        .await;
 
         assert_eq!(
             text_chunk_dimension, new_dimension,

common/src/storage/types/text_chunk_embedding.rs

@@ -0,0 +1,411 @@
+use surrealdb::RecordId;
+
+use crate::storage::types::text_chunk::TextChunk;
+use crate::{error::AppError, storage::db::SurrealDbClient, stored_object};
+
+stored_object!(TextChunkEmbedding, "text_chunk_embedding", {
+    /// Record link to the owning text_chunk
+    chunk_id: RecordId,
+    /// Denormalized source id for bulk deletes
+    source_id: String,
+    /// Embedding vector
+    embedding: Vec<f32>,
+    /// Denormalized user id (for scoping + permissions)
+    user_id: String
+});
+
+impl TextChunkEmbedding {
+    /// Recreate the HNSW index with a new embedding dimension.
+    ///
+    /// This is useful when the embedding length changes; Surreal requires the
+    /// index definition to be recreated with the updated dimension.
+    pub async fn redefine_hnsw_index(
+        db: &SurrealDbClient,
+        dimension: usize,
+    ) -> Result<(), AppError> {
+        let query = format!(
+            "BEGIN TRANSACTION;
+             REMOVE INDEX IF EXISTS idx_embedding_text_chunk_embedding ON TABLE {table};
+             DEFINE INDEX idx_embedding_text_chunk_embedding ON TABLE {table} FIELDS embedding HNSW DIMENSION {dimension};
+             COMMIT TRANSACTION;",
+            table = Self::table_name(),
+        );
+
+        let res = db.client.query(query).await.map_err(AppError::Database)?;
+        res.check().map_err(AppError::Database)?;
+
+        Ok(())
+    }
+
+    /// Create a new text chunk embedding
+    ///
+    /// `chunk_id` is the **key** part of the text_chunk id (e.g. the UUID),
+    /// not "text_chunk:uuid".
+    pub fn new(chunk_id: &str, source_id: String, embedding: Vec<f32>, user_id: String) -> Self {
+        let now = Utc::now();
+
+        Self {
+            // NOTE: `stored_object!` macro defines `id` as `String`
+            id: uuid::Uuid::new_v4().to_string(),
+            created_at: now,
+            updated_at: now,
+            // Create a record<text_chunk> link: text_chunk:<chunk_id>
+            chunk_id: RecordId::from_table_key(TextChunk::table_name(), chunk_id),
+            source_id,
+            embedding,
+            user_id,
+        }
+    }
+
+    /// Get a single embedding by its chunk RecordId
+    pub async fn get_by_chunk_id(
+        chunk_id: &RecordId,
+        db: &SurrealDbClient,
+    ) -> Result<Option<Self>, AppError> {
+        let query = format!(
+            "SELECT * FROM {} WHERE chunk_id = $chunk_id LIMIT 1",
+            Self::table_name()
+        );
+
+        let mut result = db
+            .client
+            .query(query)
+            .bind(("chunk_id", chunk_id.clone()))
+            .await
+            .map_err(AppError::Database)?;
+
+        let embeddings: Vec<Self> = result.take(0).map_err(AppError::Database)?;
+
+        Ok(embeddings.into_iter().next())
+    }
+
+    /// Delete embeddings for a given chunk RecordId
+    pub async fn delete_by_chunk_id(
+        chunk_id: &RecordId,
+        db: &SurrealDbClient,
+    ) -> Result<(), AppError> {
+        let query = format!(
+            "DELETE FROM {} WHERE chunk_id = $chunk_id",
+            Self::table_name()
+        );
+
+        db.client
+            .query(query)
+            .bind(("chunk_id", chunk_id.clone()))
+            .await
+            .map_err(AppError::Database)?
+            .check()
+            .map_err(AppError::Database)?;
+
+        Ok(())
+    }
+
+    /// Delete all embeddings that belong to chunks with a given `source_id`
+    ///
+    /// This uses the denormalized `source_id` on the embedding table.
+    pub async fn delete_by_source_id(
+        source_id: &str,
+        db: &SurrealDbClient,
+    ) -> Result<(), AppError> {
+        let query = format!(
+            "DELETE FROM {} WHERE source_id = $source_id",
+            Self::table_name()
+        );
+
+        db.client
+            .query(query)
+            .bind(("source_id", source_id.to_owned()))
+            .await
+            .map_err(AppError::Database)?
+            .check()
+            .map_err(AppError::Database)?;
+
+        Ok(())
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::storage::db::SurrealDbClient;
+    use surrealdb::Value as SurrealValue;
+    use uuid::Uuid;
+
+    /// Helper to create an in-memory DB and apply migrations
+    async fn setup_test_db() -> SurrealDbClient {
+        let namespace = "test_ns";
+        let database = Uuid::new_v4().to_string();
+        let db = SurrealDbClient::memory(namespace, &database)
+            .await
+            .expect("Failed to start in-memory surrealdb");
+
+        db.apply_migrations()
+            .await
+            .expect("Failed to apply migrations");
+
+        db
+    }
+
+    /// Helper: create a text_chunk with a known key, return its RecordId
+    async fn create_text_chunk_with_id(
+        db: &SurrealDbClient,
+        key: &str,
+        source_id: &str,
+        user_id: &str,
+    ) -> RecordId {
+        let chunk = TextChunk {
+            id: key.to_owned(),
+            created_at: Utc::now(),
+            updated_at: Utc::now(),
+            source_id: source_id.to_owned(),
+            chunk: "Some test chunk text".to_owned(),
+            user_id: user_id.to_owned(),
+        };
+
+        db.store_item(chunk)
+            .await
+            .expect("Failed to create text_chunk");
+
+        RecordId::from_table_key(TextChunk::table_name(), key)
+    }
+
+    #[tokio::test]
+    async fn test_create_and_get_by_chunk_id() {
+        let db = setup_test_db().await;
+
+        let user_id = "user_a";
+        let chunk_key = "chunk-123";
+        let source_id = "source-1";
+
+        // 1) Create a text_chunk with a known key
+        let chunk_rid = create_text_chunk_with_id(&db, chunk_key, source_id, user_id).await;
+
+        // 2) Create and store an embedding for that chunk
+        let embedding_vec = vec![0.1_f32, 0.2, 0.3];
+        let emb = TextChunkEmbedding::new(
+            chunk_key,
+            source_id.to_string(),
+            embedding_vec.clone(),
+            user_id.to_string(),
+        );
+
+        TextChunkEmbedding::redefine_hnsw_index(&db, emb.embedding.len())
+            .await
+            .expect("Failed to redefine index length");
+
+        let _: Option<TextChunkEmbedding> = db
+            .client
+            .create(TextChunkEmbedding::table_name())
+            .content(emb)
+            .await
+            .expect("Failed to store embedding")
+            .take()
+            .expect("Failed to deserialize stored embedding");
+
+        // 3) Fetch it via get_by_chunk_id
+        let fetched = TextChunkEmbedding::get_by_chunk_id(&chunk_rid, &db)
+            .await
+            .expect("Failed to get embedding by chunk_id");
+
+        assert!(fetched.is_some(), "Expected an embedding to be found");
+        let fetched = fetched.unwrap();
+
+        assert_eq!(fetched.user_id, user_id);
+        assert_eq!(fetched.chunk_id, chunk_rid);
+        assert_eq!(fetched.embedding, embedding_vec);
+    }
+
+    #[tokio::test]
+    async fn test_delete_by_chunk_id() {
+        let db = setup_test_db().await;
+
+        let user_id = "user_b";
+        let chunk_key = "chunk-delete";
+        let source_id = "source-del";
+
+        let chunk_rid = create_text_chunk_with_id(&db, chunk_key, source_id, user_id).await;
+
+        let emb = TextChunkEmbedding::new(
+            chunk_key,
+            source_id.to_string(),
+            vec![0.4_f32, 0.5, 0.6],
+            user_id.to_string(),
+        );
+
+        TextChunkEmbedding::redefine_hnsw_index(&db, emb.embedding.len())
+            .await
+            .expect("Failed to redefine index length");
+
+        let _: Option<TextChunkEmbedding> = db
+            .client
+            .create(TextChunkEmbedding::table_name())
+            .content(emb)
+            .await
+            .expect("Failed to store embedding")
+            .take()
+            .expect("Failed to deserialize stored embedding");
+
+        // Ensure it exists
+        let existing = TextChunkEmbedding::get_by_chunk_id(&chunk_rid, &db)
+            .await
+            .expect("Failed to get embedding before delete");
+        assert!(existing.is_some(), "Embedding should exist before delete");
+
+        // Delete by chunk_id
+        TextChunkEmbedding::delete_by_chunk_id(&chunk_rid, &db)
+            .await
+            .expect("Failed to delete by chunk_id");
+
+        // Ensure it no longer exists
+        let after = TextChunkEmbedding::get_by_chunk_id(&chunk_rid, &db)
+            .await
+            .expect("Failed to get embedding after delete");
+        assert!(after.is_none(), "Embedding should have been deleted");
+    }
+
+    #[tokio::test]
+    async fn test_delete_by_source_id() {
+        let db = setup_test_db().await;
+
+        let user_id = "user_c";
+        let source_id = "shared-source";
+        let other_source = "other-source";
+
+        // Two chunks with the same source_id
+        let chunk1_rid = create_text_chunk_with_id(&db, "chunk-s1", source_id, user_id).await;
+        let chunk2_rid = create_text_chunk_with_id(&db, "chunk-s2", source_id, user_id).await;
+
+        // One chunk with a different source_id
+        let chunk_other_rid =
+            create_text_chunk_with_id(&db, "chunk-other", other_source, user_id).await;
+
+        // Create embeddings for all three
+        let emb1 = TextChunkEmbedding::new(
+            "chunk-s1",
+            source_id.to_string(),
+            vec![0.1],
+            user_id.to_string(),
+        );
+        let emb2 = TextChunkEmbedding::new(
+            "chunk-s2",
+            source_id.to_string(),
+            vec![0.2],
+            user_id.to_string(),
+        );
+        let emb3 = TextChunkEmbedding::new(
+            "chunk-other",
+            other_source.to_string(),
+            vec![0.3],
+            user_id.to_string(),
+        );
+
+        // Update length on index
+        TextChunkEmbedding::redefine_hnsw_index(&db, emb1.embedding.len())
+            .await
+            .expect("Failed to redefine index length");
+
+        for emb in [emb1, emb2, emb3] {
+            let _: Option<TextChunkEmbedding> = db
+                .client
+                .create(TextChunkEmbedding::table_name())
+                .content(emb)
+                .await
+                .expect("Failed to store embedding")
+                .take()
+                .expect("Failed to deserialize stored embedding");
+        }
+
+        // Sanity check: they all exist
+        assert!(TextChunkEmbedding::get_by_chunk_id(&chunk1_rid, &db)
+            .await
+            .unwrap()
+            .is_some());
+        assert!(TextChunkEmbedding::get_by_chunk_id(&chunk2_rid, &db)
+            .await
+            .unwrap()
+            .is_some());
+        assert!(TextChunkEmbedding::get_by_chunk_id(&chunk_other_rid, &db)
+            .await
+            .unwrap()
+            .is_some());
+
+        // Delete embeddings by source_id (shared-source)
+        TextChunkEmbedding::delete_by_source_id(source_id, &db)
+            .await
+            .expect("Failed to delete by source_id");
+
+        // Chunks from shared-source should have no embeddings
+        assert!(TextChunkEmbedding::get_by_chunk_id(&chunk1_rid, &db)
+            .await
+            .unwrap()
+            .is_none());
+        assert!(TextChunkEmbedding::get_by_chunk_id(&chunk2_rid, &db)
+            .await
+            .unwrap()
+            .is_none());
+
+        // The other chunk should still have its embedding
+        assert!(TextChunkEmbedding::get_by_chunk_id(&chunk_other_rid, &db)
+            .await
+            .unwrap()
+            .is_some());
+    }
+
+    #[tokio::test]
+    async fn test_redefine_hnsw_index_updates_dimension() {
+        let db = setup_test_db().await;
+
+        // Change the index dimension from default (1536) to a smaller test value.
+        TextChunkEmbedding::redefine_hnsw_index(&db, 8)
+            .await
+            .expect("failed to redefine index");
+
+        let mut info_res = db
+            .client
+            .query("INFO FOR TABLE text_chunk_embedding;")
+            .await
+            .expect("info query failed");
+        let info: SurrealValue = info_res.take(0).expect("failed to take info result");
+        let info_json: serde_json::Value =
+            serde_json::to_value(info).expect("failed to convert info to json");
+        let idx_sql = info_json["Object"]["indexes"]["Object"]
+            ["idx_embedding_text_chunk_embedding"]["Strand"]
+            .as_str()
+            .unwrap_or_default();
+
+        assert!(
+            idx_sql.contains("DIMENSION 8"),
+            "expected index definition to contain new dimension, got: {idx_sql}"
+        );
+    }
+
+    #[tokio::test]
+    async fn test_redefine_hnsw_index_is_idempotent() {
+        let db = setup_test_db().await;
+
+        TextChunkEmbedding::redefine_hnsw_index(&db, 4)
+            .await
+            .expect("first redefine failed");
+        TextChunkEmbedding::redefine_hnsw_index(&db, 4)
+            .await
+            .expect("second redefine failed");
+
+        let mut info_res = db
+            .client
+            .query("INFO FOR TABLE text_chunk_embedding;")
+            .await
+            .expect("info query failed");
+        let info: SurrealValue = info_res.take(0).expect("failed to take info result");
+        let info_json: serde_json::Value =
+            serde_json::to_value(info).expect("failed to convert info to json");
+        let idx_sql = info_json["Object"]["indexes"]["Object"]
+            ["idx_embedding_text_chunk_embedding"]["Strand"]
+            .as_str()
+            .unwrap_or_default();
+
+        assert!(
+            idx_sql.contains("DIMENSION 4"),
+            "expected index definition to retain dimension 4, got: {idx_sql}"
+        );
+    }
+}

common/src/storage/types/text_content.rs

@@ -5,6 +5,7 @@ use crate::{error::AppError, storage::db::SurrealDbClient, stored_object};
 
 use super::file_info::FileInfo;
 
+#[allow(clippy::module_name_repetitions)]
 #[derive(Debug, Deserialize, Serialize)]
 pub struct TextContentSearchResult {
     #[serde(deserialize_with = "deserialize_flexible_id")]
@@ -50,8 +51,11 @@ pub struct TextContentSearchResult {
 
 #[derive(Serialize, Deserialize, Debug, Clone, PartialEq)]
 pub struct UrlInfo {
+    #[serde(default)]
     pub url: String,
+    #[serde(default)]
     pub title: String,
+    #[serde(default)]
     pub image_id: String,
 }
 
@@ -146,12 +150,12 @@ impl TextContent {
                 search::highlight('<b>', '</b>', 4) AS highlighted_url,       
                 search::highlight('<b>', '</b>', 5) AS highlighted_url_title, 
                 (
-                    search::score(0) +  
-                    search::score(1) +  
-                    search::score(2) +  
-                    search::score(3) +  
-                    search::score(4) +  
-                    search::score(5)    
+                    IF search::score(0) != NONE THEN search::score(0) ELSE 0 END +  
+                    IF search::score(1) != NONE THEN search::score(1) ELSE 0 END +  
+                    IF search::score(2) != NONE THEN search::score(2) ELSE 0 END +  
+                    IF search::score(3) != NONE THEN search::score(3) ELSE 0 END +  
+                    IF search::score(4) != NONE THEN search::score(4) ELSE 0 END +  
+                    IF search::score(5) != NONE THEN search::score(5) ELSE 0 END    
                 ) AS score  
             FROM text_content
             WHERE

common/src/storage/types/user.rs

@@ -1,4 +1,5 @@
 use crate::{error::AppError, storage::db::SurrealDbClient, stored_object};
+use anyhow::anyhow;
 use async_trait::async_trait;
 use axum_session_auth::Authentication;
 use chrono_tz::Tz;
@@ -17,29 +18,95 @@ use super::{
 use chrono::Duration;
 use futures::try_join;
 
+/// Result row for returning user category.
 #[derive(Deserialize)]
 pub struct CategoryResponse {
+    /// Category name tied to the user.
     category: String,
 }
 
-stored_object!(User, "user", {
+use std::str::FromStr;
+
+/// Supported UI themes.
+#[derive(Serialize, Deserialize, Debug, Clone, Copy, PartialEq, Eq, Default)]
+#[serde(rename_all = "kebab-case")]
+pub enum Theme {
+    Light,
+    Dark,
+    WarmPaper,
+    ObsidianPrism,
+    #[default]
+    System,
+}
+
+impl FromStr for Theme {
+    type Err = ();
+
+    fn from_str(s: &str) -> Result<Self, Self::Err> {
+        match s {
+            "light" => Ok(Self::Light),
+            "dark" => Ok(Self::Dark),
+            "warm-paper" => Ok(Self::WarmPaper),
+            "obsidian-prism" => Ok(Self::ObsidianPrism),
+            "system" => Ok(Self::System),
+            _ => Err(()),
+        }
+    }
+}
+
+impl Theme {
+    pub fn as_str(&self) -> &'static str {
+        match self {
+            Self::Light => "light",
+            Self::Dark => "dark",
+            Self::WarmPaper => "warm-paper",
+            Self::ObsidianPrism => "obsidian-prism",
+            Self::System => "system",
+        }
+    }
+
+    /// Returns the theme that should be initially applied.
+    /// For "system", defaults to "light".
+    pub fn initial_theme(&self) -> &'static str {
+        match self {
+            Self::System => "light",
+            other => other.as_str(),
+        }
+    }
+}
+
+stored_object!(
+    #[allow(clippy::unsafe_derive_deserialize)]
+    User, "user", {
     email: String,
     password: String,
     anonymous: bool,
     api_key: Option<String>,
     admin: bool,
     #[serde(default)]
-    timezone: String
+    timezone: String,
+    #[serde(default, deserialize_with = "deserialize_theme_or_default")]
+    theme: Theme
 });
 
+fn deserialize_theme_or_default<'de, D>(deserializer: D) -> Result<Theme, D::Error>
+where
+    D: serde::Deserializer<'de>,
+{
+    let raw = Option::<String>::deserialize(deserializer)?;
+    Ok(raw
+        .and_then(|value| Theme::from_str(value.as_str()).ok())
+        .unwrap_or_default())
+}
+
 #[async_trait]
 impl Authentication<User, String, Surreal<Any>> for User {
     async fn load_user(userid: String, db: Option<&Surreal<Any>>) -> Result<User, anyhow::Error> {
-        let db = db.unwrap();
+        let db = db.ok_or_else(|| anyhow!("Database handle missing"))?;
         Ok(db
             .select((Self::table_name(), userid.as_str()))
             .await?
-            .unwrap())
+            .ok_or_else(|| anyhow!("User {userid} not found"))?)
     }
 
     fn is_authenticated(&self) -> bool {
@@ -55,14 +122,19 @@ impl Authentication<User, String, Surreal<Any>> for User {
     }
 }
 
+/// Ensures a timezone string parses, defaulting to UTC when invalid.
 fn validate_timezone(input: &str) -> String {
-    match input.parse::<Tz>() {
-        Ok(_) => input.to_owned(),
-        Err(_) => {
-            tracing::warn!("Invalid timezone '{}' received, defaulting to UTC", input);
-            "UTC".to_owned()
-        }
+    if input.parse::<Tz>().is_ok() {
+        return input.to_owned();
     }
+
+    tracing::warn!("Invalid timezone '{}' received, defaulting to UTC", input);
+    "UTC".to_owned()
+}
+
+/// Ensures a theme string is valid, defaulting to "system" when invalid.
+fn validate_theme(input: &str) -> Theme {
+    Theme::from_str(input).unwrap_or_default()
 }
 
 #[derive(Serialize, Deserialize, Debug, Clone)]
@@ -77,12 +149,15 @@ pub struct DashboardStats {
     pub new_text_chunks_week: i64,
 }
 
+/// Helper for aggregating `SurrealDB` count responses.
 #[derive(Deserialize)]
 struct CountResult {
+    /// Row count returned by the query.
     count: i64,
 }
 
 impl User {
+    /// Counts all objects of a given type belonging to the user.
     async fn count_total<T: crate::storage::types::StoredObject>(
         db: &SurrealDbClient,
         user_id: &str,
@@ -94,9 +169,10 @@ impl User {
             .bind(("user_id", user_id.to_string()))
             .await?
             .take(0)?;
-        Ok(result.map(|r| r.count).unwrap_or(0))
+        Ok(result.map_or(0, |r| r.count))
     }
 
+    /// Counts objects of a given type created after a specific timestamp.
     async fn count_since<T: crate::storage::types::StoredObject>(
         db: &SurrealDbClient,
         user_id: &str,
@@ -112,14 +188,16 @@ impl User {
             .bind(("since", surrealdb::Datetime::from(since)))
             .await?
             .take(0)?;
-        Ok(result.map(|r| r.count).unwrap_or(0))
+        Ok(result.map_or(0, |r| r.count))
     }
 
     pub async fn get_dashboard_stats(
         user_id: &str,
         db: &SurrealDbClient,
     ) -> Result<DashboardStats, AppError> {
-        let since = chrono::Utc::now() - Duration::days(7);
+        let since = chrono::Utc::now()
+            .checked_sub_signed(Duration::days(7))
+            .unwrap_or_else(chrono::Utc::now);
 
         let (
             total_documents,
@@ -157,6 +235,7 @@ impl User {
         password: String,
         db: &SurrealDbClient,
         timezone: String,
+        theme: String,
     ) -> Result<Self, AppError> {
         // verify that the application allows new creations
         let systemsettings = SystemSettings::get_current(db).await?;
@@ -165,6 +244,7 @@ impl User {
         }
 
         let validated_tz = validate_timezone(&timezone);
+        let validated_theme = validate_theme(&theme);
         let now = Utc::now();
         let id = Uuid::new_v4().to_string();
 
@@ -179,7 +259,8 @@ impl User {
                 anonymous = false,
                 created_at = $created_at,
                 updated_at = $updated_at,
-                timezone = $timezone",
+                timezone = $timezone,
+                theme = $theme",
             )
             .bind(("table", "user"))
             .bind(("id", id))
@@ -188,6 +269,7 @@ impl User {
             .bind(("created_at", surrealdb::Datetime::from(now)))
             .bind(("updated_at", surrealdb::Datetime::from(now)))
             .bind(("timezone", validated_tz))
+            .bind(("theme", validated_theme.as_str()))
             .await?
             .take(1)?;
 
@@ -261,7 +343,7 @@ impl User {
 
     pub async fn set_api_key(id: &str, db: &SurrealDbClient) -> Result<String, AppError> {
         // Generate a secure random API key
-        let api_key = format!("sk_{}", Uuid::new_v4().to_string().replace("-", ""));
+        let api_key = format!("sk_{}", Uuid::new_v4().to_string().replace('-', ""));
 
         // Update the user record with the new API key
         let user: Option<Self> = db
@@ -341,6 +423,7 @@ impl User {
     ) -> Result<Vec<String>, AppError> {
         #[derive(Deserialize)]
         struct EntityTypeResponse {
+            /// Raw entity type value from the database.
             entity_type: String,
         }
 
@@ -358,7 +441,7 @@ impl User {
             .into_iter()
             .map(|item| {
                 let normalized = KnowledgeEntityType::from(item.entity_type);
-                format!("{:?}", normalized)
+                format!("{normalized:?}")
             })
             .collect();
 
@@ -456,6 +539,19 @@ impl User {
         Ok(())
     }
 
+    pub async fn update_theme(
+        user_id: &str,
+        theme: &str,
+        db: &SurrealDbClient,
+    ) -> Result<(), AppError> {
+        let validated_theme = validate_theme(theme);
+        db.query("UPDATE type::thing('user', $user_id) SET theme = $theme")
+            .bind(("user_id", user_id.to_string()))
+            .bind(("theme", validated_theme.as_str()))
+            .await?;
+        Ok(())
+    }
+
     pub async fn get_user_categories(
         user_id: &str,
         db: &SurrealDbClient,
@@ -662,6 +758,7 @@ mod tests {
             password.to_string(),
             &db,
             timezone.to_string(),
+            "system".to_string(),
         )
         .await
         .expect("Failed to create user");
@@ -699,6 +796,7 @@ mod tests {
             password.to_string(),
             &db,
             "UTC".to_string(),
+            "system".to_string(),
         )
         .await
         .expect("Failed to create user");
@@ -846,6 +944,7 @@ mod tests {
             password.to_string(),
             &db,
             "UTC".to_string(),
+            "system".to_string(),
         )
         .await
         .expect("Failed to create user");
@@ -880,6 +979,7 @@ mod tests {
             password.to_string(),
             &db,
             "UTC".to_string(),
+            "system".to_string(),
         )
         .await
         .expect("Failed to create user");
@@ -932,6 +1032,42 @@ mod tests {
         assert!(not_found.is_none());
     }
 
+    #[tokio::test]
+    async fn test_set_api_key_with_none_theme() {
+        let db = setup_test_db().await;
+
+        let user = User::create_new(
+            "legacy_theme@example.com".to_string(),
+            "apikey_password".to_string(),
+            &db,
+            "UTC".to_string(),
+            "system".to_string(),
+        )
+        .await
+        .expect("Failed to create user");
+
+        db.client
+            .query("UPDATE type::thing('user', $id) SET theme = NONE")
+            .bind(("id", user.id.clone()))
+            .await
+            .expect("Failed to set user theme to NONE");
+
+        let api_key = User::set_api_key(&user.id, &db)
+            .await
+            .expect("set_api_key should tolerate NONE theme");
+
+        assert!(api_key.starts_with("sk_"));
+
+        let updated_user = db
+            .get_item::<User>(&user.id)
+            .await
+            .expect("Failed to retrieve user")
+            .expect("User should still exist");
+
+        assert_eq!(updated_user.theme, Theme::System);
+        assert_eq!(updated_user.api_key, Some(api_key));
+    }
+
     #[tokio::test]
     async fn test_password_update() {
         // Setup test database
@@ -947,6 +1083,7 @@ mod tests {
             old_password.to_string(),
             &db,
             "UTC".to_string(),
+            "system".to_string(),
         )
         .await
         .expect("Failed to create user");
@@ -994,6 +1131,7 @@ mod tests {
             "password".to_string(),
             &db,
             "UTC".to_string(),
+            "system".to_string(),
         )
         .await
         .expect("Failed to create user");
@@ -1104,4 +1242,51 @@ mod tests {
             );
         }
     }
+
+    #[tokio::test]
+    async fn test_validate_theme() {
+        assert_eq!(validate_theme("light"), Theme::Light);
+        assert_eq!(validate_theme("dark"), Theme::Dark);
+        assert_eq!(validate_theme("system"), Theme::System);
+        assert_eq!(validate_theme("invalid"), Theme::System);
+    }
+
+    #[tokio::test]
+    async fn test_theme_update() {
+        let db = setup_test_db().await;
+        let email = "theme_test@example.com";
+        let user = User::create_new(
+            email.to_string(),
+            "password".to_string(),
+            &db,
+            "UTC".to_string(),
+            "system".to_string(),
+        )
+        .await
+        .expect("Failed to create user");
+
+        assert_eq!(user.theme, Theme::System);
+
+        User::update_theme(&user.id, "dark", &db)
+            .await
+            .expect("update theme");
+
+        let updated = db
+            .get_item::<User>(&user.id)
+            .await
+            .expect("get user")
+            .unwrap();
+        assert_eq!(updated.theme, Theme::Dark);
+
+        // Invalid theme should default to system (but update_theme calls validate_theme)
+        User::update_theme(&user.id, "invalid", &db)
+            .await
+            .expect("update theme invalid");
+        let updated2 = db
+            .get_item::<User>(&user.id)
+            .await
+            .expect("get user")
+            .unwrap();
+        assert_eq!(updated2.theme, Theme::System);
+    }
 }

common/src/utils/config.rs

@@ -2,17 +2,36 @@ use config::{Config, ConfigError, Environment, File};
 use serde::Deserialize;
 use std::env;
 
+/// Selects the embedding backend for vector generation.
+#[derive(Clone, Deserialize, Debug, Default, PartialEq)]
+#[serde(rename_all = "lowercase")]
+pub enum EmbeddingBackend {
+    /// Use OpenAI-compatible API for embeddings.
+    OpenAI,
+    /// Use FastEmbed local embeddings (default).
+    #[default]
+    FastEmbed,
+    /// Use deterministic hashed embeddings (for testing).
+    Hashed,
+}
+
 #[derive(Clone, Deserialize, Debug, PartialEq)]
 #[serde(rename_all = "lowercase")]
 pub enum StorageKind {
     Local,
     Memory,
+    S3,
 }
 
+/// Default storage backend when none is configured.
 fn default_storage_kind() -> StorageKind {
     StorageKind::Local
 }
 
+fn default_s3_region() -> Option<String> {
+    Some("us-east-1".to_string())
+}
+
 /// Selects the strategy used for PDF ingestion.
 #[derive(Clone, Deserialize, Debug)]
 #[serde(rename_all = "kebab-case")]
@@ -23,10 +42,13 @@ pub enum PdfIngestMode {
     LlmFirst,
 }
 
+/// Default PDF ingestion mode when unset.
 fn default_pdf_ingest_mode() -> PdfIngestMode {
     PdfIngestMode::LlmFirst
 }
 
+/// Application configuration loaded from files and environment variables.
+#[allow(clippy::module_name_repetitions)]
 #[derive(Clone, Deserialize, Debug)]
 pub struct AppConfig {
     pub openai_api_key: String,
@@ -42,6 +64,12 @@ pub struct AppConfig {
     pub openai_base_url: String,
     #[serde(default = "default_storage_kind")]
     pub storage: StorageKind,
+    #[serde(default)]
+    pub s3_bucket: Option<String>,
+    #[serde(default)]
+    pub s3_endpoint: Option<String>,
+    #[serde(default = "default_s3_region")]
+    pub s3_region: Option<String>,
     #[serde(default = "default_pdf_ingest_mode")]
     pub pdf_ingest_mode: PdfIngestMode,
     #[serde(default = "default_reranking_enabled")]
@@ -54,20 +82,57 @@ pub struct AppConfig {
     pub fastembed_show_download_progress: Option<bool>,
     #[serde(default)]
     pub fastembed_max_length: Option<usize>,
+    #[serde(default)]
+    pub retrieval_strategy: Option<String>,
+    #[serde(default)]
+    pub embedding_backend: EmbeddingBackend,
+    #[serde(default = "default_ingest_max_body_bytes")]
+    pub ingest_max_body_bytes: usize,
+    #[serde(default = "default_ingest_max_files")]
+    pub ingest_max_files: usize,
+    #[serde(default = "default_ingest_max_content_bytes")]
+    pub ingest_max_content_bytes: usize,
+    #[serde(default = "default_ingest_max_context_bytes")]
+    pub ingest_max_context_bytes: usize,
+    #[serde(default = "default_ingest_max_category_bytes")]
+    pub ingest_max_category_bytes: usize,
 }
 
+/// Default data directory for persisted assets.
 fn default_data_dir() -> String {
     "./data".to_string()
 }
 
+/// Default base URL used for OpenAI-compatible APIs.
 fn default_base_url() -> String {
     "https://api.openai.com/v1".to_string()
 }
 
+/// Whether reranking is enabled by default.
 fn default_reranking_enabled() -> bool {
     false
 }
 
+fn default_ingest_max_body_bytes() -> usize {
+    20_000_000
+}
+
+fn default_ingest_max_files() -> usize {
+    5
+}
+
+fn default_ingest_max_content_bytes() -> usize {
+    262_144
+}
+
+fn default_ingest_max_context_bytes() -> usize {
+    16_384
+}
+
+fn default_ingest_max_category_bytes() -> usize {
+    128
+}
+
 pub fn ensure_ort_path() {
     if env::var_os("ORT_DYLIB_PATH").is_some() {
         return;
@@ -111,16 +176,28 @@ impl Default for AppConfig {
             http_port: 0,
             openai_base_url: default_base_url(),
             storage: default_storage_kind(),
+            s3_bucket: None,
+            s3_endpoint: None,
+            s3_region: default_s3_region(),
             pdf_ingest_mode: default_pdf_ingest_mode(),
             reranking_enabled: default_reranking_enabled(),
             reranking_pool_size: None,
             fastembed_cache_dir: None,
             fastembed_show_download_progress: None,
             fastembed_max_length: None,
+            retrieval_strategy: None,
+            embedding_backend: EmbeddingBackend::default(),
+            ingest_max_body_bytes: default_ingest_max_body_bytes(),
+            ingest_max_files: default_ingest_max_files(),
+            ingest_max_content_bytes: default_ingest_max_content_bytes(),
+            ingest_max_context_bytes: default_ingest_max_context_bytes(),
+            ingest_max_category_bytes: default_ingest_max_category_bytes(),
         }
     }
 }
 
+/// Loads the application configuration from the environment and optional config file.
+#[allow(clippy::module_name_repetitions)]
 pub fn get_config() -> Result<AppConfig, ConfigError> {
     ensure_ort_path();
 

common/src/utils/embedding.rs

@@ -1,19 +1,328 @@
-use async_openai::types::CreateEmbeddingRequestArgs;
+use std::{
+    collections::hash_map::DefaultHasher,
+    hash::{Hash, Hasher},
+    str::FromStr,
+    sync::Arc,
+};
+
+use anyhow::{anyhow, Context, Result};
+use async_openai::{types::CreateEmbeddingRequestArgs, Client};
+use fastembed::{EmbeddingModel, ModelTrait, TextEmbedding, TextInitOptions};
+use tokio::sync::Mutex;
 use tracing::debug;
 
 use crate::{
     error::AppError,
     storage::{db::SurrealDbClient, types::system_settings::SystemSettings},
 };
-/// Generates an embedding vector for the given input text using OpenAI's embedding model.
+
+/// Supported embedding backends.
+#[allow(clippy::module_name_repetitions)]
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Default)]
+pub enum EmbeddingBackend {
+    #[default]
+    OpenAI,
+    FastEmbed,
+    Hashed,
+}
+
+impl std::str::FromStr for EmbeddingBackend {
+    type Err = anyhow::Error;
+
+    fn from_str(s: &str) -> Result<Self, Self::Err> {
+        match s.to_ascii_lowercase().as_str() {
+            "openai" => Ok(Self::OpenAI),
+            "hashed" => Ok(Self::Hashed),
+            "fastembed" | "fast-embed" | "fast" => Ok(Self::FastEmbed),
+            other => Err(anyhow!(
+                "unknown embedding backend '{other}'. Expected 'openai', 'hashed', or 'fastembed'."
+            )),
+        }
+    }
+}
+
+/// Wrapper around the chosen embedding backend.
+#[allow(clippy::module_name_repetitions)]
+#[derive(Clone)]
+pub struct EmbeddingProvider {
+    /// Concrete backend implementation.
+    inner: EmbeddingInner,
+}
+
+/// Concrete embedding implementations.
+#[derive(Clone)]
+enum EmbeddingInner {
+    /// Uses an `OpenAI`-compatible API.
+    OpenAI {
+        /// Client used to issue embedding requests.
+        client: Arc<Client<async_openai::config::OpenAIConfig>>,
+        /// Model identifier for the API.
+        model: String,
+        /// Expected output dimensions.
+        dimensions: u32,
+    },
+    /// Generates deterministic hashed embeddings without external calls.
+    Hashed {
+        /// Output vector length.
+        dimension: usize,
+    },
+    /// Uses `FastEmbed` running locally.
+    FastEmbed {
+        /// Shared `FastEmbed` model.
+        model: Arc<Mutex<TextEmbedding>>,
+        /// Model metadata used for info logging.
+        model_name: EmbeddingModel,
+        /// Output vector length.
+        dimension: usize,
+    },
+}
+
+impl EmbeddingProvider {
+    pub fn backend_label(&self) -> &'static str {
+        match self.inner {
+            EmbeddingInner::Hashed { .. } => "hashed",
+            EmbeddingInner::FastEmbed { .. } => "fastembed",
+            EmbeddingInner::OpenAI { .. } => "openai",
+        }
+    }
+
+    pub fn dimension(&self) -> usize {
+        match &self.inner {
+            EmbeddingInner::Hashed { dimension } | EmbeddingInner::FastEmbed { dimension, .. } => {
+                *dimension
+            }
+            EmbeddingInner::OpenAI { dimensions, .. } => *dimensions as usize,
+        }
+    }
+
+    pub fn model_code(&self) -> Option<String> {
+        match &self.inner {
+            EmbeddingInner::FastEmbed { model_name, .. } => Some(model_name.to_string()),
+            EmbeddingInner::OpenAI { model, .. } => Some(model.clone()),
+            EmbeddingInner::Hashed { .. } => None,
+        }
+    }
+
+    pub async fn embed(&self, text: &str) -> Result<Vec<f32>> {
+        match &self.inner {
+            EmbeddingInner::Hashed { dimension } => Ok(hashed_embedding(text, *dimension)),
+            EmbeddingInner::FastEmbed { model, .. } => {
+                let mut guard = model.lock().await;
+                let embeddings = guard
+                    .embed(vec![text.to_owned()], None)
+                    .context("generating fastembed vector")?;
+                embeddings
+                    .into_iter()
+                    .next()
+                    .ok_or_else(|| anyhow!("fastembed returned no embedding for input"))
+            }
+            EmbeddingInner::OpenAI {
+                client,
+                model,
+                dimensions,
+            } => {
+                let request = CreateEmbeddingRequestArgs::default()
+                    .model(model.clone())
+                    .input([text])
+                    .dimensions(*dimensions)
+                    .build()?;
+
+                let response = client.embeddings().create(request).await?;
+
+                let embedding = response
+                    .data
+                    .first()
+                    .ok_or_else(|| anyhow!("No embedding data received from OpenAI API"))?
+                    .embedding
+                    .clone();
+
+                Ok(embedding)
+            }
+        }
+    }
+
+    pub async fn embed_batch(&self, texts: Vec<String>) -> Result<Vec<Vec<f32>>> {
+        match &self.inner {
+            EmbeddingInner::Hashed { dimension } => Ok(texts
+                .into_iter()
+                .map(|text| hashed_embedding(&text, *dimension))
+                .collect()),
+            EmbeddingInner::FastEmbed { model, .. } => {
+                if texts.is_empty() {
+                    return Ok(Vec::new());
+                }
+                let mut guard = model.lock().await;
+                guard
+                    .embed(texts, None)
+                    .context("generating fastembed batch embeddings")
+            }
+            EmbeddingInner::OpenAI {
+                client,
+                model,
+                dimensions,
+            } => {
+                if texts.is_empty() {
+                    return Ok(Vec::new());
+                }
+
+                let request = CreateEmbeddingRequestArgs::default()
+                    .model(model.clone())
+                    .input(texts)
+                    .dimensions(*dimensions)
+                    .build()?;
+
+                let response = client.embeddings().create(request).await?;
+
+                let embeddings: Vec<Vec<f32>> = response
+                    .data
+                    .into_iter()
+                    .map(|item| item.embedding)
+                    .collect();
+
+                Ok(embeddings)
+            }
+        }
+    }
+
+    pub fn new_openai(
+        client: Arc<Client<async_openai::config::OpenAIConfig>>,
+        model: String,
+        dimensions: u32,
+    ) -> Result<Self> {
+        Ok(Self {
+            inner: EmbeddingInner::OpenAI {
+                client,
+                model,
+                dimensions,
+            },
+        })
+    }
+
+    pub async fn new_fastembed(model_override: Option<String>) -> Result<Self> {
+        let model_name = if let Some(code) = model_override {
+            EmbeddingModel::from_str(&code).map_err(|err| anyhow!(err))?
+        } else {
+            EmbeddingModel::default()
+        };
+
+        let options = TextInitOptions::new(model_name.clone()).with_show_download_progress(true);
+        let model_name_for_task = model_name.clone();
+        let model_name_code = model_name.to_string();
+
+        let (model, dimension) = tokio::task::spawn_blocking(move || -> Result<_> {
+            let model =
+                TextEmbedding::try_new(options).context("initialising FastEmbed text model")?;
+            let info = EmbeddingModel::get_model_info(&model_name_for_task)
+                .ok_or_else(|| anyhow!("FastEmbed model metadata missing for {model_name_code}"))?;
+            Ok((model, info.dim))
+        })
+        .await
+        .context("joining FastEmbed initialisation task")??;
+
+        Ok(EmbeddingProvider {
+            inner: EmbeddingInner::FastEmbed {
+                model: Arc::new(Mutex::new(model)),
+                model_name,
+                dimension,
+            },
+        })
+    }
+
+    pub fn new_hashed(dimension: usize) -> Result<Self> {
+        Ok(EmbeddingProvider {
+            inner: EmbeddingInner::Hashed {
+                dimension: dimension.max(1),
+            },
+        })
+    }
+
+    /// Creates an embedding provider based on application configuration.
+    ///
+    /// Dispatches to the appropriate constructor based on `config.embedding_backend`:
+    /// - `OpenAI`: Requires a valid OpenAI client
+    /// - `FastEmbed`: Uses local embedding model
+    /// - `Hashed`: Uses deterministic hashed embeddings (for testing)
+    pub async fn from_config(
+        config: &crate::utils::config::AppConfig,
+        openai_client: Option<Arc<Client<async_openai::config::OpenAIConfig>>>,
+    ) -> Result<Self> {
+        use crate::utils::config::EmbeddingBackend;
+
+        match config.embedding_backend {
+            EmbeddingBackend::OpenAI => {
+                let client = openai_client
+                    .ok_or_else(|| anyhow!("OpenAI embedding backend requires an OpenAI client"))?;
+                // Use defaults that match SystemSettings initial values
+                Self::new_openai(client, "text-embedding-3-small".to_string(), 1536)
+            }
+            EmbeddingBackend::FastEmbed => {
+                // Use nomic-embed-text-v1.5 as the default FastEmbed model
+                Self::new_fastembed(Some("nomic-ai/nomic-embed-text-v1.5".to_string())).await
+            }
+            EmbeddingBackend::Hashed => Self::new_hashed(384),
+        }
+    }
+}
+
+// Helper functions for hashed embeddings
+/// Generates a hashed embedding vector without external dependencies.
+fn hashed_embedding(text: &str, dimension: usize) -> Vec<f32> {
+    let dim = dimension.max(1);
+    let mut vector = vec![0.0f32; dim];
+    if text.is_empty() {
+        return vector;
+    }
+
+    for token in tokens(text) {
+        let idx = bucket(&token, dim);
+        if let Some(slot) = vector.get_mut(idx) {
+            *slot += 1.0;
+        }
+    }
+
+    let norm = vector.iter().map(|v| v * v).sum::<f32>().sqrt();
+    if norm > 0.0 {
+        for value in &mut vector {
+            *value /= norm;
+        }
+    }
+
+    vector
+}
+
+/// Tokenizes the text into alphanumeric lowercase tokens.
+fn tokens(text: &str) -> impl Iterator<Item = String> + '_ {
+    text.split(|c: char| !c.is_ascii_alphanumeric())
+        .filter(|token| !token.is_empty())
+        .map(str::to_ascii_lowercase)
+}
+
+/// Buckets a token into the hashed embedding vector.
+#[allow(clippy::arithmetic_side_effects)]
+fn bucket(token: &str, dimension: usize) -> usize {
+    let safe_dimension = dimension.max(1);
+    let mut hasher = DefaultHasher::new();
+    token.hash(&mut hasher);
+    usize::try_from(hasher.finish()).unwrap_or_default() % safe_dimension
+}
+
+// Backward compatibility function
+pub async fn generate_embedding_with_provider(
+    provider: &EmbeddingProvider,
+    input: &str,
+) -> Result<Vec<f32>, AppError> {
+    provider.embed(input).await.map_err(AppError::from)
+}
+
+/// Generates an embedding vector for the given input text using `OpenAI`'s embedding model.
 ///
 /// This function takes a text input and converts it into a numerical vector representation (embedding)
-/// using OpenAI's text-embedding-3-small model. These embeddings can be used for semantic similarity
+/// using `OpenAI`'s text-embedding-3-small model. These embeddings can be used for semantic similarity
 /// comparisons, vector search, and other natural language processing tasks.
 ///
 /// # Arguments
 ///
-/// * `client`: The OpenAI client instance used to make API requests.
+/// * `client`: The `OpenAI` client instance used to make API requests.
 /// * `input`: The text string to generate embeddings for.
 ///
 /// # Returns
@@ -25,9 +334,10 @@ use crate::{
 /// # Errors
 ///
 /// This function can return a `AppError` in the following cases:
-/// * If the OpenAI API request fails
+/// * If the `OpenAI` API request fails
 /// * If the request building fails
 /// * If no embedding data is received in the response
+#[allow(clippy::module_name_repetitions)]
 pub async fn generate_embedding(
     client: &async_openai::Client<async_openai::config::OpenAIConfig>,
     input: &str,

common/src/utils/ingest_limits.rs

@@ -0,0 +1,113 @@
+use super::config::AppConfig;
+
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub enum IngestValidationError {
+    PayloadTooLarge(String),
+    BadRequest(String),
+}
+
+pub fn validate_ingest_input(
+    config: &AppConfig,
+    content: Option<&str>,
+    context: &str,
+    category: &str,
+    file_count: usize,
+) -> Result<(), IngestValidationError> {
+    if file_count > config.ingest_max_files {
+        return Err(IngestValidationError::BadRequest(format!(
+            "Too many files. Maximum allowed is {}",
+            config.ingest_max_files
+        )));
+    }
+
+    if let Some(content) = content {
+        if content.len() > config.ingest_max_content_bytes {
+            return Err(IngestValidationError::PayloadTooLarge(format!(
+                "Content is too large. Maximum allowed is {} bytes",
+                config.ingest_max_content_bytes
+            )));
+        }
+    }
+
+    if context.len() > config.ingest_max_context_bytes {
+        return Err(IngestValidationError::PayloadTooLarge(format!(
+            "Context is too large. Maximum allowed is {} bytes",
+            config.ingest_max_context_bytes
+        )));
+    }
+
+    if category.len() > config.ingest_max_category_bytes {
+        return Err(IngestValidationError::PayloadTooLarge(format!(
+            "Category is too large. Maximum allowed is {} bytes",
+            config.ingest_max_category_bytes
+        )));
+    }
+
+    Ok(())
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn validate_ingest_input_rejects_too_many_files() {
+        let config = AppConfig {
+            ingest_max_files: 1,
+            ..Default::default()
+        };
+        let result = validate_ingest_input(&config, Some("ok"), "ctx", "cat", 2);
+
+        assert!(matches!(result, Err(IngestValidationError::BadRequest(_))));
+    }
+
+    #[test]
+    fn validate_ingest_input_rejects_oversized_content() {
+        let config = AppConfig {
+            ingest_max_content_bytes: 4,
+            ..Default::default()
+        };
+        let result = validate_ingest_input(&config, Some("12345"), "ctx", "cat", 0);
+
+        assert!(matches!(
+            result,
+            Err(IngestValidationError::PayloadTooLarge(_))
+        ));
+    }
+
+    #[test]
+    fn validate_ingest_input_rejects_oversized_context() {
+        let config = AppConfig {
+            ingest_max_context_bytes: 2,
+            ..Default::default()
+        };
+        let result = validate_ingest_input(&config, None, "long", "cat", 0);
+
+        assert!(matches!(
+            result,
+            Err(IngestValidationError::PayloadTooLarge(_))
+        ));
+    }
+
+    #[test]
+    fn validate_ingest_input_rejects_oversized_category() {
+        let config = AppConfig {
+            ingest_max_category_bytes: 2,
+            ..Default::default()
+        };
+        let result = validate_ingest_input(&config, None, "ok", "long", 0);
+
+        assert!(matches!(
+            result,
+            Err(IngestValidationError::PayloadTooLarge(_))
+        ));
+    }
+
+    #[test]
+    fn validate_ingest_input_accepts_valid_payload() {
+        let config = AppConfig::default();
+        let result = validate_ingest_input(&config, Some("ok"), "ctx", "cat", 1);
+
+        assert!(result.is_ok());
+    }
+}

common/src/utils/mod.rs

@@ -1,3 +1,4 @@
 pub mod config;
 pub mod embedding;
+pub mod ingest_limits;
 pub mod template_engine;

common/src/utils/template_engine.rs

@@ -4,6 +4,7 @@ pub use minijinja_contrib;
 pub use minijinja_embed;
 use std::sync::Arc;
 
+#[allow(clippy::module_name_repetitions)]
 pub trait ProvidesTemplateEngine {
     fn template_engine(&self) -> &Arc<TemplateEngine>;
 }
@@ -20,19 +21,49 @@ pub enum TemplateEngine {
 
 #[macro_export]
 macro_rules! create_template_engine {
-    // Macro takes the relative path to the templates dir as input
-    ($relative_path:expr) => {{
+    // Single path argument
+    ($relative_path:expr) => {
+        $crate::create_template_engine!($relative_path, Option::<&str>::None)
+    };
+
+    // Path + Fallback argument
+    ($relative_path:expr, $fallback_path:expr) => {{
         // Code for debug builds (AutoReload)
         #[cfg(debug_assertions)]
         {
             // These lines execute in the CALLING crate's context
             let crate_dir = std::path::PathBuf::from(env!("CARGO_MANIFEST_DIR"));
             let template_path = crate_dir.join($relative_path);
+            let fallback_path = $fallback_path.map(|p| crate_dir.join(p));
+
             let reloader = $crate::utils::template_engine::AutoReloader::new(move |notifier| {
                 let mut env = $crate::utils::template_engine::Environment::new();
-                env.set_loader($crate::utils::template_engine::path_loader(&template_path));
+
+                let loader_primary = $crate::utils::template_engine::path_loader(&template_path);
+
+                // Clone fallback_path for the closure
+                let fallback = fallback_path.clone();
+
+                env.set_loader(move |name| match loader_primary(name) {
+                    Ok(Some(tmpl)) => Ok(Some(tmpl)),
+                    Ok(None) => {
+                        if let Some(ref fb_path) = fallback {
+                            let loader_fallback =
+                                $crate::utils::template_engine::path_loader(fb_path);
+                            loader_fallback(name)
+                        } else {
+                            Ok(None)
+                        }
+                    }
+                    Err(e) => Err(e),
+                });
+
                 notifier.set_fast_reload(true);
                 notifier.watch_path(&template_path, true);
+                if let Some(ref fb) = fallback_path {
+                    notifier.watch_path(fb, true);
+                }
+
                 // Add contrib filters/functions
                 $crate::utils::template_engine::minijinja_contrib::add_to_environment(&mut env);
                 Ok(env)

composite-retrieval/src/fts.rs

@@ -1,265 +0,0 @@
-use std::collections::HashMap;
-
-use serde::Deserialize;
-use tracing::debug;
-
-use common::{
-    error::AppError,
-    storage::{db::SurrealDbClient, types::StoredObject},
-};
-
-use crate::scoring::Scored;
-use common::storage::types::file_info::deserialize_flexible_id;
-use surrealdb::sql::Thing;
-
-#[derive(Debug, Deserialize)]
-struct FtsScoreRow {
-    #[serde(deserialize_with = "deserialize_flexible_id")]
-    id: String,
-    fts_score: Option<f32>,
-}
-
-/// Executes a full-text search query against SurrealDB and returns scored results.
-///
-/// The function expects FTS indexes to exist for the provided table. Currently supports
-/// `knowledge_entity` (name + description) and `text_chunk` (chunk).
-pub async fn find_items_by_fts<T>(
-    take: usize,
-    query: &str,
-    db_client: &SurrealDbClient,
-    table: &str,
-    user_id: &str,
-) -> Result<Vec<Scored<T>>, AppError>
-where
-    T: for<'de> serde::Deserialize<'de> + StoredObject,
-{
-    let (filter_clause, score_clause) = match table {
-        "knowledge_entity" => (
-            "(name @0@ $terms OR description @1@ $terms)",
-            "(IF search::score(0) != NONE THEN search::score(0) ELSE 0 END) + \
-             (IF search::score(1) != NONE THEN search::score(1) ELSE 0 END)",
-        ),
-        "text_chunk" => (
-            "(chunk @0@ $terms)",
-            "IF search::score(0) != NONE THEN search::score(0) ELSE 0 END",
-        ),
-        _ => {
-            return Err(AppError::Validation(format!(
-                "FTS not configured for table '{table}'"
-            )))
-        }
-    };
-
-    let sql = format!(
-        "SELECT id, {score_clause} AS fts_score \
-         FROM {table} \
-         WHERE {filter_clause} \
-           AND user_id = $user_id \
-         ORDER BY fts_score DESC \
-         LIMIT $limit",
-        table = table,
-        filter_clause = filter_clause,
-        score_clause = score_clause
-    );
-
-    debug!(
-        table = table,
-        limit = take,
-        "Executing FTS query with filter clause: {}",
-        filter_clause
-    );
-
-    let mut response = db_client
-        .query(sql)
-        .bind(("terms", query.to_owned()))
-        .bind(("user_id", user_id.to_owned()))
-        .bind(("limit", take as i64))
-        .await?;
-
-    let score_rows: Vec<FtsScoreRow> = response.take(0)?;
-
-    if score_rows.is_empty() {
-        return Ok(Vec::new());
-    }
-
-    let ids: Vec<String> = score_rows.iter().map(|row| row.id.clone()).collect();
-    let thing_ids: Vec<Thing> = ids
-        .iter()
-        .map(|id| Thing::from((table, id.as_str())))
-        .collect();
-
-    let mut items_response = db_client
-        .query("SELECT * FROM type::table($table) WHERE id IN $things AND user_id = $user_id")
-        .bind(("table", table.to_owned()))
-        .bind(("things", thing_ids.clone()))
-        .bind(("user_id", user_id.to_owned()))
-        .await?;
-
-    let items: Vec<T> = items_response.take(0)?;
-
-    let mut item_map: HashMap<String, T> = items
-        .into_iter()
-        .map(|item| (item.get_id().to_owned(), item))
-        .collect();
-
-    let mut results = Vec::with_capacity(score_rows.len());
-    for row in score_rows {
-        if let Some(item) = item_map.remove(&row.id) {
-            let score = row.fts_score.unwrap_or_default();
-            results.push(Scored::new(item).with_fts_score(score));
-        }
-    }
-
-    Ok(results)
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-    use common::storage::types::{
-        knowledge_entity::{KnowledgeEntity, KnowledgeEntityType},
-        text_chunk::TextChunk,
-        StoredObject,
-    };
-    use uuid::Uuid;
-
-    fn dummy_embedding() -> Vec<f32> {
-        vec![0.0; 1536]
-    }
-
-    #[tokio::test]
-    async fn fts_preserves_single_field_score_for_name() {
-        let namespace = "fts_test_ns";
-        let database = &Uuid::new_v4().to_string();
-        let db = SurrealDbClient::memory(namespace, database)
-            .await
-            .expect("failed to create in-memory surreal");
-
-        db.apply_migrations()
-            .await
-            .expect("failed to apply migrations");
-
-        let user_id = "user_fts";
-        let entity = KnowledgeEntity::new(
-            "source_a".into(),
-            "Rustacean handbook".into(),
-            "completely unrelated description".into(),
-            KnowledgeEntityType::Document,
-            None,
-            dummy_embedding(),
-            user_id.into(),
-        );
-
-        db.store_item(entity.clone())
-            .await
-            .expect("failed to insert entity");
-
-        db.rebuild_indexes()
-            .await
-            .expect("failed to rebuild indexes");
-
-        let results = find_items_by_fts::<KnowledgeEntity>(
-            5,
-            "rustacean",
-            &db,
-            KnowledgeEntity::table_name(),
-            user_id,
-        )
-        .await
-        .expect("fts query failed");
-
-        assert!(!results.is_empty(), "expected at least one FTS result");
-        assert!(
-            results[0].scores.fts.is_some(),
-            "expected an FTS score when only the name matched"
-        );
-    }
-
-    #[tokio::test]
-    async fn fts_preserves_single_field_score_for_description() {
-        let namespace = "fts_test_ns_desc";
-        let database = &Uuid::new_v4().to_string();
-        let db = SurrealDbClient::memory(namespace, database)
-            .await
-            .expect("failed to create in-memory surreal");
-
-        db.apply_migrations()
-            .await
-            .expect("failed to apply migrations");
-
-        let user_id = "user_fts_desc";
-        let entity = KnowledgeEntity::new(
-            "source_b".into(),
-            "neutral name".into(),
-            "Detailed notes about async runtimes".into(),
-            KnowledgeEntityType::Document,
-            None,
-            dummy_embedding(),
-            user_id.into(),
-        );
-
-        db.store_item(entity.clone())
-            .await
-            .expect("failed to insert entity");
-
-        db.rebuild_indexes()
-            .await
-            .expect("failed to rebuild indexes");
-
-        let results = find_items_by_fts::<KnowledgeEntity>(
-            5,
-            "async",
-            &db,
-            KnowledgeEntity::table_name(),
-            user_id,
-        )
-        .await
-        .expect("fts query failed");
-
-        assert!(!results.is_empty(), "expected at least one FTS result");
-        assert!(
-            results[0].scores.fts.is_some(),
-            "expected an FTS score when only the description matched"
-        );
-    }
-
-    #[tokio::test]
-    async fn fts_preserves_scores_for_text_chunks() {
-        let namespace = "fts_test_ns_chunks";
-        let database = &Uuid::new_v4().to_string();
-        let db = SurrealDbClient::memory(namespace, database)
-            .await
-            .expect("failed to create in-memory surreal");
-
-        db.apply_migrations()
-            .await
-            .expect("failed to apply migrations");
-
-        let user_id = "user_fts_chunk";
-        let chunk = TextChunk::new(
-            "source_chunk".into(),
-            "GraphQL documentation reference".into(),
-            dummy_embedding(),
-            user_id.into(),
-        );
-
-        db.store_item(chunk.clone())
-            .await
-            .expect("failed to insert chunk");
-
-        db.rebuild_indexes()
-            .await
-            .expect("failed to rebuild indexes");
-
-        let results =
-            find_items_by_fts::<TextChunk>(5, "graphql", &db, TextChunk::table_name(), user_id)
-                .await
-                .expect("fts query failed");
-
-        assert!(!results.is_empty(), "expected at least one FTS result");
-        assert!(
-            results[0].scores.fts.is_some(),
-            "expected an FTS score when chunk field matched"
-        );
-    }
-}

composite-retrieval/src/lib.rs

@@ -1,267 +0,0 @@
-pub mod answer_retrieval;
-pub mod answer_retrieval_helper;
-pub mod fts;
-pub mod graph;
-pub mod pipeline;
-pub mod reranking;
-pub mod scoring;
-pub mod vector;
-
-use common::{
-    error::AppError,
-    storage::{
-        db::SurrealDbClient,
-        types::{knowledge_entity::KnowledgeEntity, text_chunk::TextChunk},
-    },
-};
-use reranking::RerankerLease;
-use tracing::instrument;
-
-pub use pipeline::{retrieved_entities_to_json, RetrievalConfig, RetrievalTuning};
-
-// Captures a supporting chunk plus its fused retrieval score for downstream prompts.
-#[derive(Debug, Clone)]
-pub struct RetrievedChunk {
-    pub chunk: TextChunk,
-    pub score: f32,
-}
-
-// Final entity representation returned to callers, enriched with ranked chunks.
-#[derive(Debug, Clone)]
-pub struct RetrievedEntity {
-    pub entity: KnowledgeEntity,
-    pub score: f32,
-    pub chunks: Vec<RetrievedChunk>,
-}
-
-// Primary orchestrator for the process of retrieving KnowledgeEntitities related to a input_text
-#[instrument(skip_all, fields(user_id))]
-pub async fn retrieve_entities(
-    db_client: &SurrealDbClient,
-    openai_client: &async_openai::Client<async_openai::config::OpenAIConfig>,
-    input_text: &str,
-    user_id: &str,
-    reranker: Option<RerankerLease>,
-) -> Result<Vec<RetrievedEntity>, AppError> {
-    pipeline::run_pipeline(
-        db_client,
-        openai_client,
-        input_text,
-        user_id,
-        RetrievalConfig::default(),
-        reranker,
-    )
-    .await
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-    use async_openai::Client;
-    use common::storage::types::{
-        knowledge_entity::{KnowledgeEntity, KnowledgeEntityType},
-        knowledge_relationship::KnowledgeRelationship,
-        text_chunk::TextChunk,
-    };
-    use pipeline::RetrievalConfig;
-    use uuid::Uuid;
-
-    fn test_embedding() -> Vec<f32> {
-        vec![0.9, 0.1, 0.0]
-    }
-
-    fn entity_embedding_high() -> Vec<f32> {
-        vec![0.8, 0.2, 0.0]
-    }
-
-    fn entity_embedding_low() -> Vec<f32> {
-        vec![0.1, 0.9, 0.0]
-    }
-
-    fn chunk_embedding_primary() -> Vec<f32> {
-        vec![0.85, 0.15, 0.0]
-    }
-
-    fn chunk_embedding_secondary() -> Vec<f32> {
-        vec![0.2, 0.8, 0.0]
-    }
-
-    async fn setup_test_db() -> SurrealDbClient {
-        let namespace = "test_ns";
-        let database = &Uuid::new_v4().to_string();
-        let db = SurrealDbClient::memory(namespace, database)
-            .await
-            .expect("Failed to start in-memory surrealdb");
-
-        db.apply_migrations()
-            .await
-            .expect("Failed to apply migrations");
-
-        db.query(
-            "BEGIN TRANSACTION;
-             REMOVE INDEX IF EXISTS idx_embedding_chunks ON TABLE text_chunk;
-             DEFINE INDEX idx_embedding_chunks ON TABLE text_chunk FIELDS embedding HNSW DIMENSION 3;
-             REMOVE INDEX IF EXISTS idx_embedding_entities ON TABLE knowledge_entity;
-             DEFINE INDEX idx_embedding_entities ON TABLE knowledge_entity FIELDS embedding HNSW DIMENSION 3;
-             COMMIT TRANSACTION;",
-        )
-        .await
-        .expect("Failed to configure indices");
-
-        db
-    }
-
-    #[tokio::test]
-    async fn test_retrieve_entities_with_embedding_basic_flow() {
-        let db = setup_test_db().await;
-        let user_id = "test_user";
-        let entity = KnowledgeEntity::new(
-            "source_1".into(),
-            "Rust async guide".into(),
-            "Detailed notes about async runtimes".into(),
-            KnowledgeEntityType::Document,
-            None,
-            entity_embedding_high(),
-            user_id.into(),
-        );
-        let chunk = TextChunk::new(
-            entity.source_id.clone(),
-            "Tokio uses cooperative scheduling for fairness.".into(),
-            chunk_embedding_primary(),
-            user_id.into(),
-        );
-
-        db.store_item(entity.clone())
-            .await
-            .expect("Failed to store entity");
-        db.store_item(chunk.clone())
-            .await
-            .expect("Failed to store chunk");
-
-        let openai_client = Client::new();
-        let results = pipeline::run_pipeline_with_embedding(
-            &db,
-            &openai_client,
-            test_embedding(),
-            "Rust concurrency async tasks",
-            user_id,
-            RetrievalConfig::default(),
-            None,
-        )
-        .await
-        .expect("Hybrid retrieval failed");
-
-        assert!(
-            !results.is_empty(),
-            "Expected at least one retrieval result"
-        );
-        let top = &results[0];
-        assert!(
-            top.entity.name.contains("Rust"),
-            "Expected Rust entity to be ranked first"
-        );
-        assert!(
-            !top.chunks.is_empty(),
-            "Expected Rust entity to include supporting chunks"
-        );
-    }
-
-    #[tokio::test]
-    async fn test_graph_relationship_enriches_results() {
-        let db = setup_test_db().await;
-        let user_id = "graph_user";
-
-        let primary = KnowledgeEntity::new(
-            "primary_source".into(),
-            "Async Rust patterns".into(),
-            "Explores async runtimes and scheduling strategies.".into(),
-            KnowledgeEntityType::Document,
-            None,
-            entity_embedding_high(),
-            user_id.into(),
-        );
-        let neighbor = KnowledgeEntity::new(
-            "neighbor_source".into(),
-            "Tokio Scheduler Deep Dive".into(),
-            "Details on Tokio's cooperative scheduler.".into(),
-            KnowledgeEntityType::Document,
-            None,
-            entity_embedding_low(),
-            user_id.into(),
-        );
-
-        db.store_item(primary.clone())
-            .await
-            .expect("Failed to store primary entity");
-        db.store_item(neighbor.clone())
-            .await
-            .expect("Failed to store neighbor entity");
-
-        let primary_chunk = TextChunk::new(
-            primary.source_id.clone(),
-            "Rust async tasks use Tokio's cooperative scheduler.".into(),
-            chunk_embedding_primary(),
-            user_id.into(),
-        );
-        let neighbor_chunk = TextChunk::new(
-            neighbor.source_id.clone(),
-            "Tokio's scheduler manages task fairness across executors.".into(),
-            chunk_embedding_secondary(),
-            user_id.into(),
-        );
-
-        db.store_item(primary_chunk)
-            .await
-            .expect("Failed to store primary chunk");
-        db.store_item(neighbor_chunk)
-            .await
-            .expect("Failed to store neighbor chunk");
-
-        let openai_client = Client::new();
-        let relationship = KnowledgeRelationship::new(
-            primary.id.clone(),
-            neighbor.id.clone(),
-            user_id.into(),
-            "relationship_source".into(),
-            "references".into(),
-        );
-        relationship
-            .store_relationship(&db)
-            .await
-            .expect("Failed to store relationship");
-
-        let results = pipeline::run_pipeline_with_embedding(
-            &db,
-            &openai_client,
-            test_embedding(),
-            "Rust concurrency async tasks",
-            user_id,
-            RetrievalConfig::default(),
-            None,
-        )
-        .await
-        .expect("Hybrid retrieval failed");
-
-        let mut neighbor_entry = None;
-        for entity in &results {
-            if entity.entity.id == neighbor.id {
-                neighbor_entry = Some(entity.clone());
-            }
-        }
-
-        let neighbor_entry =
-            neighbor_entry.expect("Graph-enriched neighbor should appear in results");
-
-        assert!(
-            neighbor_entry.score > 0.2,
-            "Graph-enriched entity should have a meaningful fused score"
-        );
-        assert!(
-            neighbor_entry
-                .chunks
-                .iter()
-                .all(|chunk| chunk.chunk.source_id == neighbor.source_id),
-            "Neighbor entity should surface its own supporting chunks"
-        );
-    }
-}

composite-retrieval/src/pipeline/config.rs

@@ -1,67 +0,0 @@
-use serde::{Deserialize, Serialize};
-
-/// Tunable parameters that govern each retrieval stage.
-#[derive(Debug, Clone, Serialize, Deserialize)]
-pub struct RetrievalTuning {
-    pub entity_vector_take: usize,
-    pub chunk_vector_take: usize,
-    pub entity_fts_take: usize,
-    pub chunk_fts_take: usize,
-    pub score_threshold: f32,
-    pub fallback_min_results: usize,
-    pub token_budget_estimate: usize,
-    pub avg_chars_per_token: usize,
-    pub max_chunks_per_entity: usize,
-    pub graph_traversal_seed_limit: usize,
-    pub graph_neighbor_limit: usize,
-    pub graph_score_decay: f32,
-    pub graph_seed_min_score: f32,
-    pub graph_vector_inheritance: f32,
-    pub rerank_blend_weight: f32,
-    pub rerank_scores_only: bool,
-    pub rerank_keep_top: usize,
-}
-
-impl Default for RetrievalTuning {
-    fn default() -> Self {
-        Self {
-            entity_vector_take: 15,
-            chunk_vector_take: 20,
-            entity_fts_take: 10,
-            chunk_fts_take: 20,
-            score_threshold: 0.35,
-            fallback_min_results: 10,
-            token_budget_estimate: 2800,
-            avg_chars_per_token: 4,
-            max_chunks_per_entity: 4,
-            graph_traversal_seed_limit: 5,
-            graph_neighbor_limit: 6,
-            graph_score_decay: 0.75,
-            graph_seed_min_score: 0.4,
-            graph_vector_inheritance: 0.6,
-            rerank_blend_weight: 0.65,
-            rerank_scores_only: false,
-            rerank_keep_top: 8,
-        }
-    }
-}
-
-/// Wrapper containing tuning plus future flags for per-request overrides.
-#[derive(Debug, Clone)]
-pub struct RetrievalConfig {
-    pub tuning: RetrievalTuning,
-}
-
-impl RetrievalConfig {
-    pub fn new(tuning: RetrievalTuning) -> Self {
-        Self { tuning }
-    }
-}
-
-impl Default for RetrievalConfig {
-    fn default() -> Self {
-        Self {
-            tuning: RetrievalTuning::default(),
-        }
-    }
-}

composite-retrieval/src/pipeline/mod.rs

@@ -1,106 +0,0 @@
-mod config;
-mod stages;
-mod state;
-
-pub use config::{RetrievalConfig, RetrievalTuning};
-
-use crate::{reranking::RerankerLease, RetrievedEntity};
-use async_openai::Client;
-use common::{error::AppError, storage::db::SurrealDbClient};
-use tracing::info;
-
-/// Drives the retrieval pipeline from embedding through final assembly.
-pub async fn run_pipeline(
-    db_client: &SurrealDbClient,
-    openai_client: &Client<async_openai::config::OpenAIConfig>,
-    input_text: &str,
-    user_id: &str,
-    config: RetrievalConfig,
-    reranker: Option<RerankerLease>,
-) -> Result<Vec<RetrievedEntity>, AppError> {
-    let machine = state::ready();
-    let input_chars = input_text.chars().count();
-    let input_preview: String = input_text.chars().take(120).collect();
-    let input_preview_clean = input_preview.replace('\n', " ");
-    let preview_len = input_preview_clean.chars().count();
-    info!(
-        %user_id,
-        input_chars,
-        preview_truncated = input_chars > preview_len,
-        preview = %input_preview_clean,
-        "Starting ingestion retrieval pipeline"
-    );
-    let mut ctx = stages::PipelineContext::new(
-        db_client,
-        openai_client,
-        input_text.to_owned(),
-        user_id.to_owned(),
-        config,
-        reranker,
-    );
-    let machine = stages::embed(machine, &mut ctx).await?;
-    let machine = stages::collect_candidates(machine, &mut ctx).await?;
-    let machine = stages::expand_graph(machine, &mut ctx).await?;
-    let machine = stages::attach_chunks(machine, &mut ctx).await?;
-    let machine = stages::rerank(machine, &mut ctx).await?;
-    let results = stages::assemble(machine, &mut ctx)?;
-
-    Ok(results)
-}
-
-#[cfg(test)]
-pub async fn run_pipeline_with_embedding(
-    db_client: &SurrealDbClient,
-    openai_client: &Client<async_openai::config::OpenAIConfig>,
-    query_embedding: Vec<f32>,
-    input_text: &str,
-    user_id: &str,
-    config: RetrievalConfig,
-    reranker: Option<RerankerLease>,
-) -> Result<Vec<RetrievedEntity>, AppError> {
-    let machine = state::ready();
-    let mut ctx = stages::PipelineContext::with_embedding(
-        db_client,
-        openai_client,
-        query_embedding,
-        input_text.to_owned(),
-        user_id.to_owned(),
-        config,
-        reranker,
-    );
-    let machine = stages::embed(machine, &mut ctx).await?;
-    let machine = stages::collect_candidates(machine, &mut ctx).await?;
-    let machine = stages::expand_graph(machine, &mut ctx).await?;
-    let machine = stages::attach_chunks(machine, &mut ctx).await?;
-    let machine = stages::rerank(machine, &mut ctx).await?;
-    let results = stages::assemble(machine, &mut ctx)?;
-
-    Ok(results)
-}
-
-/// Helper exposed for tests to convert retrieved entities into downstream prompt JSON.
-pub fn retrieved_entities_to_json(entities: &[RetrievedEntity]) -> serde_json::Value {
-    serde_json::json!(entities
-        .iter()
-        .map(|entry| {
-            serde_json::json!({
-                "KnowledgeEntity": {
-                    "id": entry.entity.id,
-                    "name": entry.entity.name,
-                    "description": entry.entity.description,
-                    "score": round_score(entry.score),
-                    "chunks": entry.chunks.iter().map(|chunk| {
-                        serde_json::json!({
-                            "score": round_score(chunk.score),
-                            "content": chunk.chunk.chunk
-                        })
-                    }).collect::<Vec<_>>()
-                }
-            })
-        })
-        .collect::<Vec<_>>())
-}
-
-fn round_score(value: f32) -> f64 {
-    (f64::from(value) * 1000.0).round() / 1000.0
-}

composite-retrieval/src/pipeline/stages/mod.rs

@@ -1,769 +0,0 @@
-use async_openai::Client;
-use common::{
-    error::AppError,
-    storage::{
-        db::SurrealDbClient,
-        types::{knowledge_entity::KnowledgeEntity, text_chunk::TextChunk, StoredObject},
-    },
-    utils::embedding::generate_embedding,
-};
-use fastembed::RerankResult;
-use futures::{stream::FuturesUnordered, StreamExt};
-use state_machines::core::GuardError;
-use std::collections::{HashMap, HashSet};
-use tracing::{debug, instrument, warn};
-
-use crate::{
-    fts::find_items_by_fts,
-    graph::{find_entities_by_relationship_by_id, find_entities_by_source_ids},
-    reranking::RerankerLease,
-    scoring::{
-        clamp_unit, fuse_scores, merge_scored_by_id, min_max_normalize, sort_by_fused_desc,
-        FusionWeights, Scored,
-    },
-    vector::find_items_by_vector_similarity_with_embedding,
-    RetrievedChunk, RetrievedEntity,
-};
-
-use super::{
-    config::RetrievalConfig,
-    state::{
-        CandidatesLoaded, ChunksAttached, Embedded, GraphExpanded, HybridRetrievalMachine, Ready,
-        Reranked,
-    },
-};
-
-pub struct PipelineContext<'a> {
-    pub db_client: &'a SurrealDbClient,
-    pub openai_client: &'a Client<async_openai::config::OpenAIConfig>,
-    pub input_text: String,
-    pub user_id: String,
-    pub config: RetrievalConfig,
-    pub query_embedding: Option<Vec<f32>>,
-    pub entity_candidates: HashMap<String, Scored<KnowledgeEntity>>,
-    pub chunk_candidates: HashMap<String, Scored<TextChunk>>,
-    pub filtered_entities: Vec<Scored<KnowledgeEntity>>,
-    pub chunk_values: Vec<Scored<TextChunk>>,
-    pub reranker: Option<RerankerLease>,
-}
-
-impl<'a> PipelineContext<'a> {
-    pub fn new(
-        db_client: &'a SurrealDbClient,
-        openai_client: &'a Client<async_openai::config::OpenAIConfig>,
-        input_text: String,
-        user_id: String,
-        config: RetrievalConfig,
-        reranker: Option<RerankerLease>,
-    ) -> Self {
-        Self {
-            db_client,
-            openai_client,
-            input_text,
-            user_id,
-            config,
-            query_embedding: None,
-            entity_candidates: HashMap::new(),
-            chunk_candidates: HashMap::new(),
-            filtered_entities: Vec::new(),
-            chunk_values: Vec::new(),
-            reranker,
-        }
-    }
-
-    #[cfg(test)]
-    pub fn with_embedding(
-        db_client: &'a SurrealDbClient,
-        openai_client: &'a Client<async_openai::config::OpenAIConfig>,
-        query_embedding: Vec<f32>,
-        input_text: String,
-        user_id: String,
-        config: RetrievalConfig,
-        reranker: Option<RerankerLease>,
-    ) -> Self {
-        let mut ctx = Self::new(
-            db_client,
-            openai_client,
-            input_text,
-            user_id,
-            config,
-            reranker,
-        );
-        ctx.query_embedding = Some(query_embedding);
-        ctx
-    }
-
-    fn ensure_embedding(&self) -> Result<&Vec<f32>, AppError> {
-        self.query_embedding.as_ref().ok_or_else(|| {
-            AppError::InternalError(
-                "query embedding missing before candidate collection".to_string(),
-            )
-        })
-    }
-}
-
-#[instrument(level = "trace", skip_all)]
-pub async fn embed(
-    machine: HybridRetrievalMachine<(), Ready>,
-    ctx: &mut PipelineContext<'_>,
-) -> Result<HybridRetrievalMachine<(), Embedded>, AppError> {
-    let embedding_cached = ctx.query_embedding.is_some();
-    if embedding_cached {
-        debug!("Reusing cached query embedding for hybrid retrieval");
-    } else {
-        debug!("Generating query embedding for hybrid retrieval");
-        let embedding =
-            generate_embedding(ctx.openai_client, &ctx.input_text, ctx.db_client).await?;
-        ctx.query_embedding = Some(embedding);
-    }
-
-    machine
-        .embed()
-        .map_err(|(_, guard)| map_guard_error("embed", guard))
-}
-
-#[instrument(level = "trace", skip_all)]
-pub async fn collect_candidates(
-    machine: HybridRetrievalMachine<(), Embedded>,
-    ctx: &mut PipelineContext<'_>,
-) -> Result<HybridRetrievalMachine<(), CandidatesLoaded>, AppError> {
-    debug!("Collecting initial candidates via vector and FTS search");
-    let embedding = ctx.ensure_embedding()?.clone();
-    let tuning = &ctx.config.tuning;
-
-    let weights = FusionWeights::default();
-
-    let (vector_entities, vector_chunks, mut fts_entities, mut fts_chunks) = tokio::try_join!(
-        find_items_by_vector_similarity_with_embedding(
-            tuning.entity_vector_take,
-            embedding.clone(),
-            ctx.db_client,
-            "knowledge_entity",
-            &ctx.user_id,
-        ),
-        find_items_by_vector_similarity_with_embedding(
-            tuning.chunk_vector_take,
-            embedding,
-            ctx.db_client,
-            "text_chunk",
-            &ctx.user_id,
-        ),
-        find_items_by_fts(
-            tuning.entity_fts_take,
-            &ctx.input_text,
-            ctx.db_client,
-            "knowledge_entity",
-            &ctx.user_id,
-        ),
-        find_items_by_fts(
-            tuning.chunk_fts_take,
-            &ctx.input_text,
-            ctx.db_client,
-            "text_chunk",
-            &ctx.user_id
-        ),
-    )?;
-
-    debug!(
-        vector_entities = vector_entities.len(),
-        vector_chunks = vector_chunks.len(),
-        fts_entities = fts_entities.len(),
-        fts_chunks = fts_chunks.len(),
-        "Hybrid retrieval initial candidate counts"
-    );
-
-    normalize_fts_scores(&mut fts_entities);
-    normalize_fts_scores(&mut fts_chunks);
-
-    merge_scored_by_id(&mut ctx.entity_candidates, vector_entities);
-    merge_scored_by_id(&mut ctx.entity_candidates, fts_entities);
-    merge_scored_by_id(&mut ctx.chunk_candidates, vector_chunks);
-    merge_scored_by_id(&mut ctx.chunk_candidates, fts_chunks);
-
-    apply_fusion(&mut ctx.entity_candidates, weights);
-    apply_fusion(&mut ctx.chunk_candidates, weights);
-
-    machine
-        .collect_candidates()
-        .map_err(|(_, guard)| map_guard_error("collect_candidates", guard))
-}
-
-#[instrument(level = "trace", skip_all)]
-pub async fn expand_graph(
-    machine: HybridRetrievalMachine<(), CandidatesLoaded>,
-    ctx: &mut PipelineContext<'_>,
-) -> Result<HybridRetrievalMachine<(), GraphExpanded>, AppError> {
-    debug!("Expanding candidates using graph relationships");
-    let tuning = &ctx.config.tuning;
-    let weights = FusionWeights::default();
-
-    if ctx.entity_candidates.is_empty() {
-        return machine
-            .expand_graph()
-            .map_err(|(_, guard)| map_guard_error("expand_graph", guard));
-    }
-
-    let graph_seeds = seeds_from_candidates(
-        &ctx.entity_candidates,
-        tuning.graph_seed_min_score,
-        tuning.graph_traversal_seed_limit,
-    );
-
-    if graph_seeds.is_empty() {
-        return machine
-            .expand_graph()
-            .map_err(|(_, guard)| map_guard_error("expand_graph", guard));
-    }
-
-    let mut futures = FuturesUnordered::new();
-    for seed in graph_seeds {
-        let db = ctx.db_client;
-        let user = ctx.user_id.clone();
-        futures.push(async move {
-            let neighbors = find_entities_by_relationship_by_id(
-                db,
-                &seed.id,
-                &user,
-                tuning.graph_neighbor_limit,
-            )
-            .await;
-            (seed, neighbors)
-        });
-    }
-
-    while let Some((seed, neighbors_result)) = futures.next().await {
-        let neighbors = neighbors_result.map_err(AppError::from)?;
-        if neighbors.is_empty() {
-            continue;
-        }
-
-        for neighbor in neighbors {
-            if neighbor.id == seed.id {
-                continue;
-            }
-
-            let graph_score = clamp_unit(seed.fused * tuning.graph_score_decay);
-            let entry = ctx
-                .entity_candidates
-                .entry(neighbor.id.clone())
-                .or_insert_with(|| Scored::new(neighbor.clone()));
-
-            entry.item = neighbor;
-
-            let inherited_vector = clamp_unit(graph_score * tuning.graph_vector_inheritance);
-            let vector_existing = entry.scores.vector.unwrap_or(0.0);
-            if inherited_vector > vector_existing {
-                entry.scores.vector = Some(inherited_vector);
-            }
-
-            let existing_graph = entry.scores.graph.unwrap_or(f32::MIN);
-            if graph_score > existing_graph || entry.scores.graph.is_none() {
-                entry.scores.graph = Some(graph_score);
-            }
-
-            let fused = fuse_scores(&entry.scores, weights);
-            entry.update_fused(fused);
-        }
-    }
-
-    machine
-        .expand_graph()
-        .map_err(|(_, guard)| map_guard_error("expand_graph", guard))
-}
-
-#[instrument(level = "trace", skip_all)]
-pub async fn attach_chunks(
-    machine: HybridRetrievalMachine<(), GraphExpanded>,
-    ctx: &mut PipelineContext<'_>,
-) -> Result<HybridRetrievalMachine<(), ChunksAttached>, AppError> {
-    debug!("Attaching chunks to surviving entities");
-    let tuning = &ctx.config.tuning;
-    let weights = FusionWeights::default();
-
-    let chunk_by_source = group_chunks_by_source(&ctx.chunk_candidates);
-
-    backfill_entities_from_chunks(
-        &mut ctx.entity_candidates,
-        &chunk_by_source,
-        ctx.db_client,
-        &ctx.user_id,
-        weights,
-    )
-    .await?;
-
-    boost_entities_with_chunks(&mut ctx.entity_candidates, &chunk_by_source, weights);
-
-    let mut entity_results: Vec<Scored<KnowledgeEntity>> =
-        ctx.entity_candidates.values().cloned().collect();
-    sort_by_fused_desc(&mut entity_results);
-
-    let mut filtered_entities: Vec<Scored<KnowledgeEntity>> = entity_results
-        .iter()
-        .filter(|candidate| candidate.fused >= tuning.score_threshold)
-        .cloned()
-        .collect();
-
-    if filtered_entities.len() < tuning.fallback_min_results {
-        filtered_entities = entity_results
-            .into_iter()
-            .take(tuning.fallback_min_results)
-            .collect();
-    }
-
-    ctx.filtered_entities = filtered_entities;
-
-    let mut chunk_results: Vec<Scored<TextChunk>> =
-        ctx.chunk_candidates.values().cloned().collect();
-    sort_by_fused_desc(&mut chunk_results);
-
-    let mut chunk_by_id: HashMap<String, Scored<TextChunk>> = HashMap::new();
-    for chunk in chunk_results {
-        chunk_by_id.insert(chunk.item.id.clone(), chunk);
-    }
-
-    enrich_chunks_from_entities(
-        &mut chunk_by_id,
-        &ctx.filtered_entities,
-        ctx.db_client,
-        &ctx.user_id,
-        weights,
-    )
-    .await?;
-
-    let mut chunk_values: Vec<Scored<TextChunk>> = chunk_by_id.into_values().collect();
-    sort_by_fused_desc(&mut chunk_values);
-
-    ctx.chunk_values = chunk_values;
-
-    machine
-        .attach_chunks()
-        .map_err(|(_, guard)| map_guard_error("attach_chunks", guard))
-}
-
-#[instrument(level = "trace", skip_all)]
-pub async fn rerank(
-    machine: HybridRetrievalMachine<(), ChunksAttached>,
-    ctx: &mut PipelineContext<'_>,
-) -> Result<HybridRetrievalMachine<(), Reranked>, AppError> {
-    let mut applied = false;
-
-    if let Some(reranker) = ctx.reranker.as_ref() {
-        if ctx.filtered_entities.len() > 1 {
-            let documents = build_rerank_documents(ctx, ctx.config.tuning.max_chunks_per_entity);
-
-            if documents.len() > 1 {
-                match reranker.rerank(&ctx.input_text, documents).await {
-                    Ok(results) if !results.is_empty() => {
-                        apply_rerank_results(ctx, results);
-                        applied = true;
-                    }
-                    Ok(_) => {
-                        debug!("Reranker returned no results; retaining original ordering");
-                    }
-                    Err(err) => {
-                        warn!(
-                            error = %err,
-                            "Reranking failed; continuing with original ordering"
-                        );
-                    }
-                }
-            } else {
-                debug!(
-                    document_count = documents.len(),
-                    "Skipping reranking stage; insufficient document context"
-                );
-            }
-        } else {
-            debug!("Skipping reranking stage; less than two entities available");
-        }
-    } else {
-        debug!("No reranker lease provided; skipping reranking stage");
-    }
-
-    if applied {
-        debug!("Applied reranking adjustments to candidate ordering");
-    }
-
-    machine
-        .rerank()
-        .map_err(|(_, guard)| map_guard_error("rerank", guard))
-}
-
-#[instrument(level = "trace", skip_all)]
-pub fn assemble(
-    machine: HybridRetrievalMachine<(), Reranked>,
-    ctx: &mut PipelineContext<'_>,
-) -> Result<Vec<RetrievedEntity>, AppError> {
-    debug!("Assembling final retrieved entities");
-    let tuning = &ctx.config.tuning;
-
-    let mut chunk_by_source: HashMap<String, Vec<Scored<TextChunk>>> = HashMap::new();
-    for chunk in ctx.chunk_values.drain(..) {
-        chunk_by_source
-            .entry(chunk.item.source_id.clone())
-            .or_default()
-            .push(chunk);
-    }
-
-    for chunk_list in chunk_by_source.values_mut() {
-        sort_by_fused_desc(chunk_list);
-    }
-
-    let mut token_budget_remaining = tuning.token_budget_estimate;
-    let mut results = Vec::new();
-
-    for entity in &ctx.filtered_entities {
-        let mut selected_chunks = Vec::new();
-        if let Some(candidates) = chunk_by_source.get_mut(&entity.item.source_id) {
-            let mut per_entity_count = 0;
-            candidates.sort_by(|a, b| {
-                b.fused
-                    .partial_cmp(&a.fused)
-                    .unwrap_or(std::cmp::Ordering::Equal)
-            });
-
-            for candidate in candidates.iter() {
-                if per_entity_count >= tuning.max_chunks_per_entity {
-                    break;
-                }
-                let estimated_tokens =
-                    estimate_tokens(&candidate.item.chunk, tuning.avg_chars_per_token);
-                if estimated_tokens > token_budget_remaining {
-                    continue;
-                }
-
-                token_budget_remaining = token_budget_remaining.saturating_sub(estimated_tokens);
-                per_entity_count += 1;
-
-                selected_chunks.push(RetrievedChunk {
-                    chunk: candidate.item.clone(),
-                    score: candidate.fused,
-                });
-            }
-        }
-
-        results.push(RetrievedEntity {
-            entity: entity.item.clone(),
-            score: entity.fused,
-            chunks: selected_chunks,
-        });
-
-        if token_budget_remaining == 0 {
-            break;
-        }
-    }
-
-    machine
-        .assemble()
-        .map_err(|(_, guard)| map_guard_error("assemble", guard))?;
-    Ok(results)
-}
-
-fn map_guard_error(stage: &'static str, err: GuardError) -> AppError {
-    AppError::InternalError(format!(
-        "state machine guard '{stage}' failed: guard={}, event={}, kind={:?}",
-        err.guard, err.event, err.kind
-    ))
-}
-fn normalize_fts_scores<T>(results: &mut [Scored<T>]) {
-    let raw_scores: Vec<f32> = results
-        .iter()
-        .map(|candidate| candidate.scores.fts.unwrap_or(0.0))
-        .collect();
-
-    let normalized = min_max_normalize(&raw_scores);
-    for (candidate, normalized_score) in results.iter_mut().zip(normalized.into_iter()) {
-        candidate.scores.fts = Some(normalized_score);
-        candidate.update_fused(0.0);
-    }
-}
-
-fn apply_fusion<T>(candidates: &mut HashMap<String, Scored<T>>, weights: FusionWeights)
-where
-    T: StoredObject,
-{
-    for candidate in candidates.values_mut() {
-        let fused = fuse_scores(&candidate.scores, weights);
-        candidate.update_fused(fused);
-    }
-}
-
-fn group_chunks_by_source(
-    chunks: &HashMap<String, Scored<TextChunk>>,
-) -> HashMap<String, Vec<Scored<TextChunk>>> {
-    let mut by_source: HashMap<String, Vec<Scored<TextChunk>>> = HashMap::new();
-
-    for chunk in chunks.values() {
-        by_source
-            .entry(chunk.item.source_id.clone())
-            .or_default()
-            .push(chunk.clone());
-    }
-    by_source
-}
-
-async fn backfill_entities_from_chunks(
-    entity_candidates: &mut HashMap<String, Scored<KnowledgeEntity>>,
-    chunk_by_source: &HashMap<String, Vec<Scored<TextChunk>>>,
-    db_client: &SurrealDbClient,
-    user_id: &str,
-    weights: FusionWeights,
-) -> Result<(), AppError> {
-    let mut missing_sources = Vec::new();
-
-    for source_id in chunk_by_source.keys() {
-        if !entity_candidates
-            .values()
-            .any(|entity| entity.item.source_id == *source_id)
-        {
-            missing_sources.push(source_id.clone());
-        }
-    }
-
-    if missing_sources.is_empty() {
-        return Ok(());
-    }
-
-    let related_entities: Vec<KnowledgeEntity> = find_entities_by_source_ids(
-        missing_sources.clone(),
-        "knowledge_entity",
-        user_id,
-        db_client,
-    )
-    .await
-    .unwrap_or_default();
-
-    if related_entities.is_empty() {
-        warn!("expected related entities for missing chunk sources, but none were found");
-    }
-
-    for entity in related_entities {
-        if let Some(chunks) = chunk_by_source.get(&entity.source_id) {
-            let best_chunk_score = chunks
-                .iter()
-                .map(|chunk| chunk.fused)
-                .fold(0.0f32, f32::max);
-
-            let mut scored = Scored::new(entity.clone()).with_vector_score(best_chunk_score);
-            let fused = fuse_scores(&scored.scores, weights);
-            scored.update_fused(fused);
-            entity_candidates.insert(entity.id.clone(), scored);
-        }
-    }
-
-    Ok(())
-}
-
-fn boost_entities_with_chunks(
-    entity_candidates: &mut HashMap<String, Scored<KnowledgeEntity>>,
-    chunk_by_source: &HashMap<String, Vec<Scored<TextChunk>>>,
-    weights: FusionWeights,
-) {
-    for entity in entity_candidates.values_mut() {
-        if let Some(chunks) = chunk_by_source.get(&entity.item.source_id) {
-            let best_chunk_score = chunks
-                .iter()
-                .map(|chunk| chunk.fused)
-                .fold(0.0f32, f32::max);
-
-            if best_chunk_score > 0.0 {
-                let boosted = entity.scores.vector.unwrap_or(0.0).max(best_chunk_score);
-                entity.scores.vector = Some(boosted);
-                let fused = fuse_scores(&entity.scores, weights);
-                entity.update_fused(fused);
-            }
-        }
-    }
-}
-
-async fn enrich_chunks_from_entities(
-    chunk_candidates: &mut HashMap<String, Scored<TextChunk>>,
-    entities: &[Scored<KnowledgeEntity>],
-    db_client: &SurrealDbClient,
-    user_id: &str,
-    weights: FusionWeights,
-) -> Result<(), AppError> {
-    let mut source_ids: HashSet<String> = HashSet::new();
-    for entity in entities {
-        source_ids.insert(entity.item.source_id.clone());
-    }
-
-    if source_ids.is_empty() {
-        return Ok(());
-    }
-
-    let chunks = find_entities_by_source_ids::<TextChunk>(
-        source_ids.into_iter().collect(),
-        "text_chunk",
-        user_id,
-        db_client,
-    )
-    .await?;
-
-    let mut entity_score_lookup: HashMap<String, f32> = HashMap::new();
-    for entity in entities {
-        entity_score_lookup.insert(entity.item.source_id.clone(), entity.fused);
-    }
-
-    for chunk in chunks {
-        let entry = chunk_candidates
-            .entry(chunk.id.clone())
-            .or_insert_with(|| Scored::new(chunk.clone()).with_vector_score(0.0));
-
-        let entity_score = entity_score_lookup
-            .get(&chunk.source_id)
-            .copied()
-            .unwrap_or(0.0);
-
-        entry.scores.vector = Some(entry.scores.vector.unwrap_or(0.0).max(entity_score * 0.8));
-        let fused = fuse_scores(&entry.scores, weights);
-        entry.update_fused(fused);
-        entry.item = chunk;
-    }
-
-    Ok(())
-}
-
-fn build_rerank_documents(ctx: &PipelineContext<'_>, max_chunks_per_entity: usize) -> Vec<String> {
-    if ctx.filtered_entities.is_empty() {
-        return Vec::new();
-    }
-
-    let mut chunk_by_source: HashMap<&str, Vec<&Scored<TextChunk>>> = HashMap::new();
-    for chunk in &ctx.chunk_values {
-        chunk_by_source
-            .entry(chunk.item.source_id.as_str())
-            .or_default()
-            .push(chunk);
-    }
-
-    ctx.filtered_entities
-        .iter()
-        .map(|entity| {
-            let mut doc = format!(
-                "Name: {}\nType: {:?}\nDescription: {}\n",
-                entity.item.name, entity.item.entity_type, entity.item.description
-            );
-
-            if let Some(chunks) = chunk_by_source.get(entity.item.source_id.as_str()) {
-                let mut chunk_refs = chunks.clone();
-                chunk_refs.sort_by(|a, b| {
-                    b.fused
-                        .partial_cmp(&a.fused)
-                        .unwrap_or(std::cmp::Ordering::Equal)
-                });
-
-                let mut header_added = false;
-                for chunk in chunk_refs.into_iter().take(max_chunks_per_entity.max(1)) {
-                    let snippet = chunk.item.chunk.trim();
-                    if snippet.is_empty() {
-                        continue;
-                    }
-                    if !header_added {
-                        doc.push_str("Chunks:\n");
-                        header_added = true;
-                    }
-                    doc.push_str("- ");
-                    doc.push_str(snippet);
-                    doc.push('\n');
-                }
-            }
-
-            doc
-        })
-        .collect()
-}
-
-fn apply_rerank_results(ctx: &mut PipelineContext<'_>, results: Vec<RerankResult>) {
-    if results.is_empty() || ctx.filtered_entities.is_empty() {
-        return;
-    }
-
-    let mut remaining: Vec<Option<Scored<KnowledgeEntity>>> =
-        std::mem::take(&mut ctx.filtered_entities)
-            .into_iter()
-            .map(Some)
-            .collect();
-
-    let raw_scores: Vec<f32> = results.iter().map(|r| r.score).collect();
-    let normalized_scores = min_max_normalize(&raw_scores);
-
-    let use_only = ctx.config.tuning.rerank_scores_only;
-    let blend = if use_only {
-        1.0
-    } else {
-        clamp_unit(ctx.config.tuning.rerank_blend_weight)
-    };
-    let mut reranked: Vec<Scored<KnowledgeEntity>> = Vec::with_capacity(remaining.len());
-    for (result, normalized) in results.into_iter().zip(normalized_scores.into_iter()) {
-        if let Some(slot) = remaining.get_mut(result.index) {
-            if let Some(mut candidate) = slot.take() {
-                let original = candidate.fused;
-                let blended = if use_only {
-                    clamp_unit(normalized)
-                } else {
-                    clamp_unit(original * (1.0 - blend) + normalized * blend)
-                };
-                candidate.update_fused(blended);
-                reranked.push(candidate);
-            }
-        } else {
-            warn!(
-                result_index = result.index,
-                "Reranker returned out-of-range index; skipping"
-            );
-        }
-        if reranked.len() == remaining.len() {
-            break;
-        }
-    }
-
-    for slot in remaining.into_iter() {
-        if let Some(candidate) = slot {
-            reranked.push(candidate);
-        }
-    }
-
-    ctx.filtered_entities = reranked;
-    let keep_top = ctx.config.tuning.rerank_keep_top;
-    if keep_top > 0 && ctx.filtered_entities.len() > keep_top {
-        ctx.filtered_entities.truncate(keep_top);
-    }
-}
-
-fn estimate_tokens(text: &str, avg_chars_per_token: usize) -> usize {
-    let chars = text.chars().count().max(1);
-    (chars / avg_chars_per_token).max(1)
-}
-
-#[derive(Clone)]
-struct GraphSeed {
-    id: String,
-    fused: f32,
-}
-
-fn seeds_from_candidates(
-    entity_candidates: &HashMap<String, Scored<KnowledgeEntity>>,
-    min_score: f32,
-    limit: usize,
-) -> Vec<GraphSeed> {
-    let mut seeds: Vec<GraphSeed> = entity_candidates
-        .values()
-        .filter(|entity| entity.fused >= min_score)
-        .map(|entity| GraphSeed {
-            id: entity.item.id.clone(),
-            fused: entity.fused,
-        })
-        .collect();
-
-    seeds.sort_by(|a, b| {
-        b.fused
-            .partial_cmp(&a.fused)
-            .unwrap_or(std::cmp::Ordering::Equal)
-    });
-    if seeds.len() > limit {
-        seeds.truncate(limit);
-    }
-
-    seeds
-}

composite-retrieval/src/pipeline/state.rs

@@ -1,27 +0,0 @@
-use state_machines::state_machine;
-
-state_machine! {
-    name: HybridRetrievalMachine,
-    state: HybridRetrievalState,
-    initial: Ready,
-    states: [Ready, Embedded, CandidatesLoaded, GraphExpanded, ChunksAttached, Reranked, Completed, Failed],
-    events {
-        embed { transition: { from: Ready, to: Embedded } }
-        collect_candidates { transition: { from: Embedded, to: CandidatesLoaded } }
-        expand_graph { transition: { from: CandidatesLoaded, to: GraphExpanded } }
-        attach_chunks { transition: { from: GraphExpanded, to: ChunksAttached } }
-        rerank { transition: { from: ChunksAttached, to: Reranked } }
-        assemble { transition: { from: Reranked, to: Completed } }
-        abort {
-            transition: { from: Ready, to: Failed }
-            transition: { from: CandidatesLoaded, to: Failed }
-            transition: { from: GraphExpanded, to: Failed }
-            transition: { from: ChunksAttached, to: Failed }
-            transition: { from: Reranked, to: Failed }
-        }
-    }
-}
-
-pub fn ready() -> HybridRetrievalMachine<(), Ready> {
-    HybridRetrievalMachine::new(())
-}

composite-retrieval/src/scoring.rs

@@ -1,183 +0,0 @@
-use std::cmp::Ordering;
-
-use common::storage::types::StoredObject;
-
-/// Holds optional subscores gathered from different retrieval signals.
-#[derive(Debug, Clone, Copy, Default)]
-pub struct Scores {
-    pub fts: Option<f32>,
-    pub vector: Option<f32>,
-    pub graph: Option<f32>,
-}
-
-/// Generic wrapper combining an item with its accumulated retrieval scores.
-#[derive(Debug, Clone)]
-pub struct Scored<T> {
-    pub item: T,
-    pub scores: Scores,
-    pub fused: f32,
-}
-
-impl<T> Scored<T> {
-    pub fn new(item: T) -> Self {
-        Self {
-            item,
-            scores: Scores::default(),
-            fused: 0.0,
-        }
-    }
-
-    pub const fn with_vector_score(mut self, score: f32) -> Self {
-        self.scores.vector = Some(score);
-        self
-    }
-
-    pub const fn with_fts_score(mut self, score: f32) -> Self {
-        self.scores.fts = Some(score);
-        self
-    }
-
-    pub const fn with_graph_score(mut self, score: f32) -> Self {
-        self.scores.graph = Some(score);
-        self
-    }
-
-    pub const fn update_fused(&mut self, fused: f32) {
-        self.fused = fused;
-    }
-}
-
-/// Weights used for linear score fusion.
-#[derive(Debug, Clone, Copy)]
-pub struct FusionWeights {
-    pub vector: f32,
-    pub fts: f32,
-    pub graph: f32,
-    pub multi_bonus: f32,
-}
-
-impl Default for FusionWeights {
-    fn default() -> Self {
-        Self {
-            vector: 0.5,
-            fts: 0.3,
-            graph: 0.2,
-            multi_bonus: 0.02,
-        }
-    }
-}
-
-pub const fn clamp_unit(value: f32) -> f32 {
-    value.clamp(0.0, 1.0)
-}
-
-pub fn distance_to_similarity(distance: f32) -> f32 {
-    if !distance.is_finite() {
-        return 0.0;
-    }
-    clamp_unit(1.0 / (1.0 + distance.max(0.0)))
-}
-
-pub fn min_max_normalize(scores: &[f32]) -> Vec<f32> {
-    if scores.is_empty() {
-        return Vec::new();
-    }
-
-    let mut min = f32::MAX;
-    let mut max = f32::MIN;
-
-    for s in scores {
-        if !s.is_finite() {
-            continue;
-        }
-        if *s < min {
-            min = *s;
-        }
-        if *s > max {
-            max = *s;
-        }
-    }
-
-    if !min.is_finite() || !max.is_finite() {
-        return scores.iter().map(|_| 0.0).collect();
-    }
-
-    if (max - min).abs() < f32::EPSILON {
-        return vec![1.0; scores.len()];
-    }
-
-    scores
-        .iter()
-        .map(|score| {
-            if score.is_finite() {
-                clamp_unit((score - min) / (max - min))
-            } else {
-                0.0
-            }
-        })
-        .collect()
-}
-
-pub fn fuse_scores(scores: &Scores, weights: FusionWeights) -> f32 {
-    let vector = scores.vector.unwrap_or(0.0);
-    let fts = scores.fts.unwrap_or(0.0);
-    let graph = scores.graph.unwrap_or(0.0);
-
-    let mut fused = graph.mul_add(
-        weights.graph,
-        vector.mul_add(weights.vector, fts * weights.fts),
-    );
-
-    let signals_present = scores
-        .vector
-        .iter()
-        .chain(scores.fts.iter())
-        .chain(scores.graph.iter())
-        .count();
-    if signals_present >= 2 {
-        fused += weights.multi_bonus;
-    }
-
-    clamp_unit(fused)
-}
-
-pub fn merge_scored_by_id<T>(
-    target: &mut std::collections::HashMap<String, Scored<T>>,
-    incoming: Vec<Scored<T>>,
-) where
-    T: StoredObject + Clone,
-{
-    for scored in incoming {
-        let id = scored.item.get_id().to_owned();
-        target
-            .entry(id)
-            .and_modify(|existing| {
-                if let Some(score) = scored.scores.vector {
-                    existing.scores.vector = Some(score);
-                }
-                if let Some(score) = scored.scores.fts {
-                    existing.scores.fts = Some(score);
-                }
-                if let Some(score) = scored.scores.graph {
-                    existing.scores.graph = Some(score);
-                }
-            })
-            .or_insert_with(|| Scored {
-                item: scored.item.clone(),
-                scores: scored.scores,
-                fused: scored.fused,
-            });
-    }
-}
-
-pub fn sort_by_fused_desc<T>(items: &mut [Scored<T>])
-where
-    T: StoredObject,
-{
-    items.sort_by(|a, b| {
-        b.fused
-            .partial_cmp(&a.fused)
-            .unwrap_or(Ordering::Equal)
-            .then_with(|| a.item.get_id().cmp(b.item.get_id()))
-    });
-}

composite-retrieval/src/vector.rs

@@ -1,157 +0,0 @@
-use std::collections::HashMap;
-
-use common::storage::types::file_info::deserialize_flexible_id;
-use common::{
-    error::AppError,
-    storage::{db::SurrealDbClient, types::StoredObject},
-    utils::embedding::generate_embedding,
-};
-use serde::Deserialize;
-use surrealdb::sql::Thing;
-
-use crate::scoring::{clamp_unit, distance_to_similarity, Scored};
-
-/// Compares vectors and retrieves a number of items from the specified table.
-///
-/// This function generates embeddings for the input text, constructs a query to find the closest matches in the database,
-/// and then deserializes the results into the specified type `T`.
-///
-/// # Arguments
-///
-/// * `take` - The number of items to retrieve from the database.
-/// * `input_text` - The text to generate embeddings for.
-/// * `db_client` - The SurrealDB client to use for querying the database.
-/// * `table` - The table to query in the database.
-/// * `openai_client` - The OpenAI client to use for generating embeddings.
-/// * 'user_id`-  The user id of the current user.
-///
-/// # Returns
-///
-/// A vector of type `T` containing the closest matches to the input text. Returns a `ProcessingError` if an error occurs.
-///
-/// # Type Parameters
-///
-/// * `T` - The type to deserialize the query results into. Must implement `serde::Deserialize`.
-pub async fn find_items_by_vector_similarity<T>(
-    take: usize,
-    input_text: &str,
-    db_client: &SurrealDbClient,
-    table: &str,
-    openai_client: &async_openai::Client<async_openai::config::OpenAIConfig>,
-    user_id: &str,
-) -> Result<Vec<Scored<T>>, AppError>
-where
-    T: for<'de> serde::Deserialize<'de> + StoredObject,
-{
-    // Generate embeddings
-    let input_embedding = generate_embedding(openai_client, input_text, db_client).await?;
-    find_items_by_vector_similarity_with_embedding(take, input_embedding, db_client, table, user_id)
-        .await
-}
-
-#[derive(Debug, Deserialize)]
-struct DistanceRow {
-    #[serde(deserialize_with = "deserialize_flexible_id")]
-    id: String,
-    distance: Option<f32>,
-}
-
-pub async fn find_items_by_vector_similarity_with_embedding<T>(
-    take: usize,
-    query_embedding: Vec<f32>,
-    db_client: &SurrealDbClient,
-    table: &str,
-    user_id: &str,
-) -> Result<Vec<Scored<T>>, AppError>
-where
-    T: for<'de> serde::Deserialize<'de> + StoredObject,
-{
-    let embedding_literal = serde_json::to_string(&query_embedding)
-        .map_err(|err| AppError::InternalError(format!("Failed to serialize embedding: {err}")))?;
-    let closest_query = format!(
-        "SELECT id, vector::distance::knn() AS distance \
-         FROM {table} \
-         WHERE user_id = $user_id AND embedding <|{take},40|> {embedding} \
-         LIMIT $limit",
-        table = table,
-        take = take,
-        embedding = embedding_literal
-    );
-
-    let mut response = db_client
-        .query(closest_query)
-        .bind(("user_id", user_id.to_owned()))
-        .bind(("limit", take as i64))
-        .await?;
-
-    let distance_rows: Vec<DistanceRow> = response.take(0)?;
-
-    if distance_rows.is_empty() {
-        return Ok(Vec::new());
-    }
-
-    let ids: Vec<String> = distance_rows.iter().map(|row| row.id.clone()).collect();
-    let thing_ids: Vec<Thing> = ids
-        .iter()
-        .map(|id| Thing::from((table, id.as_str())))
-        .collect();
-
-    let mut items_response = db_client
-        .query("SELECT * FROM type::table($table) WHERE id IN $things AND user_id = $user_id")
-        .bind(("table", table.to_owned()))
-        .bind(("things", thing_ids.clone()))
-        .bind(("user_id", user_id.to_owned()))
-        .await?;
-
-    let items: Vec<T> = items_response.take(0)?;
-
-    let mut item_map: HashMap<String, T> = items
-        .into_iter()
-        .map(|item| (item.get_id().to_owned(), item))
-        .collect();
-
-    let mut min_distance = f32::MAX;
-    let mut max_distance = f32::MIN;
-
-    for row in &distance_rows {
-        if let Some(distance) = row.distance {
-            if distance.is_finite() {
-                if distance < min_distance {
-                    min_distance = distance;
-                }
-                if distance > max_distance {
-                    max_distance = distance;
-                }
-            }
-        }
-    }
-
-    let normalize = min_distance.is_finite()
-        && max_distance.is_finite()
-        && (max_distance - min_distance).abs() > f32::EPSILON;
-
-    let mut scored = Vec::with_capacity(distance_rows.len());
-    for row in distance_rows {
-        if let Some(item) = item_map.remove(&row.id) {
-            let similarity = row
-                .distance
-                .map(|distance| {
-                    if normalize {
-                        let span = max_distance - min_distance;
-                        if span.abs() < f32::EPSILON {
-                            1.0
-                        } else {
-                            let normalized = 1.0 - ((distance - min_distance) / span);
-                            clamp_unit(normalized)
-                        }
-                    } else {
-                        distance_to_similarity(distance)
-                    }
-                })
-                .unwrap_or_default();
-            scored.push(Scored::new(item).with_vector_score(similarity));
-        }
-    }
-
-    Ok(scored)
-}

devenv.lock

@@ -3,10 +3,10 @@
     "devenv": {
       "locked": {
         "dir": "src/modules",
-        "lastModified": 1761839147,
+        "lastModified": 1771066302,
         "owner": "cachix",
         "repo": "devenv",
-        "rev": "bb7849648b68035f6b910120252c22b28195cf54",
+        "rev": "1b355dec9bddbaddbe4966d6fc30d7aa3af8575b",
         "type": "github"
       },
       "original": {
@@ -22,10 +22,10 @@
         "rust-analyzer-src": "rust-analyzer-src"
       },
       "locked": {
-        "lastModified": 1761893049,
+        "lastModified": 1771052630,
         "owner": "nix-community",
         "repo": "fenix",
-        "rev": "c2ac9a5c0d6d16630c3b225b874bd14528d1abe6",
+        "rev": "d0555da98576b8611c25df0c208e51e9a182d95f",
         "type": "github"
       },
       "original": {
@@ -37,14 +37,14 @@
     "flake-compat": {
       "flake": false,
       "locked": {
-        "lastModified": 1761588595,
-        "owner": "edolstra",
+        "lastModified": 1767039857,
+        "owner": "NixOS",
         "repo": "flake-compat",
-        "rev": "f387cd2afec9419c8ee37694406ca490c3f34ee5",
+        "rev": "5edf11c44bc78a0d334f6334cdaf7d60d732daab",
         "type": "github"
       },
       "original": {
-        "owner": "edolstra",
+        "owner": "NixOS",
         "repo": "flake-compat",
         "type": "github"
       }
@@ -58,10 +58,10 @@
         ]
       },
       "locked": {
-        "lastModified": 1760663237,
+        "lastModified": 1770726378,
         "owner": "cachix",
         "repo": "git-hooks.nix",
-        "rev": "ca5b894d3e3e151ffc1db040b6ce4dcc75d31c37",
+        "rev": "5eaaedde414f6eb1aea8b8525c466dc37bba95ae",
         "type": "github"
       },
       "original": {
@@ -78,10 +78,10 @@
         ]
       },
       "locked": {
-        "lastModified": 1709087332,
+        "lastModified": 1762808025,
         "owner": "hercules-ci",
         "repo": "gitignore.nix",
-        "rev": "637db329424fd7e46cf4185293b9cc8c88c95394",
+        "rev": "cb5e3fdca1de58ccbc3ef53de65bd372b48f567c",
         "type": "github"
       },
       "original": {
@@ -92,10 +92,10 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1761672384,
+        "lastModified": 1771008912,
         "owner": "nixos",
         "repo": "nixpkgs",
-        "rev": "08dacfca559e1d7da38f3cf05f1f45ee9bfd213c",
+        "rev": "a82ccc39b39b621151d6732718e3e250109076fa",
         "type": "github"
       },
       "original": {
@@ -107,10 +107,10 @@
     },
     "nixpkgs_2": {
       "locked": {
-        "lastModified": 1761880412,
+        "lastModified": 1770843696,
         "owner": "nixos",
         "repo": "nixpkgs",
-        "rev": "a7fc11be66bdfb5cdde611ee5ce381c183da8386",
+        "rev": "2343bbb58f99267223bc2aac4fc9ea301a155a16",
         "type": "github"
       },
       "original": {
@@ -135,10 +135,10 @@
     "rust-analyzer-src": {
       "flake": false,
       "locked": {
-        "lastModified": 1761849405,
+        "lastModified": 1771007332,
         "owner": "rust-lang",
         "repo": "rust-analyzer",
-        "rev": "f7de8ae045a5fe80f1203c5a1c3015b05f7c3550",
+        "rev": "bbc84d335fbbd9b3099d3e40c7469ee57dbd1873",
         "type": "github"
       },
       "original": {
@@ -155,10 +155,10 @@
         ]
       },
       "locked": {
-        "lastModified": 1761878277,
+        "lastModified": 1771038269,
         "owner": "oxalica",
         "repo": "rust-overlay",
-        "rev": "6604534e44090c917db714faa58d47861657690c",
+        "rev": "d7a86c8a4df49002446737603a3e0d7ef91a9637",
         "type": "github"
       },
       "original": {

devenv.nix

@@ -10,11 +10,14 @@
   packages = [
     pkgs.openssl
     pkgs.nodejs
+    pkgs.watchman
     pkgs.vscode-langservers-extracted
     pkgs.cargo-dist
     pkgs.cargo-xwin
     pkgs.clang
     pkgs.onnxruntime
+    pkgs.cargo-watch
+    pkgs.tailwindcss_4
   ];
 
   languages.rust = {
@@ -27,9 +30,25 @@
 
   env = {
     ORT_DYLIB_PATH = "${pkgs.onnxruntime}/lib/libonnxruntime.so";
+    S3_ENDPOINT = "http://127.0.0.1:19000";
+    S3_BUCKET = "minne-tests";
+    MINNE_TEST_S3_ENDPOINT = "http://127.0.0.1:19000";
+    MINNE_TEST_S3_BUCKET = "minne-tests";
+  };
+
+  services.minio = {
+    enable = true;
+    listenAddress = "127.0.0.1:19000";
+    consoleAddress = "127.0.0.1:19001";
+    buckets = ["minne-tests"];
+    accessKey = "minioadmin";
+    secretKey = "minioadmin";
+    region = "us-east-1";
   };
 
   processes = {
     surreal_db.exec = "docker run --rm --pull always -p 8000:8000 --net=host --user $(id -u) -v $(pwd)/database:/database surrealdb/surrealdb:latest-dev start rocksdb:/database/database.db --user root_user --pass root_password";
+    server.exec = "cargo watch -x 'run --bin main'";
+    tailwind.exec = "tailwindcss --cwd html-router -i app.css -o assets/style.css --watch=always";
   };
 }

dist-workspace.toml

@@ -12,11 +12,13 @@ include = ["lib"]
 # The installers to generate for each app
 installers = []
 # Target platforms to build apps for (Rust target-triple syntax)
-targets = ["aarch64-apple-darwin", "x86_64-unknown-linux-gnu", "x86_64-pc-windows-msvc"]
+targets = ["aarch64-apple-darwin", "x86_64-apple-darwin", "x86_64-unknown-linux-gnu", "x86_64-pc-windows-msvc"]
 # Skip checking whether the specified configuration files are up to date
 allow-dirty = ["ci"]
 
 [dist.github-custom-runners]
+aarch64-apple-darwin = "macos-latest"
+x86_64-apple-darwin = "macos-15-intel"
 x86_64-unknown-linux-gnu = "ubuntu-22.04"
 x86_64-unknown-linux-musl = "ubuntu-22.04"
 x86_64-pc-windows-msvc = "windows-latest"

docs/architecture.md

@@ -0,0 +1,74 @@
+# Architecture
+
+## Tech Stack
+
+| Layer | Technology |
+|-------|------------|
+| Backend | Rust with Axum (SSR) |
+| Frontend | HTML + HTMX + minimal JS |
+| Database | SurrealDB (graph, document, vector) |
+| AI | OpenAI-compatible API |
+| Web Processing | Headless Chromium |
+
+## Crate Structure
+
+```
+minne/
+├── main/                 # Combined server + worker binary
+├── api-router/           # REST API routes
+├── html-router/          # SSR web interface
+├── ingestion-pipeline/   # Content processing pipeline
+├── retrieval-pipeline/   # Search and retrieval logic
+├── common/               # Shared types, storage, utilities
+├── evaluations/          # Benchmarking framework
+└── json-stream-parser/   # Streaming JSON utilities
+```
+
+## Process Modes
+
+| Binary | Purpose |
+|--------|---------|
+| `main` | All-in-one: serves UI and processes content |
+| `server` | UI and API only (no background processing) |
+| `worker` | Background processing only (no UI) |
+
+Split deployment is useful for scaling or resource isolation.
+
+## Data Flow
+
+```
+Content In → Ingestion Pipeline → SurrealDB
+                    ↓
+            Entity Extraction
+                    ↓
+            Embedding Generation
+                    ↓
+            Graph Relationships
+
+Query → Retrieval Pipeline → Results
+              ↓
+       Vector Search + FTS
+              ↓
+       RRF Fusion → (Optional Rerank) → Response
+```
+
+## Database Schema
+
+SurrealDB stores:
+
+- **TextContent** — Raw ingested content
+- **TextChunk** — Chunked content with embeddings
+- **KnowledgeEntity** — Extracted entities (people, concepts, etc.)
+- **KnowledgeRelationship** — Connections between entities
+- **User** — Authentication and preferences
+- **SystemSettings** — Model configuration
+
+Embeddings are stored in dedicated tables with HNSW indexes for fast vector search.
+
+## Retrieval Strategy
+
+1. **Collect candidates** — Vector similarity + full-text search
+2. **Merge ranks** — Reciprocal Rank Fusion (RRF)
+3. **Attach context** — Link chunks to parent entities
+4. **Rerank** (optional) — Cross-encoder reranking
+5. **Return** — Top-k results with metadata

docs/configuration.md

@@ -0,0 +1,117 @@
+# Configuration
+
+Minne can be configured via environment variables or a `config.yaml` file. Environment variables take precedence.
+
+## Required Settings
+
+| Variable | Description | Example |
+|----------|-------------|---------|
+| `OPENAI_API_KEY` | API key for OpenAI-compatible endpoint | `sk-...` |
+| `SURREALDB_ADDRESS` | WebSocket address of SurrealDB | `ws://127.0.0.1:8000` |
+| `SURREALDB_USERNAME` | SurrealDB username | `root_user` |
+| `SURREALDB_PASSWORD` | SurrealDB password | `root_password` |
+| `SURREALDB_DATABASE` | Database name | `minne_db` |
+| `SURREALDB_NAMESPACE` | Namespace | `minne_ns` |
+
+
+## Optional Settings
+
+| Variable | Description | Default |
+|----------|-------------|---------|
+| `HTTP_PORT` | Server port | `3000` |
+| `DATA_DIR` | Local data directory | `./data` |
+| `OPENAI_BASE_URL` | Custom AI provider URL | OpenAI default |
+| `RUST_LOG` | Logging level | `info` |
+| `STORAGE` | Storage backend (`local`, `memory`, `s3`) | `local` |
+| `PDF_INGEST_MODE` | PDF ingestion strategy (`classic`, `llm-first`) | `llm-first` |
+| `RETRIEVAL_STRATEGY` | Default retrieval strategy | - |
+| `EMBEDDING_BACKEND` | Embedding provider (`openai`, `fastembed`) | `fastembed` |
+| `FASTEMBED_CACHE_DIR` | Model cache directory | `<data_dir>/fastembed` |
+| `FASTEMBED_SHOW_DOWNLOAD_PROGRESS` | Show progress bar for model downloads | `false` |
+| `FASTEMBED_MAX_LENGTH` | Max sequence length for FastEmbed models | - |
+| `INGEST_MAX_BODY_BYTES` | Max request body size for ingest endpoints | `20000000` |
+| `INGEST_MAX_FILES` | Max files allowed per ingest request | `5` |
+| `INGEST_MAX_CONTENT_BYTES` | Max `content` field size for ingest requests | `262144` |
+| `INGEST_MAX_CONTEXT_BYTES` | Max `context` field size for ingest requests | `16384` |
+| `INGEST_MAX_CATEGORY_BYTES` | Max `category` field size for ingest requests | `128` |
+
+### S3 Storage (Optional)
+
+Used when `STORAGE` is set to `s3`.
+
+| Variable | Description | Default |
+|----------|-------------|---------|
+| `S3_BUCKET` | S3 bucket name | - |
+| `S3_ENDPOINT` | Custom endpoint (e.g. MinIO) | AWS default |
+| `S3_REGION` | AWS Region | `us-east-1` |
+| `AWS_ACCESS_KEY_ID` | Access key | - |
+| `AWS_SECRET_ACCESS_KEY` | Secret key | - |
+
+### Reranking (Optional)
+
+| Variable | Description | Default |
+|----------|-------------|---------|
+| `RERANKING_ENABLED` | Enable FastEmbed reranking | `false` |
+| `RERANKING_POOL_SIZE` | Concurrent reranker workers | - |
+
+> [!NOTE]
+> Enabling reranking downloads ~1.1 GB of model data on first startup.
+
+## Example config.yaml
+
+```yaml
+surrealdb_address: "ws://127.0.0.1:8000"
+surrealdb_username: "root_user"
+surrealdb_password: "root_password"
+surrealdb_database: "minne_db"
+surrealdb_namespace: "minne_ns"
+openai_api_key: "sk-your-key-here"
+data_dir: "./minne_data"
+http_port: 3000
+
+# New settings
+storage: "local"
+# storage: "s3"
+# s3_bucket: "my-bucket"
+# s3_endpoint: "http://localhost:9000" # Optional, for MinIO etc.
+# s3_region: "us-east-1"
+pdf_ingest_mode: "llm-first"
+embedding_backend: "fastembed"
+
+# Optional reranking
+reranking_enabled: true
+reranking_pool_size: 2
+
+# Ingest safety limits
+ingest_max_body_bytes: 20000000
+ingest_max_files: 5
+ingest_max_content_bytes: 262144
+ingest_max_context_bytes: 16384
+ingest_max_category_bytes: 128
+```
+
+## AI Provider Setup
+
+Minne works with any OpenAI-compatible API that supports structured outputs.
+
+### OpenAI (Default)
+
+Set `OPENAI_API_KEY` only. The default base URL points to OpenAI.
+
+### Ollama
+
+```bash
+OPENAI_API_KEY="ollama"
+OPENAI_BASE_URL="http://localhost:11434/v1"
+```
+
+### Other Providers
+
+Any provider exposing an OpenAI-compatible endpoint works. Set `OPENAI_BASE_URL` accordingly.
+
+## Model Selection
+
+1. Access `/admin` in your Minne instance
+2. Select models for content processing and chat
+3. **Content Processing**: Must support structured outputs
+4. **Embedding Dimensions**: Update when changing embedding models (e.g., 1536 for `text-embedding-3-small`)

docs/features.md

@@ -0,0 +1,64 @@
+# Features
+
+## Search vs Chat
+
+**Search** — Use when you know what you're looking for. Full-text search matches query terms across your content.
+
+**Chat** — Use when exploring concepts or reasoning about your knowledge. The AI analyzes your query and retrieves relevant context from your entire knowledge base.
+
+## Content Processing
+
+Minne automatically processes saved content:
+
+1. **Web scraping** extracts readable text from URLs (via headless Chrome)
+2. **Text analysis** identifies key concepts and relationships
+3. **Graph creation** builds connections between related content
+4. **Embedding generation** enables semantic search
+
+## Knowledge Graph
+
+Explore your knowledge as an interactive network:
+
+- **Manual curation** — Create entities and relationships yourself
+- **AI automation** — Let AI extract entities and discover relationships
+- **Hybrid approach** — AI suggests connections for your approval
+
+The D3-based graph visualization shows entities as nodes and relationships as edges.
+
+## Hybrid Retrieval
+
+Minne combines multiple retrieval strategies:
+
+- **Vector similarity** — Semantic matching via embeddings
+- **Full-text search** — Keyword matching with BM25
+- **Graph traversal** — Following relationships between entities
+
+Results are merged using Reciprocal Rank Fusion (RRF) for optimal relevance.
+
+## Reranking (Optional)
+
+When enabled, retrieval results are rescored with a cross-encoder model for improved relevance. Powered by [fastembed-rs](https://github.com/Anush008/fastembed-rs).
+
+**Trade-offs:**
+- Downloads ~1.1 GB of model data
+- Adds latency per query
+- Potentially improves answer quality, see [blog post](https://blog.stark.pub/posts/eval-retrieval-refactor/)
+
+Enable via `RERANKING_ENABLED=true`. See [Configuration](./configuration.md).
+
+## Multi-Format Ingestion
+
+Supported content types:
+- Plain text and notes
+- URLs (web pages)
+- PDF documents
+- Audio files
+- Images
+
+## Scratchpad
+
+Quickly capture content without committing to permanent storage. Convert to full content when ready.
+
+## iOS Shortcut
+
+Use the [Minne iOS Shortcut](https://www.icloud.com/shortcuts/e433fbd7602f4e2eaa70dca162323477) for quick content capture from your phone.

docs/installation.md

@@ -0,0 +1,67 @@
+# Installation
+
+Minne can be installed through several methods. Choose the one that best fits your setup.
+
+## Docker Compose (Recommended)
+
+The fastest way to get Minne running with all dependencies:
+
+```bash
+git clone https://github.com/perstarkse/minne.git
+cd minne
+docker compose up -d
+```
+
+The included `docker-compose.yml` handles SurrealDB and Chromium automatically.
+
+**Required:** Set your `OPENAI_API_KEY` in `docker-compose.yml` before starting.
+
+## Nix
+
+Run Minne directly with Nix (includes Chromium):
+
+```bash
+nix run 'github:perstarkse/minne#main'
+```
+
+Configure via environment variables or a `config.yaml` file. See [Configuration](./configuration.md).
+
+## Pre-built Binaries
+
+Download binaries for Windows, macOS, and Linux from [GitHub Releases](https://github.com/perstarkse/minne/releases/latest).
+
+**Requirements:**
+- SurrealDB instance (local or remote)
+- Chromium (for web scraping)
+
+## Build from Source
+
+```bash
+git clone https://github.com/perstarkse/minne.git
+cd minne
+cargo build --release --bin main
+```
+
+The binary will be at `target/release/main`.
+
+**Requirements:**
+- Rust toolchain
+- SurrealDB accessible at configured address
+- Chromium in PATH
+
+## Process Modes
+
+Minne offers flexible deployment:
+
+| Binary | Description |
+|--------|-------------|
+| `main` | Combined server + worker (recommended) |
+| `server` | Web interface and API only |
+| `worker` | Background processing only |
+
+For most users, `main` is the right choice. Split deployments are useful for resource optimization or scaling.
+
+## Next Steps
+
+- [Configuration](./configuration.md) — Environment variables and config.yaml
+- [Features](./features.md) — What Minne can do

docs/vision.md

@@ -0,0 +1,48 @@
+# Vision
+
+## The "Why" Behind Minne
+
+Personal knowledge management has always fascinated me. I wanted something that made it incredibly easy to capture content—snippets of text, URLs, media—while automatically discovering connections between ideas. But I also wanted control over my knowledge structure.
+
+Traditional tools like Logseq and Obsidian are excellent, but manual linking often becomes a hindrance. Fully automated systems sometimes miss important context or create relationships I wouldn't have chosen.
+
+Minne offers the best of both worlds: effortless capture with AI-assisted relationship discovery, but with flexibility to manually curate, edit, or override connections. Let AI handle the heavy lifting, take full control yourself, or use a hybrid approach where AI suggests and you approve.
+
+## Design Principles
+
+- **Capture should be instant** — No friction between thought and storage
+- **Connections should emerge** — AI finds relationships you might miss
+- **Control should be optional** — Automate by default, curate when it matters
+- **Privacy should be default** — Self-hosted, your data stays yours
+
+## Roadmap
+
+### Near-term
+
+- [ ] TUI frontend with system editor integration
+- [ ] Enhanced retrieval recall via improved reranking
+- [ ] Additional content type support (e-books, research papers)
+
+### Medium-term
+
+- [ ] Embedded SurrealDB option (zero-config `nix run` with just `OPENAI_API_KEY`)
+- [ ] Browser extension for seamless capture
+- [ ] Mobile-native apps
+
+### Long-term
+
+- [ ] Federated knowledge sharing (opt-in)
+- [ ] Local LLM integration (fully offline operation)
+- [ ] Plugin system for custom entity extractors
+
+## Related Projects
+
+If Minne isn't quite right for you, check out:
+
+- [Karakeep](https://github.com/karakeep-app/karakeep) (formerly Hoarder) — Excellent bookmark/read-later with AI tagging
+- [Logseq](https://logseq.com/) — Outliner-based PKM with manual linking
+- [Obsidian](https://obsidian.md/) — Markdown-based PKM with plugin ecosystem
+
+## Contributing
+
+Feature requests and contributions are welcome. Minne was built for personal use first, but the self-hosted community benefits when we share.

evaluations/Cargo.toml

@@ -0,0 +1,36 @@
+[package]
+name = "evaluations"
+version = "0.1.0"
+edition = "2021"
+
+[dependencies]
+anyhow = { workspace = true }
+async-openai = { workspace = true }
+chrono = { workspace = true }
+common = { path = "../common" }
+retrieval-pipeline = { path = "../retrieval-pipeline" }
+ingestion-pipeline = { path = "../ingestion-pipeline" }
+futures = { workspace = true }
+fastembed = { workspace = true }
+serde = { workspace = true, features = ["derive"] }
+tokio = { workspace = true, features = ["macros", "rt-multi-thread"] }
+tracing = { workspace = true }
+tracing-subscriber = { workspace = true }
+uuid = { workspace = true }
+text-splitter = { workspace = true }
+unicode-normalization = { workspace = true }
+rand = "0.8"
+sha2 = { workspace = true }
+object_store = { workspace = true }
+surrealdb = { workspace = true }
+serde_json = { workspace = true }
+async-trait = { workspace = true }
+once_cell = "1.19"
+serde_yaml = "0.9"
+criterion = "0.5"
+state-machines = { workspace = true }
+clap = { version = "4.4", features = ["derive", "env"] }
+
+[dev-dependencies]
+tempfile = { workspace = true }
+common = { path = "../common", features = ["test-utils"] }

evaluations/README.md

@@ -0,0 +1,212 @@
+# Evaluations
+
+The `evaluations` crate provides a retrieval evaluation framework for benchmarking Minne's information retrieval pipeline against standard datasets.
+
+## Quick Start
+
+```bash
+# Run SQuAD v2.0 evaluation (vector-only, recommended)
+cargo run --package evaluations -- --ingest-chunks-only
+
+# Run a specific dataset
+cargo run --package evaluations -- --dataset fiqa --ingest-chunks-only
+
+# Convert dataset only (no evaluation)
+cargo run --package evaluations -- --convert-only
+```
+
+## Prerequisites
+
+### 1. SurrealDB
+
+Start a SurrealDB instance before running evaluations:
+
+```bash
+docker-compose up -d surrealdb
+```
+
+Or using the default endpoint configuration:
+
+```bash
+surreal start --user root_user --pass root_password
+```
+
+### 2. Download Raw Datasets
+
+Raw datasets must be downloaded manually and placed in `evaluations/data/raw/`. See [Dataset Sources](#dataset-sources) below for links and formats.
+
+## Directory Structure
+
+```
+evaluations/
+├── data/
+│   ├── raw/          # Downloaded raw datasets (manual)
+│   │   ├── squad/    # SQuAD v2.0
+│   │   ├── nq-dev/   # Natural Questions
+│   │   ├── fiqa/     # BEIR: FiQA-2018
+│   │   ├── fever/    # BEIR: FEVER
+│   │   ├── hotpotqa/ # BEIR: HotpotQA
+│   │   └── ...       # Other BEIR subsets
+│   └── converted/    # Auto-generated (Minne JSON format)
+├── cache/            # Ingestion and embedding caches
+├── reports/          # Evaluation output (JSON + Markdown)
+├── manifest.yaml     # Dataset and slice definitions
+└── src/              # Evaluation source code
+```
+
+## Dataset Sources
+
+### SQuAD v2.0
+
+Download and place at `data/raw/squad/dev-v2.0.json`:
+
+```bash
+mkdir -p evaluations/data/raw/squad
+curl -L https://rajpurkar.github.io/SQuAD-explorer/dataset/dev-v2.0.json \
+  -o evaluations/data/raw/squad/dev-v2.0.json
+```
+
+### Natural Questions (NQ)
+
+Download and place at `data/raw/nq-dev/dev-all.jsonl`:
+
+```bash
+mkdir -p evaluations/data/raw/nq-dev
+# Download from Google's Natural Questions page or HuggingFace
+# File: dev-all.jsonl (simplified JSONL format)
+```
+
+Source: [Google Natural Questions](https://ai.google.com/research/NaturalQuestions)
+
+### BEIR Datasets
+
+All BEIR datasets follow the same format structure:
+
+```
+data/raw/<dataset>/
+├── corpus.jsonl      # Document corpus
+├── queries.jsonl     # Query set
+└── qrels/
+    └── test.tsv      # Relevance judgments (or dev.tsv)
+```
+
+Download datasets from the [BEIR Benchmark repository](https://github.com/beir-cellar/beir). Each dataset zip extracts to the required directory structure.
+
+| Dataset    | Directory     |
+|------------|---------------|
+| FEVER      | `fever/`      |
+| FiQA-2018  | `fiqa/`       |
+| HotpotQA   | `hotpotqa/`   |
+| NFCorpus   | `nfcorpus/`   |
+| Quora      | `quora/`      |
+| TREC-COVID | `trec-covid/` |
+| SciFact    | `scifact/`    |
+| NQ (BEIR)  | `nq/`         |
+
+Example download:
+
+```bash
+cd evaluations/data/raw
+curl -L https://public.ukp.informatik.tu-darmstadt.de/thakur/BEIR/datasets/fiqa.zip -o fiqa.zip
+unzip fiqa.zip && rm fiqa.zip
+```
+
+## Dataset Conversion
+
+Raw datasets are automatically converted to Minne's internal JSON format on first run. To force reconversion:
+
+```bash
+cargo run --package evaluations -- --force-convert
+```
+
+Converted files are saved to `data/converted/` and cached for subsequent runs.
+
+## CLI Reference
+
+### Common Options
+
+| Flag | Description | Default |
+|------|-------------|---------|
+| `--dataset <NAME>` | Dataset to evaluate | `squad-v2` |
+| `--limit <N>` | Max questions to evaluate (0 = all) | `200` |
+| `--k <N>` | Precision@k cutoff | `5` |
+| `--slice <ID>` | Use a predefined slice from manifest | — |
+| `--rerank` | Enable FastEmbed reranking stage | disabled |
+| `--embedding-backend <BE>` | `fastembed` or `hashed` | `fastembed` |
+| `--ingest-chunks-only` | Skip entity extraction, ingest only text chunks | disabled |
+
+> [!TIP]
+> Use `--ingest-chunks-only` when evaluating vector-only retrieval strategies. This skips the LLM-based entity extraction and graph generation, significantly speeding up ingestion while focusing on pure chunk-based vector search.
+
+### Available Datasets
+
+```
+squad-v2, natural-questions, beir, fever, fiqa, hotpotqa, 
+nfcorpus, quora, trec-covid, scifact, nq-beir
+```
+
+### Database Configuration
+
+| Flag | Environment | Default |
+|------|-------------|---------|
+| `--db-endpoint` | `EVAL_DB_ENDPOINT` | `ws://127.0.0.1:8000` |
+| `--db-username` | `EVAL_DB_USERNAME` | `root_user` |
+| `--db-password` | `EVAL_DB_PASSWORD` | `root_password` |
+| `--db-namespace` | `EVAL_DB_NAMESPACE` | auto-generated |
+| `--db-database` | `EVAL_DB_DATABASE` | auto-generated |
+
+### Example Runs
+
+```bash
+# Vector-only evaluation (recommended for benchmarking)
+cargo run --package evaluations -- \
+  --dataset fiqa \
+  --ingest-chunks-only \
+  --limit 200
+
+# Full FiQA evaluation with reranking
+cargo run --package evaluations -- \
+  --dataset fiqa \
+  --ingest-chunks-only \
+  --limit 500 \
+  --rerank \
+  --k 10
+
+# Use a predefined slice for reproducibility
+cargo run --package evaluations -- --slice fiqa-test-200 --ingest-chunks-only
+
+# Run the mixed BEIR benchmark
+cargo run --package evaluations -- --dataset beir --slice beir-mix-600 --ingest-chunks-only
+```
+
+## Slices
+
+Slices are predefined, reproducible subsets defined in `manifest.yaml`. Each slice specifies:
+
+- **limit**: Number of questions
+- **corpus_limit**: Maximum corpus size
+- **seed**: Fixed RNG seed for reproducibility
+
+View available slices in [manifest.yaml](./manifest.yaml).
+
+## Reports
+
+Evaluations generate reports in `reports/`:
+
+- **JSON**: Full structured results (`*-report.json`)
+- **Markdown**: Human-readable summary with sample mismatches (`*-report.md`)
+- **History**: Timestamped run history (`history/`)
+
+## Performance Tuning
+
+```bash
+# Log per-stage performance timings
+cargo run --package evaluations -- --perf-log-console
+
+# Save telemetry to file
+cargo run --package evaluations -- --perf-log-json ./perf.json
+```
+
+## License
+
+See [../LICENSE](../LICENSE).

evaluations/manifest.yaml

@@ -0,0 +1,168 @@
+default_dataset: squad-v2
+datasets:
+  - id: squad-v2
+    label: "SQuAD v2.0"
+    category: "SQuAD v2.0"
+    entity_suffix: "SQuAD"
+    source_prefix: "squad"
+    raw: "data/raw/squad/dev-v2.0.json"
+    converted: "data/converted/squad-minne.json"
+    include_unanswerable: false
+    slices:
+      - id: squad-dev-200
+        label: "SQuAD dev (200)"
+        description: "Deterministic 200-case slice for local eval"
+        limit: 200
+        corpus_limit: 2000
+        seed: 0x5eed2025
+  - id: natural-questions-dev
+    label: "Natural Questions (dev)"
+    category: "Natural Questions"
+    entity_suffix: "Natural Questions"
+    source_prefix: "nq"
+    raw: "data/raw/nq-dev/dev-all.jsonl"
+    converted: "data/converted/nq-dev-minne.json"
+    include_unanswerable: true
+    slices:
+      - id: nq-dev-200
+        label: "NQ dev (200)"
+        description: "200-case slice of the dev set"
+        limit: 200
+        corpus_limit: 2000
+        include_unanswerable: false
+        seed: 0x5eed2025
+  - id: beir
+    label: "BEIR mix"
+    category: "BEIR"
+    entity_suffix: "BEIR"
+    source_prefix: "beir"
+    raw: "data/raw/beir"
+    converted: "data/converted/beir-minne.json"
+    include_unanswerable: false
+    slices:
+      - id: beir-mix-600
+        label: "BEIR mix (600)"
+        description: "Balanced slice across FEVER, FiQA, HotpotQA, NFCorpus, Quora, TREC-COVID, SciFact, NQ-BEIR"
+        limit: 600
+        corpus_limit: 6000
+        seed: 0x5eed2025
+  - id: fever
+    label: "FEVER (BEIR)"
+    category: "FEVER"
+    entity_suffix: "FEVER"
+    source_prefix: "fever"
+    raw: "data/raw/fever"
+    converted: "data/converted/fever-minne.json"
+    include_unanswerable: false
+    slices:
+      - id: fever-test-200
+        label: "FEVER test (200)"
+        description: "200-case slice from BEIR test qrels"
+        limit: 200
+        corpus_limit: 5000
+        seed: 0x5eed2025
+  - id: fiqa
+    label: "FiQA-2018 (BEIR)"
+    category: "FiQA-2018"
+    entity_suffix: "FiQA"
+    source_prefix: "fiqa"
+    raw: "data/raw/fiqa"
+    converted: "data/converted/fiqa-minne.json"
+    include_unanswerable: false
+    slices:
+      - id: fiqa-test-200
+        label: "FiQA test (200)"
+        description: "200-case slice from BEIR test qrels"
+        limit: 200
+        corpus_limit: 5000
+        seed: 0x5eed2025
+  - id: hotpotqa
+    label: "HotpotQA (BEIR)"
+    category: "HotpotQA"
+    entity_suffix: "HotpotQA"
+    source_prefix: "hotpotqa"
+    raw: "data/raw/hotpotqa"
+    converted: "data/converted/hotpotqa-minne.json"
+    include_unanswerable: false
+    slices:
+      - id: hotpotqa-test-200
+        label: "HotpotQA test (200)"
+        description: "200-case slice from BEIR test qrels"
+        limit: 200
+        corpus_limit: 5000
+        seed: 0x5eed2025
+  - id: nfcorpus
+    label: "NFCorpus (BEIR)"
+    category: "NFCorpus"
+    entity_suffix: "NFCorpus"
+    source_prefix: "nfcorpus"
+    raw: "data/raw/nfcorpus"
+    converted: "data/converted/nfcorpus-minne.json"
+    include_unanswerable: false
+    slices:
+      - id: nfcorpus-test-200
+        label: "NFCorpus test (200)"
+        description: "200-case slice from BEIR test qrels"
+        limit: 200
+        corpus_limit: 5000
+        seed: 0x5eed2025
+  - id: quora
+    label: "Quora (IR)"
+    category: "Quora"
+    entity_suffix: "Quora"
+    source_prefix: "quora"
+    raw: "data/raw/quora"
+    converted: "data/converted/quora-minne.json"
+    include_unanswerable: false
+    slices:
+      - id: quora-test-200
+        label: "Quora test (200)"
+        description: "200-case slice from BEIR test qrels"
+        limit: 200
+        corpus_limit: 5000
+        seed: 0x5eed2025
+  - id: trec-covid
+    label: "TREC-COVID (BEIR)"
+    category: "TREC-COVID"
+    entity_suffix: "TREC-COVID"
+    source_prefix: "trec-covid"
+    raw: "data/raw/trec-covid"
+    converted: "data/converted/trec-covid-minne.json"
+    include_unanswerable: false
+    slices:
+      - id: trec-covid-test-200
+        label: "TREC-COVID test (200)"
+        description: "200-case slice from BEIR test qrels"
+        limit: 200
+        corpus_limit: 5000
+        seed: 0x5eed2025
+  - id: scifact
+    label: "SciFact (BEIR)"
+    category: "SciFact"
+    entity_suffix: "SciFact"
+    source_prefix: "scifact"
+    raw: "data/raw/scifact"
+    converted: "data/converted/scifact-minne.json"
+    include_unanswerable: false
+    slices:
+      - id: scifact-test-200
+        label: "SciFact test (200)"
+        description: "200-case slice from BEIR test qrels"
+        limit: 200
+        corpus_limit: 3000
+        seed: 0x5eed2025
+  - id: nq-beir
+    label: "Natural Questions (BEIR)"
+    category: "Natural Questions"
+    entity_suffix: "Natural Questions"
+    source_prefix: "nq-beir"
+    raw: "data/raw/nq"
+    converted: "data/converted/nq-beir-minne.json"
+    include_unanswerable: false
+    slices:
+      - id: nq-beir-test-200
+        label: "NQ (BEIR) test (200)"
+        description: "200-case slice from BEIR test qrels"
+        limit: 200
+        corpus_limit: 5000
+        seed: 0x5eed2025

evaluations/src/args.rs

@@ -0,0 +1,506 @@
+use std::{
+    env,
+    path::{Path, PathBuf},
+};
+
+use anyhow::{anyhow, Context, Result};
+use clap::{Args, Parser, ValueEnum};
+use retrieval_pipeline::RetrievalStrategy;
+
+use crate::datasets::DatasetKind;
+
+fn workspace_root() -> PathBuf {
+    let manifest_dir = PathBuf::from(env!("CARGO_MANIFEST_DIR"));
+    manifest_dir.parent().unwrap_or(&manifest_dir).to_path_buf()
+}
+
+fn default_report_dir() -> PathBuf {
+    workspace_root().join("evaluations/reports")
+}
+
+fn default_cache_dir() -> PathBuf {
+    workspace_root().join("evaluations/cache")
+}
+
+fn default_ingestion_cache_dir() -> PathBuf {
+    workspace_root().join("evaluations/cache/ingested")
+}
+
+pub const DEFAULT_SLICE_SEED: u64 = 0x5eed_2025;
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq, ValueEnum, Default)]
+#[value(rename_all = "lowercase")]
+pub enum EmbeddingBackend {
+    Hashed,
+    #[default]
+    FastEmbed,
+}
+
+impl std::fmt::Display for EmbeddingBackend {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        match self {
+            Self::Hashed => write!(f, "hashed"),
+            Self::FastEmbed => write!(f, "fastembed"),
+        }
+    }
+}
+
+#[derive(Debug, Clone, Args)]
+pub struct RetrievalSettings {
+    /// Override chunk vector candidate cap
+    #[arg(long)]
+    pub chunk_vector_take: Option<usize>,
+
+    /// Override chunk FTS candidate cap
+    #[arg(long)]
+    pub chunk_fts_take: Option<usize>,
+
+    /// Override average characters per token used for budgeting
+    #[arg(long)]
+    pub chunk_avg_chars_per_token: Option<usize>,
+
+    /// Override maximum chunks attached per entity
+    #[arg(long)]
+    pub max_chunks_per_entity: Option<usize>,
+
+    /// Enable the FastEmbed reranking stage
+    #[arg(long = "rerank", action = clap::ArgAction::SetTrue, default_value_t = false)]
+    pub rerank: bool,
+
+    /// Reranking engine pool size / parallelism
+    #[arg(long, default_value_t = 4)]
+    pub rerank_pool_size: usize,
+
+    /// Keep top-N entities after reranking
+    #[arg(long, default_value_t = 10)]
+    pub rerank_keep_top: usize,
+
+    /// Cap the number of chunks returned by retrieval (revised strategy)
+    #[arg(long, default_value_t = 5)]
+    pub chunk_result_cap: usize,
+
+    /// Reciprocal rank fusion k value for revised chunk merging
+    #[arg(long)]
+    pub chunk_rrf_k: Option<f32>,
+
+    /// Weight for vector ranks in revised RRF
+    #[arg(long)]
+    pub chunk_rrf_vector_weight: Option<f32>,
+
+    /// Weight for chunk FTS ranks in revised RRF
+    #[arg(long)]
+    pub chunk_rrf_fts_weight: Option<f32>,
+
+    /// Include vector ranks in revised RRF (default: true)
+    #[arg(long)]
+    pub chunk_rrf_use_vector: Option<bool>,
+
+    /// Include chunk FTS ranks in revised RRF (default: true)
+    #[arg(long)]
+    pub chunk_rrf_use_fts: Option<bool>,
+
+    /// Require verified chunks (disable with --llm-mode)
+    #[arg(skip = true)]
+    pub require_verified_chunks: bool,
+
+    /// Select the retrieval pipeline strategy
+    #[arg(long, default_value_t = RetrievalStrategy::Default)]
+    pub strategy: RetrievalStrategy,
+}
+
+impl Default for RetrievalSettings {
+    fn default() -> Self {
+        Self {
+            chunk_vector_take: None,
+            chunk_fts_take: None,
+            chunk_avg_chars_per_token: None,
+            max_chunks_per_entity: None,
+            rerank: false,
+            rerank_pool_size: 4,
+            rerank_keep_top: 10,
+            chunk_result_cap: 5,
+            chunk_rrf_k: None,
+            chunk_rrf_vector_weight: None,
+            chunk_rrf_fts_weight: None,
+            chunk_rrf_use_vector: None,
+            chunk_rrf_use_fts: None,
+            require_verified_chunks: true,
+            strategy: RetrievalStrategy::Default,
+        }
+    }
+}
+
+#[derive(Debug, Clone, Args)]
+pub struct IngestConfig {
+    /// Directory for ingestion corpora caches
+    #[arg(long, default_value_os_t = default_ingestion_cache_dir())]
+    pub ingestion_cache_dir: PathBuf,
+
+    /// Minimum tokens per chunk for ingestion
+    #[arg(long, default_value_t = 256)]
+    pub ingest_chunk_min_tokens: usize,
+
+    /// Maximum tokens per chunk for ingestion
+    #[arg(long, default_value_t = 512)]
+    pub ingest_chunk_max_tokens: usize,
+
+    /// Overlap between chunks during ingestion (tokens)
+    #[arg(long, default_value_t = 50)]
+    pub ingest_chunk_overlap_tokens: usize,
+
+    /// Run ingestion in chunk-only mode (skip analyzer/graph generation)
+    #[arg(long)]
+    pub ingest_chunks_only: bool,
+
+    /// Number of paragraphs to ingest concurrently
+    #[arg(long, default_value_t = 10)]
+    pub ingestion_batch_size: usize,
+
+    /// Maximum retries for ingestion failures per paragraph
+    #[arg(long, default_value_t = 3)]
+    pub ingestion_max_retries: usize,
+
+    /// Recompute embeddings for cached corpora without re-running ingestion
+    #[arg(long, alias = "refresh-embeddings")]
+    pub refresh_embeddings_only: bool,
+
+    /// Delete cached paragraph shards before rebuilding the ingestion corpus
+    #[arg(long)]
+    pub slice_reset_ingestion: bool,
+}
+
+#[derive(Debug, Clone, Args)]
+pub struct DatabaseArgs {
+    /// SurrealDB server endpoint
+    #[arg(long, default_value = "ws://127.0.0.1:8000", env = "EVAL_DB_ENDPOINT")]
+    pub db_endpoint: String,
+
+    /// SurrealDB root username
+    #[arg(long, default_value = "root_user", env = "EVAL_DB_USERNAME")]
+    pub db_username: String,
+
+    /// SurrealDB root password
+    #[arg(long, default_value = "root_password", env = "EVAL_DB_PASSWORD")]
+    pub db_password: String,
+
+    /// Override the namespace used on the SurrealDB server
+    #[arg(long, env = "EVAL_DB_NAMESPACE")]
+    pub db_namespace: Option<String>,
+
+    /// Override the database used on the SurrealDB server
+    #[arg(long, env = "EVAL_DB_DATABASE")]
+    pub db_database: Option<String>,
+
+    /// Path to inspect DB state
+    #[arg(long)]
+    pub inspect_db_state: Option<PathBuf>,
+}
+
+#[derive(Parser, Debug, Clone)]
+#[command(author, version, about, long_about = None)]
+pub struct Config {
+    /// Convert the selected dataset and exit
+    #[arg(long)]
+    pub convert_only: bool,
+
+    /// Regenerate the converted dataset even if it already exists
+    #[arg(long, alias = "refresh")]
+    pub force_convert: bool,
+
+    /// Dataset to evaluate
+    #[arg(long, default_value_t = DatasetKind::default())]
+    pub dataset: DatasetKind,
+
+    /// Enable LLM-assisted evaluation features (includes unanswerable cases)
+    #[arg(long)]
+    pub llm_mode: bool,
+
+    /// Cap the slice corpus size (positives + negatives)
+    #[arg(long)]
+    pub corpus_limit: Option<usize>,
+
+    /// Path to the raw dataset (defaults per dataset)
+    #[arg(long)]
+    pub raw: Option<PathBuf>,
+
+    /// Path to write/read the converted dataset (defaults per dataset)
+    #[arg(long)]
+    pub converted: Option<PathBuf>,
+
+    /// Directory to write evaluation reports
+    #[arg(long, default_value_os_t = default_report_dir())]
+    pub report_dir: PathBuf,
+
+    /// Precision@k cutoff
+    #[arg(long, default_value_t = 5)]
+    pub k: usize,
+
+    /// Limit the number of questions evaluated (0 = all)
+    #[arg(long = "limit", default_value_t = 200)]
+    pub limit_arg: usize,
+
+    /// Number of mismatches to surface in the Markdown summary
+    #[arg(long, default_value_t = 5)]
+    pub sample: usize,
+
+    /// Disable context cropping when converting datasets (ingest entire documents)
+    #[arg(long)]
+    pub full_context: bool,
+
+    #[command(flatten)]
+    pub retrieval: RetrievalSettings,
+
+    /// Concurrency level
+    #[arg(long, default_value_t = 1)]
+    pub concurrency: usize,
+
+    /// Embedding backend
+    #[arg(long, default_value_t = EmbeddingBackend::FastEmbed)]
+    pub embedding_backend: EmbeddingBackend,
+
+    /// FastEmbed model code
+    #[arg(long)]
+    pub embedding_model: Option<String>,
+
+    /// Directory for embedding caches
+    #[arg(long, default_value_os_t = default_cache_dir())]
+    pub cache_dir: PathBuf,
+
+    #[command(flatten)]
+    pub ingest: IngestConfig,
+
+    /// Include entity descriptions and categories in JSON reports
+    #[arg(long)]
+    pub detailed_report: bool,
+
+    /// Use a cached dataset slice by id or path
+    #[arg(long)]
+    pub slice: Option<String>,
+
+    /// Ignore cached corpus state and rebuild the slice's SurrealDB corpus
+    #[arg(long)]
+    pub reseed_slice: bool,
+
+    /// Slice seed
+    #[arg(skip = DEFAULT_SLICE_SEED)]
+    pub slice_seed: u64,
+
+    /// Grow the slice ledger to contain at least this many answerable cases, then exit
+    #[arg(long)]
+    pub slice_grow: Option<usize>,
+
+    /// Evaluate questions starting at this offset within the slice
+    #[arg(long, default_value_t = 0)]
+    pub slice_offset: usize,
+
+    /// Target negative-to-positive paragraph ratio for slice growth
+    #[arg(long, default_value_t = crate::slice::DEFAULT_NEGATIVE_MULTIPLIER)]
+    pub negative_multiplier: f32,
+
+    /// Annotate the run; label is stored in JSON/Markdown reports
+    #[arg(long)]
+    pub label: Option<String>,
+
+    /// Write per-query chunk diagnostics JSONL to the provided path
+    #[arg(long, alias = "chunk-diagnostics")]
+    pub chunk_diagnostics_path: Option<PathBuf>,
+
+    /// Inspect an ingestion cache question and exit
+    #[arg(long)]
+    pub inspect_question: Option<String>,
+
+    /// Path to an ingestion cache manifest JSON for inspection mode
+    #[arg(long)]
+    pub inspect_manifest: Option<PathBuf>,
+
+    /// Override the SurrealDB system settings query model
+    #[arg(long)]
+    pub query_model: Option<String>,
+
+    /// Write structured performance telemetry JSON to the provided path
+    #[arg(long)]
+    pub perf_log_json: Option<PathBuf>,
+
+    /// Directory that receives timestamped perf JSON copies
+    #[arg(long)]
+    pub perf_log_dir: Option<PathBuf>,
+
+    /// Print per-stage performance timings to stdout after the run
+    #[arg(long, alias = "perf-log")]
+    pub perf_log_console: bool,
+
+    #[command(flatten)]
+    pub database: DatabaseArgs,
+
+    // Computed fields (not arguments)
+    #[arg(skip)]
+    pub raw_dataset_path: PathBuf,
+    #[arg(skip)]
+    pub converted_dataset_path: PathBuf,
+    #[arg(skip)]
+    pub limit: Option<usize>,
+    #[arg(skip)]
+    pub summary_sample: usize,
+}
+
+impl Config {
+    pub fn context_token_limit(&self) -> Option<usize> {
+        None
+    }
+
+    pub fn finalize(&mut self) -> Result<()> {
+        // Handle dataset paths
+        if let Some(raw) = &self.raw {
+            self.raw_dataset_path = raw.clone();
+        } else {
+            self.raw_dataset_path = self.dataset.default_raw_path();
+        }
+
+        if let Some(converted) = &self.converted {
+            self.converted_dataset_path = converted.clone();
+        } else {
+            self.converted_dataset_path = self.dataset.default_converted_path();
+        }
+
+        // Handle limit
+        if self.limit_arg == 0 {
+            self.limit = None;
+        } else {
+            self.limit = Some(self.limit_arg);
+        }
+
+        // Handle sample
+        self.summary_sample = self.sample.max(1);
+
+        // Handle retrieval settings
+        self.retrieval.require_verified_chunks = !self.llm_mode;
+
+        if self.dataset == DatasetKind::Beir {
+            self.negative_multiplier = 9.0;
+        }
+
+        // Validations
+        if self.ingest.ingest_chunk_min_tokens == 0
+            || self.ingest.ingest_chunk_min_tokens >= self.ingest.ingest_chunk_max_tokens
+        {
+            return Err(anyhow!(
+                "--ingest-chunk-min-tokens must be greater than zero and less than --ingest-chunk-max-tokens (got {} >= {})",
+                self.ingest.ingest_chunk_min_tokens,
+                self.ingest.ingest_chunk_max_tokens
+            ));
+        }
+
+        if self.ingest.ingest_chunk_overlap_tokens >= self.ingest.ingest_chunk_min_tokens {
+            return Err(anyhow!(
+                "--ingest-chunk-overlap-tokens ({}) must be less than --ingest-chunk-min-tokens ({})",
+                self.ingest.ingest_chunk_overlap_tokens,
+                self.ingest.ingest_chunk_min_tokens
+            ));
+        }
+
+        if self.retrieval.rerank && self.retrieval.rerank_pool_size == 0 {
+            return Err(anyhow!(
+                "--rerank-pool must be greater than zero when reranking is enabled"
+            ));
+        }
+
+        if let Some(k) = self.retrieval.chunk_rrf_k {
+            if k <= 0.0 || !k.is_finite() {
+                return Err(anyhow!(
+                    "--chunk-rrf-k must be a positive, finite number (got {k})"
+                ));
+            }
+        }
+        if let Some(weight) = self.retrieval.chunk_rrf_vector_weight {
+            if weight < 0.0 || !weight.is_finite() {
+                return Err(anyhow!(
+                    "--chunk-rrf-vector-weight must be a non-negative, finite number (got {weight})"
+                ));
+            }
+        }
+        if let Some(weight) = self.retrieval.chunk_rrf_fts_weight {
+            if weight < 0.0 || !weight.is_finite() {
+                return Err(anyhow!(
+                    "--chunk-rrf-fts-weight must be a non-negative, finite number (got {weight})"
+                ));
+            }
+        }
+
+        if self.concurrency == 0 {
+            return Err(anyhow!("--concurrency must be greater than zero"));
+        }
+
+        if self.embedding_backend == EmbeddingBackend::Hashed && self.embedding_model.is_some() {
+            return Err(anyhow!(
+                "--embedding-model cannot be used with the 'hashed' embedding backend"
+            ));
+        }
+
+        if let Some(query_model) = &self.query_model {
+            if query_model.trim().is_empty() {
+                return Err(anyhow!("--query-model requires a non-empty model name"));
+            }
+        }
+
+        if let Some(grow) = self.slice_grow {
+            if grow == 0 {
+                return Err(anyhow!("--slice-grow must be greater than zero"));
+            }
+        }
+
+        if self.negative_multiplier <= 0.0 || !self.negative_multiplier.is_finite() {
+            return Err(anyhow!(
+                "--negative-multiplier must be a positive finite number"
+            ));
+        }
+
+        // Handle corpus limit logic
+        if let Some(limit) = self.limit {
+            if let Some(corpus_limit) = self.corpus_limit {
+                if corpus_limit < limit {
+                    self.corpus_limit = Some(limit);
+                }
+            } else {
+                let default_multiplier = 10usize;
+                let mut computed = limit.saturating_mul(default_multiplier);
+                if computed < limit {
+                    computed = limit;
+                }
+                let max_cap = 1_000usize;
+                if computed > max_cap {
+                    computed = max_cap;
+                }
+                self.corpus_limit = Some(computed);
+            }
+        }
+
+        // Handle perf log dir env var fallback
+        if self.perf_log_dir.is_none() {
+            if let Ok(dir) = env::var("EVAL_PERF_LOG_DIR") {
+                if !dir.trim().is_empty() {
+                    self.perf_log_dir = Some(PathBuf::from(dir));
+                }
+            }
+        }
+
+        Ok(())
+    }
+}
+
+pub struct ParsedArgs {
+    pub config: Config,
+}
+
+pub fn parse() -> Result<ParsedArgs> {
+    let mut config = Config::parse();
+    config.finalize()?;
+    Ok(ParsedArgs { config })
+}
+
+pub fn ensure_parent(path: &Path) -> Result<()> {
+    if let Some(parent) = path.parent() {
+        std::fs::create_dir_all(parent)
+            .with_context(|| format!("creating parent directory for {}", path.display()))?;
+    }
+    Ok(())
+}

evaluations/src/cache.rs

@@ -0,0 +1,88 @@
+use std::{
+    collections::HashMap,
+    path::{Path, PathBuf},
+    sync::{
+        atomic::{AtomicBool, Ordering},
+        Arc,
+    },
+};
+
+use anyhow::{Context, Result};
+use serde::{Deserialize, Serialize};
+use tokio::sync::Mutex;
+
+#[derive(Debug, Default, Serialize, Deserialize)]
+struct EmbeddingCacheData {
+    entities: HashMap<String, Vec<f32>>,
+    chunks: HashMap<String, Vec<f32>>,
+}
+
+#[derive(Clone)]
+pub struct EmbeddingCache {
+    path: Arc<PathBuf>,
+    data: Arc<Mutex<EmbeddingCacheData>>,
+    dirty: Arc<AtomicBool>,
+}
+
+#[allow(dead_code)]
+impl EmbeddingCache {
+    pub async fn load(path: impl AsRef<Path>) -> Result<Self> {
+        let path = path.as_ref().to_path_buf();
+        let data = if path.exists() {
+            let raw = tokio::fs::read(&path)
+                .await
+                .with_context(|| format!("reading embedding cache {}", path.display()))?;
+            serde_json::from_slice(&raw)
+                .with_context(|| format!("parsing embedding cache {}", path.display()))?
+        } else {
+            EmbeddingCacheData::default()
+        };
+
+        Ok(Self {
+            path: Arc::new(path),
+            data: Arc::new(Mutex::new(data)),
+            dirty: Arc::new(AtomicBool::new(false)),
+        })
+    }
+
+    pub async fn get_entity(&self, id: &str) -> Option<Vec<f32>> {
+        let guard = self.data.lock().await;
+        guard.entities.get(id).cloned()
+    }
+
+    pub async fn insert_entity(&self, id: String, embedding: Vec<f32>) {
+        let mut guard = self.data.lock().await;
+        guard.entities.insert(id, embedding);
+        self.dirty.store(true, Ordering::Relaxed);
+    }
+
+    pub async fn get_chunk(&self, id: &str) -> Option<Vec<f32>> {
+        let guard = self.data.lock().await;
+        guard.chunks.get(id).cloned()
+    }
+
+    pub async fn insert_chunk(&self, id: String, embedding: Vec<f32>) {
+        let mut guard = self.data.lock().await;
+        guard.chunks.insert(id, embedding);
+        self.dirty.store(true, Ordering::Relaxed);
+    }
+
+    pub async fn persist(&self) -> Result<()> {
+        if !self.dirty.load(Ordering::Relaxed) {
+            return Ok(());
+        }
+
+        let guard = self.data.lock().await;
+        let body = serde_json::to_vec_pretty(&*guard).context("serialising embedding cache")?;
+        if let Some(parent) = self.path.parent() {
+            tokio::fs::create_dir_all(parent)
+                .await
+                .with_context(|| format!("creating cache directory {}", parent.display()))?;
+        }
+        tokio::fs::write(&*self.path, body)
+            .await
+            .with_context(|| format!("writing embedding cache {}", self.path.display()))?;
+        self.dirty.store(false, Ordering::Relaxed);
+        Ok(())
+    }
+}

evaluations/src/cases.rs

@@ -0,0 +1,187 @@
+//! Case generation from corpus manifests.
+
+use std::collections::HashMap;
+
+use crate::corpus;
+
+/// A test case for retrieval evaluation derived from a manifest question.
+pub(crate) struct SeededCase {
+    pub question_id: String,
+    pub question: String,
+    pub expected_source: String,
+    pub answers: Vec<String>,
+    pub paragraph_id: String,
+    pub paragraph_title: String,
+    pub expected_chunk_ids: Vec<String>,
+    pub is_impossible: bool,
+    pub has_verified_chunks: bool,
+}
+
+/// Convert a corpus manifest into seeded evaluation cases.
+pub(crate) fn cases_from_manifest(manifest: &corpus::CorpusManifest) -> Vec<SeededCase> {
+    let mut title_map = HashMap::new();
+    for paragraph in &manifest.paragraphs {
+        title_map.insert(paragraph.paragraph_id.as_str(), paragraph.title.clone());
+    }
+
+    let include_impossible = manifest.metadata.include_unanswerable;
+    let require_verified_chunks = manifest.metadata.require_verified_chunks;
+
+    manifest
+        .questions
+        .iter()
+        .filter(|question| {
+            should_include_question(question, include_impossible, require_verified_chunks)
+        })
+        .map(|question| {
+            let title = title_map
+                .get(question.paragraph_id.as_str())
+                .cloned()
+                .unwrap_or_else(|| "Untitled".to_string());
+            SeededCase {
+                question_id: question.question_id.clone(),
+                question: question.question_text.clone(),
+                expected_source: question.text_content_id.clone(),
+                answers: question.answers.clone(),
+                paragraph_id: question.paragraph_id.clone(),
+                paragraph_title: title,
+                expected_chunk_ids: question.matching_chunk_ids.clone(),
+                is_impossible: question.is_impossible,
+                has_verified_chunks: !question.matching_chunk_ids.is_empty(),
+            }
+        })
+        .collect()
+}
+
+fn should_include_question(
+    question: &corpus::CorpusQuestion,
+    include_impossible: bool,
+    require_verified_chunks: bool,
+) -> bool {
+    if !include_impossible && question.is_impossible {
+        return false;
+    }
+    if require_verified_chunks && question.matching_chunk_ids.is_empty() {
+        return false;
+    }
+    true
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::corpus::store::{CorpusParagraph, EmbeddedKnowledgeEntity, EmbeddedTextChunk};
+    use crate::corpus::{CorpusManifest, CorpusMetadata, CorpusQuestion, MANIFEST_VERSION};
+    use chrono::Utc;
+    use common::storage::types::text_content::TextContent;
+
+    fn sample_manifest() -> CorpusManifest {
+        let paragraphs = vec![
+            CorpusParagraph {
+                paragraph_id: "p1".to_string(),
+                title: "Alpha".to_string(),
+                text_content: TextContent::new(
+                    "alpha context".to_string(),
+                    None,
+                    "test".to_string(),
+                    None,
+                    None,
+                    "user".to_string(),
+                ),
+                entities: Vec::<EmbeddedKnowledgeEntity>::new(),
+                relationships: Vec::new(),
+                chunks: Vec::<EmbeddedTextChunk>::new(),
+            },
+            CorpusParagraph {
+                paragraph_id: "p2".to_string(),
+                title: "Beta".to_string(),
+                text_content: TextContent::new(
+                    "beta context".to_string(),
+                    None,
+                    "test".to_string(),
+                    None,
+                    None,
+                    "user".to_string(),
+                ),
+                entities: Vec::<EmbeddedKnowledgeEntity>::new(),
+                relationships: Vec::new(),
+                chunks: Vec::<EmbeddedTextChunk>::new(),
+            },
+        ];
+        let questions = vec![
+            CorpusQuestion {
+                question_id: "q1".to_string(),
+                paragraph_id: "p1".to_string(),
+                text_content_id: "tc-alpha".to_string(),
+                question_text: "What is Alpha?".to_string(),
+                answers: vec!["Alpha".to_string()],
+                is_impossible: false,
+                matching_chunk_ids: vec!["chunk-alpha".to_string()],
+            },
+            CorpusQuestion {
+                question_id: "q2".to_string(),
+                paragraph_id: "p1".to_string(),
+                text_content_id: "tc-alpha".to_string(),
+                question_text: "Unanswerable?".to_string(),
+                answers: Vec::new(),
+                is_impossible: true,
+                matching_chunk_ids: Vec::new(),
+            },
+            CorpusQuestion {
+                question_id: "q3".to_string(),
+                paragraph_id: "p2".to_string(),
+                text_content_id: "tc-beta".to_string(),
+                question_text: "Where is Beta?".to_string(),
+                answers: vec!["Beta".to_string()],
+                is_impossible: false,
+                matching_chunk_ids: Vec::new(),
+            },
+        ];
+        CorpusManifest {
+            version: MANIFEST_VERSION,
+            metadata: CorpusMetadata {
+                dataset_id: "ds".to_string(),
+                dataset_label: "Dataset".to_string(),
+                slice_id: "slice".to_string(),
+                include_unanswerable: true,
+                require_verified_chunks: true,
+                ingestion_fingerprint: "fp".to_string(),
+                embedding_backend: "test".to_string(),
+                embedding_model: None,
+                embedding_dimension: 3,
+                converted_checksum: "chk".to_string(),
+                generated_at: Utc::now(),
+                paragraph_count: paragraphs.len(),
+                question_count: questions.len(),
+                chunk_min_tokens: 1,
+                chunk_max_tokens: 10,
+                chunk_only: false,
+            },
+            paragraphs,
+            questions,
+        }
+    }
+
+    #[test]
+    fn cases_respect_mode_filters() {
+        let mut manifest = sample_manifest();
+        manifest.metadata.include_unanswerable = false;
+        manifest.metadata.require_verified_chunks = true;
+
+        let strict_cases = cases_from_manifest(&manifest);
+        assert_eq!(strict_cases.len(), 1);
+        assert_eq!(strict_cases[0].question_id, "q1");
+        assert_eq!(strict_cases[0].paragraph_title, "Alpha");
+
+        let mut llm_manifest = manifest.clone();
+        llm_manifest.metadata.include_unanswerable = true;
+        llm_manifest.metadata.require_verified_chunks = false;
+
+        let llm_cases = cases_from_manifest(&llm_manifest);
+        let ids: Vec<_> = llm_cases
+            .iter()
+            .map(|case| case.question_id.as_str())
+            .collect();
+        assert_eq!(ids, vec!["q1", "q2", "q3"]);
+    }
+}

evaluations/src/corpus/config.rs

@@ -0,0 +1,42 @@
+use std::path::PathBuf;
+
+use crate::args::Config;
+
+#[derive(Debug, Clone)]
+pub struct CorpusCacheConfig {
+    pub ingestion_cache_dir: PathBuf,
+    pub force_refresh: bool,
+    pub refresh_embeddings_only: bool,
+    pub ingestion_batch_size: usize,
+    pub ingestion_max_retries: usize,
+}
+
+impl CorpusCacheConfig {
+    pub fn new(
+        ingestion_cache_dir: impl Into<PathBuf>,
+        force_refresh: bool,
+        refresh_embeddings_only: bool,
+        ingestion_batch_size: usize,
+        ingestion_max_retries: usize,
+    ) -> Self {
+        Self {
+            ingestion_cache_dir: ingestion_cache_dir.into(),
+            force_refresh,
+            refresh_embeddings_only,
+            ingestion_batch_size,
+            ingestion_max_retries,
+        }
+    }
+}
+
+impl From<&Config> for CorpusCacheConfig {
+    fn from(config: &Config) -> Self {
+        CorpusCacheConfig::new(
+            config.ingest.ingestion_cache_dir.clone(),
+            config.force_convert || config.ingest.slice_reset_ingestion,
+            config.ingest.refresh_embeddings_only,
+            config.ingest.ingestion_batch_size,
+            config.ingest.ingestion_max_retries,
+        )
+    }
+}

evaluations/src/corpus/mod.rs

@@ -0,0 +1,26 @@
+mod config;
+mod orchestrator;
+pub(crate) mod store;
+
+pub use config::CorpusCacheConfig;
+pub use orchestrator::{
+    cached_corpus_dir, compute_ingestion_fingerprint, corpus_handle_from_manifest, ensure_corpus,
+    load_cached_manifest,
+};
+pub use store::{
+    seed_manifest_into_db, window_manifest, CorpusHandle, CorpusManifest, CorpusMetadata,
+    CorpusQuestion, EmbeddedKnowledgeEntity, EmbeddedTextChunk, ParagraphShard,
+    ParagraphShardStore, MANIFEST_VERSION,
+};
+
+pub fn make_ingestion_config(config: &crate::args::Config) -> ingestion_pipeline::IngestionConfig {
+    ingestion_pipeline::IngestionConfig {
+        tuning: ingestion_pipeline::IngestionTuning {
+            chunk_min_tokens: config.ingest.ingest_chunk_min_tokens,
+            chunk_max_tokens: config.ingest.ingest_chunk_max_tokens,
+            chunk_overlap_tokens: config.ingest.ingest_chunk_overlap_tokens,
+            ..Default::default()
+        },
+        chunk_only: config.ingest.ingest_chunks_only,
+    }
+}

evaluations/src/corpus/orchestrator.rs

@@ -0,0 +1,806 @@
+use std::{
+    collections::{HashMap, HashSet},
+    fs,
+    io::Read,
+    path::{Path, PathBuf},
+    sync::Arc,
+};
+
+use anyhow::{anyhow, Context, Result};
+use async_openai::Client;
+use chrono::Utc;
+#[cfg(not(test))]
+use common::utils::config::get_config;
+use common::{
+    storage::{
+        db::SurrealDbClient,
+        store::{DynStore, StorageManager},
+        types::{ingestion_payload::IngestionPayload, ingestion_task::IngestionTask, StoredObject},
+    },
+    utils::config::{AppConfig, StorageKind},
+};
+use futures::future::try_join_all;
+use ingestion_pipeline::{IngestionConfig, IngestionPipeline};
+use object_store::memory::InMemory;
+use sha2::{Digest, Sha256};
+use tracing::{info, warn};
+use uuid::Uuid;
+
+use crate::{
+    datasets::{ConvertedDataset, ConvertedParagraph, ConvertedQuestion},
+    slice::{self, ResolvedSlice, SliceParagraphKind},
+};
+
+use crate::corpus::{
+    CorpusCacheConfig, CorpusHandle, CorpusManifest, CorpusMetadata, CorpusQuestion,
+    EmbeddedKnowledgeEntity, EmbeddedTextChunk, ParagraphShard, ParagraphShardStore,
+    MANIFEST_VERSION,
+};
+
+const INGESTION_SPEC_VERSION: u32 = 2;
+
+type OpenAIClient = Client<async_openai::config::OpenAIConfig>;
+
+#[derive(Clone)]
+struct ParagraphShardRecord {
+    shard: ParagraphShard,
+    dirty: bool,
+    needs_reembed: bool,
+}
+
+#[derive(Clone)]
+struct IngestRequest<'a> {
+    slot: usize,
+    paragraph: &'a ConvertedParagraph,
+    shard_path: String,
+    question_refs: Vec<&'a ConvertedQuestion>,
+}
+
+impl<'a> IngestRequest<'a> {
+    fn from_entry(
+        slot: usize,
+        paragraph: &'a ConvertedParagraph,
+        entry: &'a slice::SliceParagraphEntry,
+    ) -> Result<Self> {
+        let shard_path = entry
+            .shard_path
+            .clone()
+            .unwrap_or_else(|| slice::default_shard_path(&entry.id));
+        let question_refs = match &entry.kind {
+            SliceParagraphKind::Positive { question_ids } => question_ids
+                .iter()
+                .map(|id| {
+                    paragraph
+                        .questions
+                        .iter()
+                        .find(|question| question.id == *id)
+                        .ok_or_else(|| {
+                            anyhow!(
+                                "paragraph '{}' missing question '{}' referenced by slice",
+                                paragraph.id,
+                                id
+                            )
+                        })
+                })
+                .collect::<Result<Vec<_>>>()?,
+            SliceParagraphKind::Negative => Vec::new(),
+        };
+        Ok(Self {
+            slot,
+            paragraph,
+            shard_path,
+            question_refs,
+        })
+    }
+}
+
+struct ParagraphPlan<'a> {
+    slot: usize,
+    entry: &'a slice::SliceParagraphEntry,
+    paragraph: &'a ConvertedParagraph,
+}
+
+#[derive(Default)]
+struct IngestionStats {
+    positive_reused: usize,
+    positive_ingested: usize,
+    negative_reused: usize,
+    negative_ingested: usize,
+}
+
+#[allow(clippy::too_many_arguments)]
+pub async fn ensure_corpus(
+    dataset: &ConvertedDataset,
+    slice: &ResolvedSlice<'_>,
+    window: &slice::SliceWindow<'_>,
+    cache: &CorpusCacheConfig,
+    embedding: Arc<common::utils::embedding::EmbeddingProvider>,
+    openai: Arc<OpenAIClient>,
+    user_id: &str,
+    converted_path: &Path,
+    ingestion_config: IngestionConfig,
+) -> Result<CorpusHandle> {
+    let checksum = compute_file_checksum(converted_path)
+        .with_context(|| format!("computing checksum for {}", converted_path.display()))?;
+    let ingestion_fingerprint =
+        build_ingestion_fingerprint(dataset, slice, &checksum, &ingestion_config);
+
+    let base_dir = cached_corpus_dir(
+        cache,
+        dataset.metadata.id.as_str(),
+        slice.manifest.slice_id.as_str(),
+    );
+    if cache.force_refresh && !cache.refresh_embeddings_only {
+        let _ = fs::remove_dir_all(&base_dir);
+    }
+    let store = ParagraphShardStore::new(base_dir.clone());
+    store.ensure_base_dir()?;
+
+    let positive_set: HashSet<&str> = window.positive_ids().collect();
+    let require_verified_chunks = slice.manifest.require_verified_chunks;
+    let embedding_backend_label = embedding.backend_label().to_string();
+    let embedding_model_code = embedding.model_code();
+    let embedding_dimension = embedding.dimension();
+    if positive_set.is_empty() {
+        return Err(anyhow!(
+            "window selection contains zero positive paragraphs for slice '{}'",
+            slice.manifest.slice_id
+        ));
+    }
+
+    let desired_negatives =
+        ((positive_set.len() as f32) * slice.manifest.negative_multiplier).ceil() as usize;
+    let mut plan = Vec::new();
+    let mut negatives_added = 0usize;
+    for (idx, entry) in slice.manifest.paragraphs.iter().enumerate() {
+        let include = match &entry.kind {
+            SliceParagraphKind::Positive { .. } => positive_set.contains(entry.id.as_str()),
+            SliceParagraphKind::Negative => {
+                negatives_added < desired_negatives && {
+                    negatives_added += 1;
+                    true
+                }
+            }
+        };
+        if include {
+            let paragraph = slice
+                .paragraphs
+                .get(idx)
+                .copied()
+                .ok_or_else(|| anyhow!("slice missing paragraph index {}", idx))?;
+            plan.push(ParagraphPlan {
+                slot: plan.len(),
+                entry,
+                paragraph,
+            });
+        }
+    }
+
+    if plan.is_empty() {
+        return Err(anyhow!(
+            "no paragraphs selected for ingestion (slice '{}')",
+            slice.manifest.slice_id
+        ));
+    }
+
+    let mut records: Vec<Option<ParagraphShardRecord>> = vec![None; plan.len()];
+    let mut ingest_requests = Vec::new();
+    let mut stats = IngestionStats::default();
+
+    for plan_entry in &plan {
+        let shard_path = plan_entry
+            .entry
+            .shard_path
+            .clone()
+            .unwrap_or_else(|| slice::default_shard_path(&plan_entry.entry.id));
+        let shard = if cache.force_refresh {
+            None
+        } else {
+            store.load(&shard_path, &ingestion_fingerprint)?
+        };
+        if let Some(shard) = shard {
+            let model_matches = shard.embedding_model.as_deref() == embedding_model_code.as_deref();
+            let needs_reembed = shard.embedding_backend != embedding_backend_label
+                || shard.embedding_dimension != embedding_dimension
+                || !model_matches;
+            match plan_entry.entry.kind {
+                SliceParagraphKind::Positive { .. } => stats.positive_reused += 1,
+                SliceParagraphKind::Negative => stats.negative_reused += 1,
+            }
+            records[plan_entry.slot] = Some(ParagraphShardRecord {
+                shard,
+                dirty: false,
+                needs_reembed,
+            });
+        } else {
+            match plan_entry.entry.kind {
+                SliceParagraphKind::Positive { .. } => stats.positive_ingested += 1,
+                SliceParagraphKind::Negative => stats.negative_ingested += 1,
+            }
+            let request =
+                IngestRequest::from_entry(plan_entry.slot, plan_entry.paragraph, plan_entry.entry)?;
+            ingest_requests.push(request);
+        }
+    }
+
+    if cache.refresh_embeddings_only && !ingest_requests.is_empty() {
+        return Err(anyhow!(
+            "--refresh-embeddings requested but {} shard(s) missing for dataset '{}' slice '{}'",
+            ingest_requests.len(),
+            dataset.metadata.id,
+            slice.manifest.slice_id
+        ));
+    }
+
+    if !ingest_requests.is_empty() {
+        let new_shards = ingest_paragraph_batch(
+            dataset,
+            &ingest_requests,
+            embedding.clone(),
+            openai.clone(),
+            user_id,
+            &ingestion_fingerprint,
+            &embedding_backend_label,
+            embedding_model_code.clone(),
+            embedding_dimension,
+            cache.ingestion_batch_size,
+            cache.ingestion_max_retries,
+            ingestion_config.clone(),
+        )
+        .await
+        .context("ingesting missing slice paragraphs")?;
+        for (request, shard) in ingest_requests.into_iter().zip(new_shards.into_iter()) {
+            store.persist(&shard)?;
+            records[request.slot] = Some(ParagraphShardRecord {
+                shard,
+                dirty: false,
+                needs_reembed: false,
+            });
+        }
+    }
+
+    for record in &mut records {
+        let shard_record = record
+            .as_mut()
+            .context("shard record missing after ingestion run")?;
+        if cache.refresh_embeddings_only || shard_record.needs_reembed {
+            // Embeddings are now generated by the pipeline using FastEmbed - no need to re-embed
+            shard_record.shard.ingestion_fingerprint = ingestion_fingerprint.clone();
+            shard_record.shard.ingested_at = Utc::now();
+            shard_record.shard.embedding_backend = embedding_backend_label.clone();
+            shard_record.shard.embedding_model = embedding_model_code.clone();
+            shard_record.shard.embedding_dimension = embedding_dimension;
+            shard_record.dirty = true;
+            shard_record.needs_reembed = false;
+        }
+    }
+
+    let mut record_index = HashMap::new();
+    for (idx, plan_entry) in plan.iter().enumerate() {
+        record_index.insert(plan_entry.entry.id.as_str(), idx);
+    }
+
+    let mut corpus_paragraphs = Vec::with_capacity(plan.len());
+    for record in &records {
+        let shard = &record.as_ref().expect("record missing").shard;
+        corpus_paragraphs.push(shard.to_corpus_paragraph());
+    }
+
+    let mut corpus_questions = Vec::with_capacity(window.cases.len());
+    for case in &window.cases {
+        let slot = record_index
+            .get(case.paragraph.id.as_str())
+            .copied()
+            .ok_or_else(|| {
+                anyhow!(
+                    "slice case references paragraph '{}' that is not part of the window",
+                    case.paragraph.id
+                )
+            })?;
+        let record_slot = records
+            .get_mut(slot)
+            .context("shard record slot missing for question binding")?;
+        let record = record_slot
+            .as_mut()
+            .context("shard record missing for question binding")?;
+        let (chunk_ids, updated) = match record.shard.ensure_question_binding(case.question) {
+            Ok(result) => result,
+            Err(err) => {
+                if require_verified_chunks {
+                    return Err(err).context(format!(
+                        "locating answer text for question '{}' in paragraph '{}'",
+                        case.question.id, case.paragraph.id
+                    ));
+                }
+                warn!(
+                    question_id = %case.question.id,
+                    paragraph_id = %case.paragraph.id,
+                    error = %err,
+                    "Failed to locate answer text in ingested content; recording empty chunk bindings"
+                );
+                record
+                    .shard
+                    .question_bindings
+                    .insert(case.question.id.clone(), Vec::new());
+                record.dirty = true;
+                (Vec::new(), true)
+            }
+        };
+        if updated {
+            record.dirty = true;
+        }
+        corpus_questions.push(CorpusQuestion {
+            question_id: case.question.id.clone(),
+            paragraph_id: case.paragraph.id.clone(),
+            text_content_id: record.shard.text_content.get_id().to_string(),
+            question_text: case.question.question.clone(),
+            answers: case.question.answers.clone(),
+            is_impossible: case.question.is_impossible,
+            matching_chunk_ids: chunk_ids,
+        });
+    }
+
+    for entry in records.iter_mut().flatten() {
+        if entry.dirty {
+            store.persist(&entry.shard)?;
+        }
+    }
+
+    let manifest = CorpusManifest {
+        version: MANIFEST_VERSION,
+        metadata: CorpusMetadata {
+            dataset_id: dataset.metadata.id.clone(),
+            dataset_label: dataset.metadata.label.clone(),
+            slice_id: slice.manifest.slice_id.clone(),
+            include_unanswerable: slice.manifest.includes_unanswerable,
+            require_verified_chunks: slice.manifest.require_verified_chunks,
+            ingestion_fingerprint: ingestion_fingerprint.clone(),
+            embedding_backend: embedding.backend_label().to_string(),
+            embedding_model: embedding.model_code(),
+            embedding_dimension: embedding.dimension(),
+            converted_checksum: checksum,
+            generated_at: Utc::now(),
+            paragraph_count: corpus_paragraphs.len(),
+            question_count: corpus_questions.len(),
+            chunk_min_tokens: ingestion_config.tuning.chunk_min_tokens,
+            chunk_max_tokens: ingestion_config.tuning.chunk_max_tokens,
+            chunk_only: ingestion_config.chunk_only,
+        },
+        paragraphs: corpus_paragraphs,
+        questions: corpus_questions,
+    };
+
+    let ingested_count = stats.positive_ingested + stats.negative_ingested;
+    let reused_ingestion = ingested_count == 0 && !cache.force_refresh;
+    let reused_embeddings = reused_ingestion && !cache.refresh_embeddings_only;
+
+    info!(
+        dataset = %dataset.metadata.id,
+        slice = %slice.manifest.slice_id,
+        fingerprint = %ingestion_fingerprint,
+        reused_ingestion,
+        reused_embeddings,
+        positive_reused = stats.positive_reused,
+        positive_ingested = stats.positive_ingested,
+        negative_reused = stats.negative_reused,
+        negative_ingested = stats.negative_ingested,
+        shard_dir = %base_dir.display(),
+        "Corpus cache outcome"
+    );
+
+    let handle = CorpusHandle {
+        manifest,
+        path: base_dir,
+        reused_ingestion,
+        reused_embeddings,
+        positive_reused: stats.positive_reused,
+        positive_ingested: stats.positive_ingested,
+        negative_reused: stats.negative_reused,
+        negative_ingested: stats.negative_ingested,
+    };
+
+    persist_manifest(&handle).context("persisting corpus manifest")?;
+
+    Ok(handle)
+}
+
+#[allow(clippy::too_many_arguments)]
+async fn ingest_paragraph_batch(
+    dataset: &ConvertedDataset,
+    targets: &[IngestRequest<'_>],
+    embedding: Arc<common::utils::embedding::EmbeddingProvider>,
+    openai: Arc<OpenAIClient>,
+    user_id: &str,
+    ingestion_fingerprint: &str,
+    embedding_backend: &str,
+    embedding_model: Option<String>,
+    embedding_dimension: usize,
+    batch_size: usize,
+    max_retries: usize,
+    ingestion_config: IngestionConfig,
+) -> Result<Vec<ParagraphShard>> {
+    if targets.is_empty() {
+        return Ok(Vec::new());
+    }
+    let namespace = format!("ingest_eval_{}", Uuid::new_v4());
+    let db = create_ingest_db(&namespace).await?;
+    db.apply_migrations()
+        .await
+        .context("applying migrations for ingestion")?;
+
+    let app_config = AppConfig {
+        storage: StorageKind::Memory,
+        ..Default::default()
+    };
+    let backend: DynStore = Arc::new(InMemory::new());
+    let storage = StorageManager::with_backend(backend, StorageKind::Memory);
+
+    let pipeline_config = ingestion_config.clone();
+    let pipeline = IngestionPipeline::new_with_config(
+        db,
+        openai.clone(),
+        app_config,
+        None::<Arc<retrieval_pipeline::reranking::RerankerPool>>,
+        storage,
+        embedding.clone(),
+        pipeline_config,
+    )?;
+    let pipeline = Arc::new(pipeline);
+
+    let mut shards = Vec::with_capacity(targets.len());
+    let category = dataset.metadata.category.clone();
+    for (batch_index, batch) in targets.chunks(batch_size).enumerate() {
+        info!(
+            batch = batch_index,
+            batch_size = batch.len(),
+            total_batches = targets.len().div_ceil(batch_size),
+            "Ingesting paragraph batch"
+        );
+        let model_clone = embedding_model.clone();
+        let backend_clone = embedding_backend.to_string();
+        let pipeline_clone = pipeline.clone();
+        let category_clone = category.clone();
+        let tasks = batch.iter().cloned().map(move |request| {
+            ingest_single_paragraph(
+                pipeline_clone.clone(),
+                request,
+                category_clone.clone(),
+                user_id,
+                ingestion_fingerprint,
+                backend_clone.clone(),
+                model_clone.clone(),
+                embedding_dimension,
+                max_retries,
+                ingestion_config.tuning.chunk_min_tokens,
+                ingestion_config.tuning.chunk_max_tokens,
+                ingestion_config.chunk_only,
+            )
+        });
+        let batch_results: Vec<ParagraphShard> = try_join_all(tasks)
+            .await
+            .context("ingesting batch of paragraphs")?;
+        shards.extend(batch_results);
+    }
+
+    Ok(shards)
+}
+
+#[cfg(test)]
+async fn create_ingest_db(namespace: &str) -> Result<Arc<SurrealDbClient>> {
+    let db = SurrealDbClient::memory(namespace, "corpus")
+        .await
+        .context("creating in-memory surrealdb for ingestion")?;
+    Ok(Arc::new(db))
+}
+
+#[cfg(not(test))]
+async fn create_ingest_db(namespace: &str) -> Result<Arc<SurrealDbClient>> {
+    let config = get_config().context("loading app config for ingestion database")?;
+    let db = SurrealDbClient::new(
+        &config.surrealdb_address,
+        &config.surrealdb_username,
+        &config.surrealdb_password,
+        namespace,
+        "corpus",
+    )
+    .await
+    .context("creating surrealdb database for ingestion")?;
+    Ok(Arc::new(db))
+}
+
+#[allow(clippy::too_many_arguments)]
+async fn ingest_single_paragraph(
+    pipeline: Arc<IngestionPipeline>,
+    request: IngestRequest<'_>,
+    category: String,
+    user_id: &str,
+    ingestion_fingerprint: &str,
+    embedding_backend: String,
+    embedding_model: Option<String>,
+    embedding_dimension: usize,
+    max_retries: usize,
+    chunk_min_tokens: usize,
+    chunk_max_tokens: usize,
+    chunk_only: bool,
+) -> Result<ParagraphShard> {
+    let paragraph = request.paragraph;
+    let mut last_err: Option<anyhow::Error> = None;
+    for attempt in 1..=max_retries {
+        let payload = IngestionPayload::Text {
+            text: paragraph.context.clone(),
+            context: paragraph.title.clone(),
+            category: category.clone(),
+            user_id: user_id.to_string(),
+        };
+        let task = IngestionTask::new(payload, user_id.to_string());
+        match pipeline.produce_artifacts(&task).await {
+            Ok(artifacts) => {
+                let entities: Vec<EmbeddedKnowledgeEntity> = artifacts
+                    .entities
+                    .into_iter()
+                    .map(|e| EmbeddedKnowledgeEntity {
+                        entity: e.entity,
+                        embedding: e.embedding,
+                    })
+                    .collect();
+                let chunks: Vec<EmbeddedTextChunk> = artifacts
+                    .chunks
+                    .into_iter()
+                    .map(|c| EmbeddedTextChunk {
+                        chunk: c.chunk,
+                        embedding: c.embedding,
+                    })
+                    .collect();
+                // No need to reembed - pipeline now uses FastEmbed internally
+                let mut shard = ParagraphShard::new(
+                    paragraph,
+                    request.shard_path,
+                    ingestion_fingerprint,
+                    artifacts.text_content,
+                    entities,
+                    artifacts.relationships,
+                    chunks,
+                    &embedding_backend,
+                    embedding_model.clone(),
+                    embedding_dimension,
+                    chunk_min_tokens,
+                    chunk_max_tokens,
+                    chunk_only,
+                );
+                for question in &request.question_refs {
+                    if let Err(err) = shard.ensure_question_binding(question) {
+                        warn!(
+                            question_id = %question.id,
+                            paragraph_id = %paragraph.id,
+                            error = %err,
+                            "Failed to locate answer text in ingested content; recording empty chunk bindings"
+                        );
+                        shard
+                            .question_bindings
+                            .insert(question.id.clone(), Vec::new());
+                    }
+                }
+                return Ok(shard);
+            }
+            Err(err) => {
+                warn!(
+                    paragraph_id = %paragraph.id,
+                    attempt,
+                    max_attempts = max_retries,
+                    error = ?err,
+                    "ingestion attempt failed for paragraph; retrying"
+                );
+                last_err = Some(err.into());
+            }
+        }
+    }
+
+    Err(last_err
+        .unwrap_or_else(|| anyhow!("ingestion failed"))
+        .context(format!("running ingestion for paragraph {}", paragraph.id)))
+}
+
+pub fn cached_corpus_dir(cache: &CorpusCacheConfig, dataset_id: &str, slice_id: &str) -> PathBuf {
+    cache.ingestion_cache_dir.join(dataset_id).join(slice_id)
+}
+
+pub fn build_ingestion_fingerprint(
+    dataset: &ConvertedDataset,
+    slice: &ResolvedSlice<'_>,
+    checksum: &str,
+    ingestion_config: &IngestionConfig,
+) -> String {
+    let config_repr = format!("{:?}", ingestion_config);
+    let mut hasher = Sha256::new();
+    hasher.update(config_repr.as_bytes());
+    let config_hash = format!("{:x}", hasher.finalize());
+
+    format!(
+        "v{INGESTION_SPEC_VERSION}:{}:{}:{}:{}:{}",
+        dataset.metadata.id,
+        slice.manifest.slice_id,
+        slice.manifest.includes_unanswerable,
+        checksum,
+        config_hash
+    )
+}
+
+pub fn compute_ingestion_fingerprint(
+    dataset: &ConvertedDataset,
+    slice: &ResolvedSlice<'_>,
+    converted_path: &Path,
+    ingestion_config: &IngestionConfig,
+) -> Result<String> {
+    let checksum = compute_file_checksum(converted_path)?;
+    Ok(build_ingestion_fingerprint(
+        dataset,
+        slice,
+        &checksum,
+        ingestion_config,
+    ))
+}
+
+pub fn load_cached_manifest(base_dir: &Path) -> Result<Option<CorpusManifest>> {
+    let path = base_dir.join("manifest.json");
+    if !path.exists() {
+        return Ok(None);
+    }
+    let mut file = fs::File::open(&path)
+        .with_context(|| format!("opening cached manifest {}", path.display()))?;
+    let mut buf = Vec::new();
+    file.read_to_end(&mut buf)
+        .with_context(|| format!("reading cached manifest {}", path.display()))?;
+    let manifest: CorpusManifest = serde_json::from_slice(&buf)
+        .with_context(|| format!("deserialising cached manifest {}", path.display()))?;
+    Ok(Some(manifest))
+}
+
+fn persist_manifest(handle: &CorpusHandle) -> Result<()> {
+    let path = handle.path.join("manifest.json");
+    if let Some(parent) = path.parent() {
+        fs::create_dir_all(parent)
+            .with_context(|| format!("creating manifest directory {}", parent.display()))?;
+    }
+    let tmp_path = path.with_extension("json.tmp");
+    let blob =
+        serde_json::to_vec_pretty(&handle.manifest).context("serialising corpus manifest")?;
+    fs::write(&tmp_path, &blob)
+        .with_context(|| format!("writing temporary manifest {}", tmp_path.display()))?;
+    fs::rename(&tmp_path, &path)
+        .with_context(|| format!("replacing manifest {}", path.display()))?;
+    Ok(())
+}
+
+pub fn corpus_handle_from_manifest(manifest: CorpusManifest, base_dir: PathBuf) -> CorpusHandle {
+    CorpusHandle {
+        manifest,
+        path: base_dir,
+        reused_ingestion: true,
+        reused_embeddings: true,
+        positive_reused: 0,
+        positive_ingested: 0,
+        negative_reused: 0,
+        negative_ingested: 0,
+    }
+}
+
+fn compute_file_checksum(path: &Path) -> Result<String> {
+    let mut file = fs::File::open(path)
+        .with_context(|| format!("opening file {} for checksum", path.display()))?;
+    let mut hasher = Sha256::new();
+    let mut buffer = [0u8; 8192];
+    loop {
+        let read = file
+            .read(&mut buffer)
+            .with_context(|| format!("reading {} for checksum", path.display()))?;
+        if read == 0 {
+            break;
+        }
+        hasher.update(&buffer[..read]);
+    }
+    Ok(format!("{:x}", hasher.finalize()))
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::{
+        datasets::{ConvertedDataset, ConvertedParagraph, ConvertedQuestion, DatasetKind},
+        slice::{CaseRef, SliceCaseEntry, SliceManifest, SliceParagraphEntry, SliceParagraphKind},
+    };
+    use chrono::Utc;
+
+    fn dummy_dataset() -> ConvertedDataset {
+        let question = ConvertedQuestion {
+            id: "q1".to_string(),
+            question: "What?".to_string(),
+            answers: vec!["A".to_string()],
+            is_impossible: false,
+        };
+        let paragraph = ConvertedParagraph {
+            id: "p1".to_string(),
+            title: "title".to_string(),
+            context: "context".to_string(),
+            questions: vec![question],
+        };
+
+        ConvertedDataset {
+            generated_at: Utc::now(),
+            metadata: crate::datasets::DatasetMetadata::for_kind(
+                DatasetKind::default(),
+                false,
+                None,
+            ),
+            source: "src".to_string(),
+            paragraphs: vec![paragraph],
+        }
+    }
+
+    fn dummy_slice<'a>(dataset: &'a ConvertedDataset) -> ResolvedSlice<'a> {
+        let paragraph = &dataset.paragraphs[0];
+        let question = &paragraph.questions[0];
+        let manifest = SliceManifest {
+            version: 1,
+            slice_id: "slice-1".to_string(),
+            dataset_id: dataset.metadata.id.clone(),
+            dataset_label: dataset.metadata.label.clone(),
+            dataset_source: dataset.source.clone(),
+            includes_unanswerable: false,
+            require_verified_chunks: false,
+            seed: 1,
+            requested_limit: Some(1),
+            requested_corpus: 1,
+            generated_at: Utc::now(),
+            case_count: 1,
+            positive_paragraphs: 1,
+            negative_paragraphs: 0,
+            total_paragraphs: 1,
+            negative_multiplier: 1.0,
+            cases: vec![SliceCaseEntry {
+                question_id: question.id.clone(),
+                paragraph_id: paragraph.id.clone(),
+            }],
+            paragraphs: vec![SliceParagraphEntry {
+                id: paragraph.id.clone(),
+                kind: SliceParagraphKind::Positive {
+                    question_ids: vec![question.id.clone()],
+                },
+                shard_path: None,
+            }],
+        };
+
+        ResolvedSlice {
+            manifest,
+            path: PathBuf::from("cache"),
+            paragraphs: dataset.paragraphs.iter().collect(),
+            cases: vec![CaseRef {
+                paragraph,
+                question,
+            }],
+        }
+    }
+
+    #[test]
+    fn fingerprint_changes_with_chunk_settings() {
+        let dataset = dummy_dataset();
+        let slice = dummy_slice(&dataset);
+        let checksum = "deadbeef";
+
+        let base_config = IngestionConfig::default();
+        let fp_base = build_ingestion_fingerprint(&dataset, &slice, checksum, &base_config);
+
+        let mut token_config = base_config.clone();
+        token_config.tuning.chunk_min_tokens += 1;
+        let fp_token = build_ingestion_fingerprint(&dataset, &slice, checksum, &token_config);
+        assert_ne!(fp_base, fp_token, "token bounds should affect fingerprint");
+
+        let mut chunk_only_config = base_config;
+        chunk_only_config.chunk_only = true;
+        let fp_chunk_only =
+            build_ingestion_fingerprint(&dataset, &slice, checksum, &chunk_only_config);
+        assert_ne!(
+            fp_base, fp_chunk_only,
+            "chunk-only mode should affect fingerprint"
+        );
+    }
+}

evaluations/src/corpus/store.rs

@@ -0,0 +1,934 @@
+use std::{
+    collections::{HashMap, HashSet},
+    fs,
+    io::BufReader,
+    path::PathBuf,
+};
+
+use anyhow::{anyhow, Context, Result};
+use chrono::{DateTime, Utc};
+use common::storage::types::StoredObject;
+use common::storage::{
+    db::SurrealDbClient,
+    types::{
+        knowledge_entity::KnowledgeEntity,
+        knowledge_entity_embedding::KnowledgeEntityEmbedding,
+        knowledge_relationship::{KnowledgeRelationship, RelationshipMetadata},
+        text_chunk::TextChunk,
+        text_chunk_embedding::TextChunkEmbedding,
+        text_content::TextContent,
+    },
+};
+use serde::Deserialize;
+use serde::Serialize;
+use surrealdb::sql::Thing;
+use tracing::{debug, warn};
+
+use crate::datasets::{ConvertedParagraph, ConvertedQuestion};
+
+pub const MANIFEST_VERSION: u32 = 3;
+pub const PARAGRAPH_SHARD_VERSION: u32 = 3;
+const MANIFEST_BATCH_SIZE: usize = 100;
+const MANIFEST_MAX_BYTES_PER_BATCH: usize = 300_000; // default cap for non-text batches
+const TEXT_CONTENT_MAX_BYTES_PER_BATCH: usize = 250_000; // text bodies can be large; limit aggressively
+const MAX_BATCHES_PER_REQUEST: usize = 24;
+const REQUEST_MAX_BYTES: usize = 800_000; // total payload cap per Surreal query request
+
+fn current_manifest_version() -> u32 {
+    MANIFEST_VERSION
+}
+
+fn current_paragraph_shard_version() -> u32 {
+    PARAGRAPH_SHARD_VERSION
+}
+
+fn default_chunk_min_tokens() -> usize {
+    500
+}
+
+fn default_chunk_max_tokens() -> usize {
+    2_000
+}
+
+fn default_chunk_only() -> bool {
+    false
+}
+
+#[derive(Debug, Clone, serde::Serialize, serde::Deserialize)]
+pub struct EmbeddedKnowledgeEntity {
+    pub entity: KnowledgeEntity,
+    pub embedding: Vec<f32>,
+}
+
+#[derive(Debug, Clone, serde::Serialize, serde::Deserialize)]
+pub struct EmbeddedTextChunk {
+    pub chunk: TextChunk,
+    pub embedding: Vec<f32>,
+}
+
+#[derive(Debug, Clone, serde::Deserialize)]
+struct LegacyKnowledgeEntity {
+    #[serde(flatten)]
+    pub entity: KnowledgeEntity,
+    #[serde(default)]
+    pub embedding: Vec<f32>,
+}
+
+#[derive(Debug, Clone, serde::Deserialize)]
+struct LegacyTextChunk {
+    #[serde(flatten)]
+    pub chunk: TextChunk,
+    #[serde(default)]
+    pub embedding: Vec<f32>,
+}
+
+fn deserialize_embedded_entities<'de, D>(
+    deserializer: D,
+) -> Result<Vec<EmbeddedKnowledgeEntity>, D::Error>
+where
+    D: serde::Deserializer<'de>,
+{
+    #[derive(serde::Deserialize)]
+    #[serde(untagged)]
+    enum EntityInput {
+        Embedded(Vec<EmbeddedKnowledgeEntity>),
+        Legacy(Vec<LegacyKnowledgeEntity>),
+    }
+
+    match EntityInput::deserialize(deserializer)? {
+        EntityInput::Embedded(items) => Ok(items),
+        EntityInput::Legacy(items) => Ok(items
+            .into_iter()
+            .map(|legacy| EmbeddedKnowledgeEntity {
+                entity: legacy.entity,
+                embedding: legacy.embedding,
+            })
+            .collect()),
+    }
+}
+
+fn deserialize_embedded_chunks<'de, D>(deserializer: D) -> Result<Vec<EmbeddedTextChunk>, D::Error>
+where
+    D: serde::Deserializer<'de>,
+{
+    #[derive(serde::Deserialize)]
+    #[serde(untagged)]
+    enum ChunkInput {
+        Embedded(Vec<EmbeddedTextChunk>),
+        Legacy(Vec<LegacyTextChunk>),
+    }
+
+    match ChunkInput::deserialize(deserializer)? {
+        ChunkInput::Embedded(items) => Ok(items),
+        ChunkInput::Legacy(items) => Ok(items
+            .into_iter()
+            .map(|legacy| EmbeddedTextChunk {
+                chunk: legacy.chunk,
+                embedding: legacy.embedding,
+            })
+            .collect()),
+    }
+}
+
+#[derive(Debug, Clone, serde::Serialize, serde::Deserialize)]
+pub struct CorpusManifest {
+    #[serde(default = "current_manifest_version")]
+    pub version: u32,
+    pub metadata: CorpusMetadata,
+    pub paragraphs: Vec<CorpusParagraph>,
+    pub questions: Vec<CorpusQuestion>,
+}
+
+#[derive(Debug, Clone, serde::Serialize, serde::Deserialize)]
+pub struct CorpusMetadata {
+    pub dataset_id: String,
+    pub dataset_label: String,
+    pub slice_id: String,
+    pub include_unanswerable: bool,
+    #[serde(default)]
+    pub require_verified_chunks: bool,
+    pub ingestion_fingerprint: String,
+    pub embedding_backend: String,
+    pub embedding_model: Option<String>,
+    pub embedding_dimension: usize,
+    pub converted_checksum: String,
+    pub generated_at: DateTime<Utc>,
+    pub paragraph_count: usize,
+    pub question_count: usize,
+    #[serde(default = "default_chunk_min_tokens")]
+    pub chunk_min_tokens: usize,
+    #[serde(default = "default_chunk_max_tokens")]
+    pub chunk_max_tokens: usize,
+    #[serde(default = "default_chunk_only")]
+    pub chunk_only: bool,
+}
+
+#[derive(Debug, Clone, serde::Serialize, serde::Deserialize)]
+pub struct CorpusParagraph {
+    pub paragraph_id: String,
+    pub title: String,
+    pub text_content: TextContent,
+    #[serde(deserialize_with = "deserialize_embedded_entities")]
+    pub entities: Vec<EmbeddedKnowledgeEntity>,
+    pub relationships: Vec<KnowledgeRelationship>,
+    #[serde(deserialize_with = "deserialize_embedded_chunks")]
+    pub chunks: Vec<EmbeddedTextChunk>,
+}
+
+#[derive(Debug, Clone, serde::Serialize, serde::Deserialize)]
+pub struct CorpusQuestion {
+    pub question_id: String,
+    pub paragraph_id: String,
+    pub text_content_id: String,
+    pub question_text: String,
+    pub answers: Vec<String>,
+    pub is_impossible: bool,
+    pub matching_chunk_ids: Vec<String>,
+}
+
+pub struct CorpusHandle {
+    pub manifest: CorpusManifest,
+    pub path: PathBuf,
+    pub reused_ingestion: bool,
+    pub reused_embeddings: bool,
+    pub positive_reused: usize,
+    pub positive_ingested: usize,
+    pub negative_reused: usize,
+    pub negative_ingested: usize,
+}
+
+pub fn window_manifest(
+    manifest: &CorpusManifest,
+    offset: usize,
+    length: usize,
+    negative_multiplier: f32,
+) -> Result<CorpusManifest> {
+    let total = manifest.questions.len();
+    if total == 0 {
+        return Err(anyhow!(
+            "manifest contains no questions; cannot select a window"
+        ));
+    }
+    if offset >= total {
+        return Err(anyhow!(
+            "window offset {} exceeds manifest questions ({})",
+            offset,
+            total
+        ));
+    }
+    let end = (offset + length).min(total);
+    let questions = manifest.questions[offset..end].to_vec();
+
+    let selected_positive_ids: HashSet<_> =
+        questions.iter().map(|q| q.paragraph_id.clone()).collect();
+    let positives_all: HashSet<_> = manifest
+        .questions
+        .iter()
+        .map(|q| q.paragraph_id.as_str())
+        .collect();
+    let available_negatives = manifest
+        .paragraphs
+        .len()
+        .saturating_sub(positives_all.len());
+    let desired_negatives =
+        ((selected_positive_ids.len() as f32) * negative_multiplier).ceil() as usize;
+    let desired_negatives = desired_negatives.min(available_negatives);
+
+    let mut paragraphs = Vec::new();
+    let mut negative_count = 0usize;
+    for paragraph in &manifest.paragraphs {
+        if selected_positive_ids.contains(&paragraph.paragraph_id) {
+            paragraphs.push(paragraph.clone());
+        } else if negative_count < desired_negatives {
+            paragraphs.push(paragraph.clone());
+            negative_count += 1;
+        }
+    }
+
+    let mut narrowed = manifest.clone();
+    narrowed.questions = questions;
+    narrowed.paragraphs = paragraphs;
+    narrowed.metadata.paragraph_count = narrowed.paragraphs.len();
+    narrowed.metadata.question_count = narrowed.questions.len();
+
+    Ok(narrowed)
+}
+
+#[derive(Debug, Clone, Serialize)]
+struct RelationInsert {
+    #[serde(rename = "in")]
+    pub in_: Thing,
+    #[serde(rename = "out")]
+    pub out: Thing,
+    pub id: String,
+    pub metadata: RelationshipMetadata,
+}
+
+#[derive(Debug)]
+struct SizedBatch<T> {
+    approx_bytes: usize,
+    items: Vec<T>,
+}
+
+struct ManifestBatches {
+    text_contents: Vec<SizedBatch<TextContent>>,
+    entities: Vec<SizedBatch<KnowledgeEntity>>,
+    entity_embeddings: Vec<SizedBatch<KnowledgeEntityEmbedding>>,
+    relationships: Vec<SizedBatch<RelationInsert>>,
+    chunks: Vec<SizedBatch<TextChunk>>,
+    chunk_embeddings: Vec<SizedBatch<TextChunkEmbedding>>,
+}
+
+fn build_manifest_batches(manifest: &CorpusManifest) -> Result<ManifestBatches> {
+    let mut text_contents = Vec::new();
+    let mut entities = Vec::new();
+    let mut entity_embeddings = Vec::new();
+    let mut relationships = Vec::new();
+    let mut chunks = Vec::new();
+    let mut chunk_embeddings = Vec::new();
+
+    let mut seen_text_content = HashSet::new();
+    let mut seen_entities = HashSet::new();
+    let mut seen_relationships = HashSet::new();
+    let mut seen_chunks = HashSet::new();
+
+    for paragraph in &manifest.paragraphs {
+        if seen_text_content.insert(paragraph.text_content.id.clone()) {
+            text_contents.push(paragraph.text_content.clone());
+        }
+
+        for embedded_entity in &paragraph.entities {
+            if seen_entities.insert(embedded_entity.entity.id.clone()) {
+                let entity = embedded_entity.entity.clone();
+                entities.push(entity.clone());
+                entity_embeddings.push(KnowledgeEntityEmbedding::new(
+                    &entity.id,
+                    embedded_entity.embedding.clone(),
+                    entity.user_id.clone(),
+                ));
+            }
+        }
+
+        for relationship in &paragraph.relationships {
+            if seen_relationships.insert(relationship.id.clone()) {
+                let table = KnowledgeEntity::table_name();
+                let in_id = relationship
+                    .in_
+                    .strip_prefix(&format!("{table}:"))
+                    .unwrap_or(&relationship.in_);
+                let out_id = relationship
+                    .out
+                    .strip_prefix(&format!("{table}:"))
+                    .unwrap_or(&relationship.out);
+                let in_thing = Thing::from((table, in_id));
+                let out_thing = Thing::from((table, out_id));
+                relationships.push(RelationInsert {
+                    in_: in_thing,
+                    out: out_thing,
+                    id: relationship.id.clone(),
+                    metadata: relationship.metadata.clone(),
+                });
+            }
+        }
+
+        for embedded_chunk in &paragraph.chunks {
+            if seen_chunks.insert(embedded_chunk.chunk.id.clone()) {
+                let chunk = embedded_chunk.chunk.clone();
+                chunks.push(chunk.clone());
+                chunk_embeddings.push(TextChunkEmbedding::new(
+                    &chunk.id,
+                    chunk.source_id.clone(),
+                    embedded_chunk.embedding.clone(),
+                    chunk.user_id.clone(),
+                ));
+            }
+        }
+    }
+
+    Ok(ManifestBatches {
+        text_contents: chunk_items(
+            &text_contents,
+            MANIFEST_BATCH_SIZE,
+            TEXT_CONTENT_MAX_BYTES_PER_BATCH,
+        )
+        .context("chunking text_content payloads")?,
+        entities: chunk_items(&entities, MANIFEST_BATCH_SIZE, MANIFEST_MAX_BYTES_PER_BATCH)
+            .context("chunking knowledge_entity payloads")?,
+        entity_embeddings: chunk_items(
+            &entity_embeddings,
+            MANIFEST_BATCH_SIZE,
+            MANIFEST_MAX_BYTES_PER_BATCH,
+        )
+        .context("chunking knowledge_entity_embedding payloads")?,
+        relationships: chunk_items(
+            &relationships,
+            MANIFEST_BATCH_SIZE,
+            MANIFEST_MAX_BYTES_PER_BATCH,
+        )
+        .context("chunking relationship payloads")?,
+        chunks: chunk_items(&chunks, MANIFEST_BATCH_SIZE, MANIFEST_MAX_BYTES_PER_BATCH)
+            .context("chunking text_chunk payloads")?,
+        chunk_embeddings: chunk_items(
+            &chunk_embeddings,
+            MANIFEST_BATCH_SIZE,
+            MANIFEST_MAX_BYTES_PER_BATCH,
+        )
+        .context("chunking text_chunk_embedding payloads")?,
+    })
+}
+
+#[derive(Debug, Clone, serde::Serialize, serde::Deserialize)]
+pub struct ParagraphShard {
+    #[serde(default = "current_paragraph_shard_version")]
+    pub version: u32,
+    pub paragraph_id: String,
+    pub shard_path: String,
+    pub ingestion_fingerprint: String,
+    pub ingested_at: DateTime<Utc>,
+    pub title: String,
+    pub text_content: TextContent,
+    #[serde(deserialize_with = "deserialize_embedded_entities")]
+    pub entities: Vec<EmbeddedKnowledgeEntity>,
+    pub relationships: Vec<KnowledgeRelationship>,
+    #[serde(deserialize_with = "deserialize_embedded_chunks")]
+    pub chunks: Vec<EmbeddedTextChunk>,
+    #[serde(default)]
+    pub question_bindings: HashMap<String, Vec<String>>,
+    #[serde(default)]
+    pub embedding_backend: String,
+    #[serde(default)]
+    pub embedding_model: Option<String>,
+    #[serde(default)]
+    pub embedding_dimension: usize,
+    #[serde(default = "default_chunk_min_tokens")]
+    pub chunk_min_tokens: usize,
+    #[serde(default = "default_chunk_max_tokens")]
+    pub chunk_max_tokens: usize,
+    #[serde(default = "default_chunk_only")]
+    pub chunk_only: bool,
+}
+
+pub struct ParagraphShardStore {
+    base_dir: PathBuf,
+}
+
+impl ParagraphShardStore {
+    pub fn new(base_dir: PathBuf) -> Self {
+        Self { base_dir }
+    }
+
+    pub fn ensure_base_dir(&self) -> Result<()> {
+        fs::create_dir_all(&self.base_dir)
+            .with_context(|| format!("creating shard base dir {}", self.base_dir.display()))
+    }
+
+    fn resolve(&self, relative: &str) -> PathBuf {
+        self.base_dir.join(relative)
+    }
+
+    pub fn load(&self, relative: &str, fingerprint: &str) -> Result<Option<ParagraphShard>> {
+        let path = self.resolve(relative);
+        let file = match fs::File::open(&path) {
+            Ok(file) => file,
+            Err(err) if err.kind() == std::io::ErrorKind::NotFound => return Ok(None),
+            Err(err) => {
+                return Err(err).with_context(|| format!("opening shard {}", path.display()))
+            }
+        };
+        let reader = BufReader::new(file);
+        let mut shard: ParagraphShard = serde_json::from_reader(reader)
+            .with_context(|| format!("parsing shard {}", path.display()))?;
+
+        if shard.ingestion_fingerprint != fingerprint {
+            debug!(
+                path = %path.display(),
+                expected = fingerprint,
+                found = shard.ingestion_fingerprint,
+                "Shard fingerprint mismatch; will rebuild"
+            );
+            return Ok(None);
+        }
+        if shard.version != PARAGRAPH_SHARD_VERSION {
+            warn!(
+                path = %path.display(),
+                version = shard.version,
+                expected = PARAGRAPH_SHARD_VERSION,
+                "Upgrading shard to current version"
+            );
+            shard.version = PARAGRAPH_SHARD_VERSION;
+        }
+        shard.shard_path = relative.to_string();
+        Ok(Some(shard))
+    }
+
+    pub fn persist(&self, shard: &ParagraphShard) -> Result<()> {
+        let mut shard = shard.clone();
+        shard.version = PARAGRAPH_SHARD_VERSION;
+
+        let path = self.resolve(&shard.shard_path);
+        if let Some(parent) = path.parent() {
+            fs::create_dir_all(parent)
+                .with_context(|| format!("creating shard dir {}", parent.display()))?;
+        }
+        let tmp_path = path.with_extension("json.tmp");
+        let body = serde_json::to_vec_pretty(&shard).context("serialising paragraph shard")?;
+        fs::write(&tmp_path, &body)
+            .with_context(|| format!("writing shard tmp {}", tmp_path.display()))?;
+        fs::rename(&tmp_path, &path)
+            .with_context(|| format!("renaming shard tmp {}", path.display()))?;
+        Ok(())
+    }
+}
+
+impl ParagraphShard {
+    #[allow(clippy::too_many_arguments)]
+    pub fn new(
+        paragraph: &ConvertedParagraph,
+        shard_path: String,
+        ingestion_fingerprint: &str,
+        text_content: TextContent,
+        entities: Vec<EmbeddedKnowledgeEntity>,
+        relationships: Vec<KnowledgeRelationship>,
+        chunks: Vec<EmbeddedTextChunk>,
+        embedding_backend: &str,
+        embedding_model: Option<String>,
+        embedding_dimension: usize,
+        chunk_min_tokens: usize,
+        chunk_max_tokens: usize,
+        chunk_only: bool,
+    ) -> Self {
+        Self {
+            version: PARAGRAPH_SHARD_VERSION,
+            paragraph_id: paragraph.id.clone(),
+            shard_path,
+            ingestion_fingerprint: ingestion_fingerprint.to_string(),
+            ingested_at: Utc::now(),
+            title: paragraph.title.clone(),
+            text_content,
+            entities,
+            relationships,
+            chunks,
+            question_bindings: HashMap::new(),
+            embedding_backend: embedding_backend.to_string(),
+            embedding_model,
+            embedding_dimension,
+            chunk_min_tokens,
+            chunk_max_tokens,
+            chunk_only,
+        }
+    }
+
+    pub fn to_corpus_paragraph(&self) -> CorpusParagraph {
+        CorpusParagraph {
+            paragraph_id: self.paragraph_id.clone(),
+            title: self.title.clone(),
+            text_content: self.text_content.clone(),
+            entities: self.entities.clone(),
+            relationships: self.relationships.clone(),
+            chunks: self.chunks.clone(),
+        }
+    }
+
+    pub fn ensure_question_binding(
+        &mut self,
+        question: &ConvertedQuestion,
+    ) -> Result<(Vec<String>, bool)> {
+        if let Some(existing) = self.question_bindings.get(&question.id) {
+            return Ok((existing.clone(), false));
+        }
+        let chunk_ids = validate_answers(&self.text_content, &self.chunks, question)?;
+        self.question_bindings
+            .insert(question.id.clone(), chunk_ids.clone());
+        Ok((chunk_ids, true))
+    }
+}
+
+fn validate_answers(
+    content: &TextContent,
+    chunks: &[EmbeddedTextChunk],
+    question: &ConvertedQuestion,
+) -> Result<Vec<String>> {
+    if question.is_impossible || question.answers.is_empty() {
+        return Ok(Vec::new());
+    }
+
+    let mut matches = std::collections::BTreeSet::new();
+    let mut found_any = false;
+    let haystack = content.text.to_ascii_lowercase();
+    let haystack_norm = normalize_answer_text(&haystack);
+    for answer in &question.answers {
+        let needle: String = answer.to_ascii_lowercase();
+        let needle_norm = normalize_answer_text(&needle);
+        let text_match = haystack.contains(&needle)
+            || (!needle_norm.is_empty() && haystack_norm.contains(&needle_norm));
+        if text_match {
+            found_any = true;
+        }
+        for chunk in chunks {
+            let chunk_text = chunk.chunk.chunk.to_ascii_lowercase();
+            let chunk_norm = normalize_answer_text(&chunk_text);
+            if chunk_text.contains(&needle)
+                || (!needle_norm.is_empty() && chunk_norm.contains(&needle_norm))
+            {
+                matches.insert(chunk.chunk.get_id().to_string());
+                found_any = true;
+            }
+        }
+    }
+
+    if !found_any {
+        Err(anyhow!(
+            "expected answer for question '{}' was not found in ingested content",
+            question.id
+        ))
+    } else {
+        Ok(matches.into_iter().collect())
+    }
+}
+
+fn normalize_answer_text(text: &str) -> String {
+    text.chars()
+        .map(|ch| {
+            if ch.is_alphanumeric() || ch.is_whitespace() {
+                ch.to_ascii_lowercase()
+            } else {
+                ' '
+            }
+        })
+        .collect::<String>()
+        .split_whitespace()
+        .collect::<Vec<_>>()
+        .join(" ")
+}
+
+fn chunk_items<T: Clone + Serialize>(
+    items: &[T],
+    max_items: usize,
+    max_bytes: usize,
+) -> Result<Vec<SizedBatch<T>>> {
+    if items.is_empty() {
+        return Ok(Vec::new());
+    }
+
+    let mut batches = Vec::new();
+    let mut current = Vec::new();
+    let mut current_bytes = 0usize;
+
+    for item in items {
+        let size = serde_json::to_vec(item)
+            .map(|buf| buf.len())
+            .context("serialising batch item for sizing")?;
+
+        let would_overflow_items = !current.is_empty() && current.len() >= max_items;
+        let would_overflow_bytes = !current.is_empty() && current_bytes + size > max_bytes;
+
+        if would_overflow_items || would_overflow_bytes {
+            batches.push(SizedBatch {
+                approx_bytes: current_bytes.max(1),
+                items: std::mem::take(&mut current),
+            });
+            current_bytes = 0;
+        }
+
+        current_bytes += size;
+        current.push(item.clone());
+    }
+
+    if !current.is_empty() {
+        batches.push(SizedBatch {
+            approx_bytes: current_bytes.max(1),
+            items: current,
+        });
+    }
+
+    Ok(batches)
+}
+
+async fn execute_batched_inserts<T: Clone + Serialize + 'static>(
+    db: &SurrealDbClient,
+    statement: impl AsRef<str>,
+    prefix: &str,
+    batches: &[SizedBatch<T>],
+) -> Result<()> {
+    if batches.is_empty() {
+        return Ok(());
+    }
+
+    let mut start = 0;
+    while start < batches.len() {
+        let mut group_bytes = 0usize;
+        let mut group_end = start;
+        let mut group_count = 0usize;
+
+        while group_end < batches.len() {
+            let batch_bytes = batches[group_end].approx_bytes.max(1);
+            if group_count > 0
+                && (group_bytes + batch_bytes > REQUEST_MAX_BYTES
+                    || group_count >= MAX_BATCHES_PER_REQUEST)
+            {
+                break;
+            }
+            group_bytes += batch_bytes;
+            group_end += 1;
+            group_count += 1;
+        }
+
+        let slice = &batches[start..group_end];
+        let mut query = db.client.query("BEGIN TRANSACTION;");
+        for (bind_index, batch) in slice.iter().enumerate() {
+            let name = format!("{prefix}{bind_index}");
+            query = query
+                .query(format!("{} ${};", statement.as_ref(), name))
+                .bind((name, batch.items.clone()));
+        }
+        let response = query
+            .query("COMMIT TRANSACTION;")
+            .await
+            .context("executing batched insert transaction")?;
+        if let Err(err) = response.check() {
+            return Err(anyhow!(
+                "batched insert failed for statement '{}': {err:?}",
+                statement.as_ref()
+            ));
+        }
+
+        start = group_end;
+    }
+
+    Ok(())
+}
+
+pub async fn seed_manifest_into_db(db: &SurrealDbClient, manifest: &CorpusManifest) -> Result<()> {
+    let batches = build_manifest_batches(manifest).context("preparing manifest batches")?;
+
+    let result = async {
+        execute_batched_inserts(
+            db,
+            format!("INSERT INTO {}", TextContent::table_name()),
+            "tc",
+            &batches.text_contents,
+        )
+        .await?;
+
+        execute_batched_inserts(
+            db,
+            format!("INSERT INTO {}", KnowledgeEntity::table_name()),
+            "ke",
+            &batches.entities,
+        )
+        .await?;
+
+        execute_batched_inserts(
+            db,
+            format!("INSERT INTO {}", TextChunk::table_name()),
+            "ch",
+            &batches.chunks,
+        )
+        .await?;
+
+        execute_batched_inserts(
+            db,
+            "INSERT RELATION INTO relates_to",
+            "rel",
+            &batches.relationships,
+        )
+        .await?;
+
+        execute_batched_inserts(
+            db,
+            format!("INSERT INTO {}", KnowledgeEntityEmbedding::table_name()),
+            "kee",
+            &batches.entity_embeddings,
+        )
+        .await?;
+
+        execute_batched_inserts(
+            db,
+            format!("INSERT INTO {}", TextChunkEmbedding::table_name()),
+            "tce",
+            &batches.chunk_embeddings,
+        )
+        .await?;
+
+        Ok(())
+    }
+    .await;
+
+    if result.is_err() {
+        // Best-effort cleanup to avoid leaving partial manifest data behind.
+        let _ = db
+            .client
+            .query(
+                "BEGIN TRANSACTION;
+                 DELETE text_chunk_embedding;
+                 DELETE knowledge_entity_embedding;
+                 DELETE relates_to;
+                 DELETE text_chunk;
+                 DELETE knowledge_entity;
+                 DELETE text_content;
+                 COMMIT TRANSACTION;",
+            )
+            .await;
+    }
+
+    result
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use chrono::Utc;
+    use common::storage::types::knowledge_entity::KnowledgeEntityType;
+    use uuid::Uuid;
+
+    fn build_manifest() -> CorpusManifest {
+        let user_id = "user-1".to_string();
+        let source_id = "source-1".to_string();
+        let now = Utc::now();
+        let text_content_id = Uuid::new_v4().to_string();
+
+        let text_content = TextContent {
+            id: text_content_id.clone(),
+            created_at: now,
+            updated_at: now,
+            text: "Hello world".to_string(),
+            file_info: None,
+            url_info: None,
+            context: None,
+            category: "test".to_string(),
+            user_id: user_id.clone(),
+        };
+
+        let entity = KnowledgeEntity {
+            id: Uuid::new_v4().to_string(),
+            created_at: now,
+            updated_at: now,
+            source_id: source_id.clone(),
+            name: "Entity".to_string(),
+            description: "A test entity".to_string(),
+            entity_type: KnowledgeEntityType::Document,
+            metadata: None,
+            user_id: user_id.clone(),
+        };
+        let relationship = KnowledgeRelationship::new(
+            format!("knowledge_entity:{}", entity.id),
+            format!("knowledge_entity:{}", entity.id),
+            user_id.clone(),
+            source_id.clone(),
+            "related".to_string(),
+        );
+
+        let chunk = TextChunk {
+            id: Uuid::new_v4().to_string(),
+            created_at: now,
+            updated_at: now,
+            source_id: source_id.clone(),
+            chunk: "chunk text".to_string(),
+            user_id: user_id.clone(),
+        };
+
+        let paragraph_one = CorpusParagraph {
+            paragraph_id: "p1".to_string(),
+            title: "Paragraph 1".to_string(),
+            text_content: text_content.clone(),
+            entities: vec![EmbeddedKnowledgeEntity {
+                entity: entity.clone(),
+                embedding: vec![0.1, 0.2, 0.3],
+            }],
+            relationships: vec![relationship],
+            chunks: vec![EmbeddedTextChunk {
+                chunk: chunk.clone(),
+                embedding: vec![0.3, 0.2, 0.1],
+            }],
+        };
+
+        // Duplicate content/entities should be de-duplicated by the loader.
+        let paragraph_two = CorpusParagraph {
+            paragraph_id: "p2".to_string(),
+            title: "Paragraph 2".to_string(),
+            text_content: text_content.clone(),
+            entities: vec![EmbeddedKnowledgeEntity {
+                entity: entity.clone(),
+                embedding: vec![0.1, 0.2, 0.3],
+            }],
+            relationships: Vec::new(),
+            chunks: vec![EmbeddedTextChunk {
+                chunk: chunk.clone(),
+                embedding: vec![0.3, 0.2, 0.1],
+            }],
+        };
+
+        let question = CorpusQuestion {
+            question_id: "q1".to_string(),
+            paragraph_id: paragraph_one.paragraph_id.clone(),
+            text_content_id: text_content_id,
+            question_text: "What is this?".to_string(),
+            answers: vec!["Hello".to_string()],
+            is_impossible: false,
+            matching_chunk_ids: vec![chunk.id.clone()],
+        };
+
+        CorpusManifest {
+            version: current_manifest_version(),
+            metadata: CorpusMetadata {
+                dataset_id: "dataset".to_string(),
+                dataset_label: "Dataset".to_string(),
+                slice_id: "slice".to_string(),
+                include_unanswerable: false,
+                require_verified_chunks: false,
+                ingestion_fingerprint: "fp".to_string(),
+                embedding_backend: "test".to_string(),
+                embedding_model: Some("model".to_string()),
+                embedding_dimension: 3,
+                converted_checksum: "checksum".to_string(),
+                generated_at: now,
+                paragraph_count: 2,
+                question_count: 1,
+                chunk_min_tokens: 1,
+                chunk_max_tokens: 10,
+                chunk_only: false,
+            },
+            paragraphs: vec![paragraph_one, paragraph_two],
+            questions: vec![question],
+        }
+    }
+
+    #[test]
+    fn window_manifest_trims_questions_and_negatives() {
+        let manifest = build_manifest();
+        // Add extra negatives to simulate multiplier ~4x
+        let mut manifest = manifest;
+        let mut extra_paragraphs = Vec::new();
+        for _ in 0..8 {
+            let mut p = manifest.paragraphs[0].clone();
+            p.paragraph_id = Uuid::new_v4().to_string();
+            p.entities.clear();
+            p.relationships.clear();
+            p.chunks.clear();
+            extra_paragraphs.push(p);
+        }
+        manifest.paragraphs.extend(extra_paragraphs);
+        manifest.metadata.paragraph_count = manifest.paragraphs.len();
+
+        let windowed = window_manifest(&manifest, 0, 1, 4.0).expect("window manifest");
+        assert_eq!(windowed.questions.len(), 1);
+        // Expect roughly 4x negatives (bounded by available paragraphs)
+        assert!(
+            windowed.paragraphs.len() <= manifest.paragraphs.len(),
+            "windowed paragraphs should never exceed original"
+        );
+        let positive_set: std::collections::HashSet<_> = windowed
+            .questions
+            .iter()
+            .map(|q| q.paragraph_id.as_str())
+            .collect();
+        let positives = windowed
+            .paragraphs
+            .iter()
+            .filter(|p| positive_set.contains(p.paragraph_id.as_str()))
+            .count();
+        let negatives = windowed.paragraphs.len().saturating_sub(positives);
+        assert_eq!(positives, 1);
+        assert!(negatives >= 1, "should include some negatives");
+    }
+}

evaluations/src/datasets/beir.rs

@@ -0,0 +1,341 @@
+use std::{
+    collections::{BTreeMap, HashMap},
+    fs::File,
+    io::{BufRead, BufReader},
+    path::{Path, PathBuf},
+};
+
+use anyhow::{anyhow, Context, Result};
+use serde::Deserialize;
+use tracing::warn;
+
+use super::{ConvertedParagraph, ConvertedQuestion, DatasetKind};
+
+const ANSWER_SNIPPET_CHARS: usize = 240;
+
+#[derive(Debug, Deserialize)]
+struct BeirCorpusRow {
+    #[serde(rename = "_id")]
+    id: String,
+    #[serde(default)]
+    title: Option<String>,
+    #[serde(default)]
+    text: Option<String>,
+}
+
+#[derive(Debug, Deserialize)]
+struct BeirQueryRow {
+    #[serde(rename = "_id")]
+    id: String,
+    text: String,
+}
+
+#[derive(Debug, Clone)]
+struct BeirParagraph {
+    title: String,
+    context: String,
+}
+
+#[derive(Debug, Clone)]
+struct BeirQuery {
+    text: String,
+}
+
+#[derive(Debug, Clone)]
+struct QrelEntry {
+    doc_id: String,
+    score: i32,
+}
+
+pub fn convert_beir(raw_dir: &Path, dataset: DatasetKind) -> Result<Vec<ConvertedParagraph>> {
+    let corpus_path = raw_dir.join("corpus.jsonl");
+    let queries_path = raw_dir.join("queries.jsonl");
+    let qrels_path = resolve_qrels_path(raw_dir)?;
+
+    let corpus = load_corpus(&corpus_path)?;
+    let queries = load_queries(&queries_path)?;
+    let qrels = load_qrels(&qrels_path)?;
+
+    let mut paragraphs = Vec::with_capacity(corpus.len());
+    let mut paragraph_index = HashMap::new();
+
+    for (doc_id, entry) in corpus.iter() {
+        let paragraph_id = format!("{}-{doc_id}", dataset.source_prefix());
+        let paragraph = ConvertedParagraph {
+            id: paragraph_id.clone(),
+            title: entry.title.clone(),
+            context: entry.context.clone(),
+            questions: Vec::new(),
+        };
+        paragraph_index.insert(doc_id.clone(), paragraphs.len());
+        paragraphs.push(paragraph);
+    }
+
+    let mut missing_queries = 0usize;
+    let mut missing_docs = 0usize;
+    let mut skipped_answers = 0usize;
+
+    for (query_id, entries) in qrels {
+        let query = match queries.get(&query_id) {
+            Some(query) => query,
+            None => {
+                missing_queries += 1;
+                warn!(query_id = %query_id, "Skipping qrels entry for missing query");
+                continue;
+            }
+        };
+
+        let best = match select_best_doc(&entries) {
+            Some(entry) => entry,
+            None => continue,
+        };
+
+        let paragraph_slot = match paragraph_index.get(&best.doc_id) {
+            Some(slot) => *slot,
+            None => {
+                missing_docs += 1;
+                warn!(
+                    query_id = %query_id,
+                    doc_id = %best.doc_id,
+                    "Skipping qrels entry referencing missing corpus document"
+                );
+                continue;
+            }
+        };
+
+        let answer = answer_snippet(&paragraphs[paragraph_slot].context);
+        let answers = match answer {
+            Some(snippet) => vec![snippet],
+            None => {
+                skipped_answers += 1;
+                warn!(
+                    query_id = %query_id,
+                    doc_id = %best.doc_id,
+                    "Skipping query because no non-empty answer snippet could be derived"
+                );
+                continue;
+            }
+        };
+
+        let question_id = format!("{}-{query_id}", dataset.source_prefix());
+        paragraphs[paragraph_slot]
+            .questions
+            .push(ConvertedQuestion {
+                id: question_id,
+                question: query.text.clone(),
+                answers,
+                is_impossible: false,
+            });
+    }
+
+    if missing_queries + missing_docs + skipped_answers > 0 {
+        warn!(
+            missing_queries,
+            missing_docs, skipped_answers, "Skipped some BEIR qrels entries during conversion"
+        );
+    }
+
+    Ok(paragraphs)
+}
+
+fn resolve_qrels_path(raw_dir: &Path) -> Result<PathBuf> {
+    let qrels_dir = raw_dir.join("qrels");
+    let candidates = ["test.tsv", "dev.tsv", "train.tsv"];
+
+    for name in candidates {
+        let candidate = qrels_dir.join(name);
+        if candidate.exists() {
+            return Ok(candidate);
+        }
+    }
+
+    Err(anyhow!(
+        "No qrels file found under {}; expected one of {:?}",
+        qrels_dir.display(),
+        candidates
+    ))
+}
+
+fn load_corpus(path: &Path) -> Result<BTreeMap<String, BeirParagraph>> {
+    let file =
+        File::open(path).with_context(|| format!("opening BEIR corpus at {}", path.display()))?;
+    let reader = BufReader::new(file);
+    let mut corpus = BTreeMap::new();
+
+    for (idx, line) in reader.lines().enumerate() {
+        let raw = line
+            .with_context(|| format!("reading corpus line {} from {}", idx + 1, path.display()))?;
+        if raw.trim().is_empty() {
+            continue;
+        }
+        let row: BeirCorpusRow = serde_json::from_str(&raw).with_context(|| {
+            format!(
+                "parsing corpus JSON on line {} from {}",
+                idx + 1,
+                path.display()
+            )
+        })?;
+        let title = row.title.unwrap_or_else(|| row.id.clone());
+        let text = row.text.unwrap_or_default();
+        let context = build_context(&title, &text);
+
+        if context.is_empty() {
+            warn!(doc_id = %row.id, "Skipping empty corpus document");
+            continue;
+        }
+
+        corpus.insert(row.id, BeirParagraph { title, context });
+    }
+
+    Ok(corpus)
+}
+
+fn load_queries(path: &Path) -> Result<BTreeMap<String, BeirQuery>> {
+    let file = File::open(path)
+        .with_context(|| format!("opening BEIR queries file at {}", path.display()))?;
+    let reader = BufReader::new(file);
+    let mut queries = BTreeMap::new();
+
+    for (idx, line) in reader.lines().enumerate() {
+        let raw = line
+            .with_context(|| format!("reading query line {} from {}", idx + 1, path.display()))?;
+        if raw.trim().is_empty() {
+            continue;
+        }
+        let row: BeirQueryRow = serde_json::from_str(&raw).with_context(|| {
+            format!(
+                "parsing query JSON on line {} from {}",
+                idx + 1,
+                path.display()
+            )
+        })?;
+        queries.insert(
+            row.id,
+            BeirQuery {
+                text: row.text.trim().to_string(),
+            },
+        );
+    }
+
+    Ok(queries)
+}
+
+fn load_qrels(path: &Path) -> Result<BTreeMap<String, Vec<QrelEntry>>> {
+    let file =
+        File::open(path).with_context(|| format!("opening BEIR qrels at {}", path.display()))?;
+    let reader = BufReader::new(file);
+    let mut qrels: BTreeMap<String, Vec<QrelEntry>> = BTreeMap::new();
+
+    for (idx, line) in reader.lines().enumerate() {
+        let raw = line
+            .with_context(|| format!("reading qrels line {} from {}", idx + 1, path.display()))?;
+        let trimmed = raw.trim();
+        if trimmed.is_empty() || trimmed.starts_with("query-id") {
+            continue;
+        }
+        let mut parts = trimmed.split_whitespace();
+        let query_id = parts
+            .next()
+            .ok_or_else(|| anyhow!("missing query id on line {}", idx + 1))?;
+        let doc_id = parts
+            .next()
+            .ok_or_else(|| anyhow!("missing document id on line {}", idx + 1))?;
+        let score_raw = parts
+            .next()
+            .ok_or_else(|| anyhow!("missing score on line {}", idx + 1))?;
+        let score: i32 = score_raw.parse().with_context(|| {
+            format!(
+                "parsing qrels score '{}' on line {} from {}",
+                score_raw,
+                idx + 1,
+                path.display()
+            )
+        })?;
+
+        qrels
+            .entry(query_id.to_string())
+            .or_default()
+            .push(QrelEntry {
+                doc_id: doc_id.to_string(),
+                score,
+            });
+    }
+
+    Ok(qrels)
+}
+
+fn select_best_doc(entries: &[QrelEntry]) -> Option<&QrelEntry> {
+    entries
+        .iter()
+        .max_by(|a, b| a.score.cmp(&b.score).then_with(|| b.doc_id.cmp(&a.doc_id)))
+}
+
+fn answer_snippet(text: &str) -> Option<String> {
+    let trimmed = text.trim();
+    if trimmed.is_empty() {
+        return None;
+    }
+    let snippet: String = trimmed.chars().take(ANSWER_SNIPPET_CHARS).collect();
+    let snippet = snippet.trim();
+    if snippet.is_empty() {
+        None
+    } else {
+        Some(snippet.to_string())
+    }
+}
+
+fn build_context(title: &str, text: &str) -> String {
+    let title = title.trim();
+    let text = text.trim();
+
+    match (title.is_empty(), text.is_empty()) {
+        (true, true) => String::new(),
+        (true, false) => text.to_string(),
+        (false, true) => title.to_string(),
+        (false, false) => format!("{title}\n\n{text}"),
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use std::fs;
+    use tempfile::tempdir;
+
+    #[test]
+    fn converts_basic_beir_layout() {
+        let dir = tempdir().unwrap();
+        let corpus = r#"
+{"_id":"d1","title":"Doc 1","text":"Doc one has some text for testing."}
+{"_id":"d2","title":"Doc 2","text":"Second document content."}
+"#;
+        let queries = r#"
+{"_id":"q1","text":"What is in doc one?"}
+"#;
+        let qrels = "query-id\tcorpus-id\tscore\nq1\td1\t2\n";
+
+        fs::write(dir.path().join("corpus.jsonl"), corpus.trim()).unwrap();
+        fs::write(dir.path().join("queries.jsonl"), queries.trim()).unwrap();
+        fs::create_dir_all(dir.path().join("qrels")).unwrap();
+        fs::write(dir.path().join("qrels/test.tsv"), qrels).unwrap();
+
+        let paragraphs = convert_beir(dir.path(), DatasetKind::Fever).unwrap();
+
+        assert_eq!(paragraphs.len(), 2);
+        let doc_one = paragraphs
+            .iter()
+            .find(|p| p.id == "fever-d1")
+            .expect("missing paragraph for d1");
+        assert_eq!(doc_one.questions.len(), 1);
+        let question = &doc_one.questions[0];
+        assert_eq!(question.id, "fever-q1");
+        assert!(!question.answers.is_empty());
+        assert!(doc_one.context.contains(&question.answers[0]));
+
+        let doc_two = paragraphs
+            .iter()
+            .find(|p| p.id == "fever-d2")
+            .expect("missing paragraph for d2");
+        assert!(doc_two.questions.is_empty());
+    }
+}

evaluations/src/datasets/mod.rs

@@ -0,0 +1,623 @@
+mod beir;
+mod nq;
+mod squad;
+
+use std::{
+    collections::{BTreeMap, HashMap},
+    fs::{self},
+    path::{Path, PathBuf},
+    str::FromStr,
+};
+
+use anyhow::{anyhow, bail, Context, Result};
+use chrono::{DateTime, TimeZone, Utc};
+use clap::ValueEnum;
+use once_cell::sync::OnceCell;
+use serde::{Deserialize, Serialize};
+use tracing::warn;
+
+const MANIFEST_PATH: &str = concat!(env!("CARGO_MANIFEST_DIR"), "/manifest.yaml");
+static DATASET_CATALOG: OnceCell<DatasetCatalog> = OnceCell::new();
+
+#[derive(Debug, Clone)]
+#[allow(dead_code)]
+pub struct DatasetCatalog {
+    datasets: BTreeMap<String, DatasetEntry>,
+    slices: HashMap<String, SliceLocation>,
+    default_dataset: String,
+}
+
+#[derive(Debug, Clone)]
+#[allow(dead_code)]
+pub struct DatasetEntry {
+    pub metadata: DatasetMetadata,
+    pub raw_path: PathBuf,
+    pub converted_path: PathBuf,
+    pub include_unanswerable: bool,
+    pub slices: Vec<SliceEntry>,
+}
+
+#[derive(Debug, Clone)]
+#[allow(dead_code)]
+pub struct SliceEntry {
+    pub id: String,
+    pub dataset_id: String,
+    pub label: String,
+    pub description: Option<String>,
+    pub limit: Option<usize>,
+    pub corpus_limit: Option<usize>,
+    pub include_unanswerable: Option<bool>,
+    pub seed: Option<u64>,
+}
+
+#[derive(Debug, Clone)]
+#[allow(dead_code)]
+struct SliceLocation {
+    dataset_id: String,
+    slice_index: usize,
+}
+
+#[derive(Debug, Deserialize)]
+struct ManifestFile {
+    default_dataset: Option<String>,
+    datasets: Vec<ManifestDataset>,
+}
+
+#[derive(Debug, Deserialize)]
+struct ManifestDataset {
+    id: String,
+    label: String,
+    category: String,
+    #[serde(default)]
+    entity_suffix: Option<String>,
+    #[serde(default)]
+    source_prefix: Option<String>,
+    raw: String,
+    converted: String,
+    #[serde(default)]
+    include_unanswerable: bool,
+    #[serde(default)]
+    slices: Vec<ManifestSlice>,
+}
+
+#[derive(Debug, Deserialize)]
+struct ManifestSlice {
+    id: String,
+    label: String,
+    #[serde(default)]
+    description: Option<String>,
+    #[serde(default)]
+    limit: Option<usize>,
+    #[serde(default)]
+    corpus_limit: Option<usize>,
+    #[serde(default)]
+    include_unanswerable: Option<bool>,
+    #[serde(default)]
+    seed: Option<u64>,
+}
+
+impl DatasetCatalog {
+    pub fn load() -> Result<Self> {
+        let manifest_raw = fs::read_to_string(MANIFEST_PATH)
+            .with_context(|| format!("reading dataset manifest at {}", MANIFEST_PATH))?;
+        let manifest: ManifestFile = serde_yaml::from_str(&manifest_raw)
+            .with_context(|| format!("parsing dataset manifest at {}", MANIFEST_PATH))?;
+
+        let root = Path::new(env!("CARGO_MANIFEST_DIR"));
+        let mut datasets = BTreeMap::new();
+        let mut slices = HashMap::new();
+
+        for dataset in manifest.datasets {
+            let raw_path = resolve_path(root, &dataset.raw);
+            let converted_path = resolve_path(root, &dataset.converted);
+
+            if !raw_path.exists() {
+                bail!(
+                    "dataset '{}' raw file missing at {}",
+                    dataset.id,
+                    raw_path.display()
+                );
+            }
+            if !converted_path.exists() {
+                warn!(
+                    "dataset '{}' converted file missing at {}; the next conversion run will regenerate it",
+                    dataset.id,
+                    converted_path.display()
+                );
+            }
+
+            let metadata = DatasetMetadata {
+                id: dataset.id.clone(),
+                label: dataset.label.clone(),
+                category: dataset.category.clone(),
+                entity_suffix: dataset
+                    .entity_suffix
+                    .clone()
+                    .unwrap_or_else(|| dataset.label.clone()),
+                source_prefix: dataset
+                    .source_prefix
+                    .clone()
+                    .unwrap_or_else(|| dataset.id.clone()),
+                include_unanswerable: dataset.include_unanswerable,
+                context_token_limit: None,
+            };
+
+            let mut entry_slices = Vec::with_capacity(dataset.slices.len());
+
+            for (index, manifest_slice) in dataset.slices.into_iter().enumerate() {
+                if slices.contains_key(&manifest_slice.id) {
+                    bail!(
+                        "slice '{}' defined multiple times in manifest",
+                        manifest_slice.id
+                    );
+                }
+                entry_slices.push(SliceEntry {
+                    id: manifest_slice.id.clone(),
+                    dataset_id: dataset.id.clone(),
+                    label: manifest_slice.label,
+                    description: manifest_slice.description,
+                    limit: manifest_slice.limit,
+                    corpus_limit: manifest_slice.corpus_limit,
+                    include_unanswerable: manifest_slice.include_unanswerable,
+                    seed: manifest_slice.seed,
+                });
+                slices.insert(
+                    manifest_slice.id,
+                    SliceLocation {
+                        dataset_id: dataset.id.clone(),
+                        slice_index: index,
+                    },
+                );
+            }
+
+            datasets.insert(
+                metadata.id.clone(),
+                DatasetEntry {
+                    metadata,
+                    raw_path,
+                    converted_path,
+                    include_unanswerable: dataset.include_unanswerable,
+                    slices: entry_slices,
+                },
+            );
+        }
+
+        let default_dataset = manifest
+            .default_dataset
+            .or_else(|| datasets.keys().next().cloned())
+            .ok_or_else(|| anyhow!("dataset manifest does not include any datasets"))?;
+
+        Ok(Self {
+            datasets,
+            slices,
+            default_dataset,
+        })
+    }
+
+    pub fn global() -> Result<&'static Self> {
+        DATASET_CATALOG.get_or_try_init(Self::load)
+    }
+
+    pub fn dataset(&self, id: &str) -> Result<&DatasetEntry> {
+        self.datasets
+            .get(id)
+            .ok_or_else(|| anyhow!("unknown dataset '{id}' in manifest"))
+    }
+
+    #[allow(dead_code)]
+    pub fn default_dataset(&self) -> Result<&DatasetEntry> {
+        self.dataset(&self.default_dataset)
+    }
+
+    #[allow(dead_code)]
+    pub fn slice(&self, slice_id: &str) -> Result<(&DatasetEntry, &SliceEntry)> {
+        let location = self
+            .slices
+            .get(slice_id)
+            .ok_or_else(|| anyhow!("unknown slice '{slice_id}' in manifest"))?;
+        let dataset = self
+            .datasets
+            .get(&location.dataset_id)
+            .ok_or_else(|| anyhow!("slice '{slice_id}' references missing dataset"))?;
+        let slice = dataset
+            .slices
+            .get(location.slice_index)
+            .ok_or_else(|| anyhow!("slice index out of bounds for '{slice_id}'"))?;
+        Ok((dataset, slice))
+    }
+}
+
+fn resolve_path(root: &Path, value: &str) -> PathBuf {
+    let path = Path::new(value);
+    if path.is_absolute() {
+        path.to_path_buf()
+    } else {
+        root.join(path)
+    }
+}
+
+pub fn catalog() -> Result<&'static DatasetCatalog> {
+    DatasetCatalog::global()
+}
+
+fn dataset_entry_for_kind(kind: DatasetKind) -> Result<&'static DatasetEntry> {
+    let catalog = catalog()?;
+    catalog.dataset(kind.id())
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq, ValueEnum, Default)]
+pub enum DatasetKind {
+    #[default]
+    SquadV2,
+    NaturalQuestions,
+    Beir,
+    #[value(name = "fever")]
+    Fever,
+    #[value(name = "fiqa")]
+    Fiqa,
+    #[value(name = "hotpotqa", alias = "hotpot-qa")]
+    HotpotQa,
+    #[value(name = "nfcorpus", alias = "nf-corpus")]
+    Nfcorpus,
+    #[value(name = "quora")]
+    Quora,
+    #[value(name = "trec-covid", alias = "treccovid", alias = "trec_covid")]
+    TrecCovid,
+    #[value(name = "scifact")]
+    Scifact,
+    #[value(name = "nq-beir", alias = "natural-questions-beir")]
+    NqBeir,
+}
+
+impl DatasetKind {
+    pub fn id(self) -> &'static str {
+        match self {
+            Self::SquadV2 => "squad-v2",
+            Self::NaturalQuestions => "natural-questions-dev",
+            Self::Beir => "beir",
+            Self::Fever => "fever",
+            Self::Fiqa => "fiqa",
+            Self::HotpotQa => "hotpotqa",
+            Self::Nfcorpus => "nfcorpus",
+            Self::Quora => "quora",
+            Self::TrecCovid => "trec-covid",
+            Self::Scifact => "scifact",
+            Self::NqBeir => "nq-beir",
+        }
+    }
+
+    pub fn label(self) -> &'static str {
+        match self {
+            Self::SquadV2 => "SQuAD v2.0",
+            Self::NaturalQuestions => "Natural Questions (dev)",
+            Self::Beir => "BEIR mix",
+            Self::Fever => "FEVER (BEIR)",
+            Self::Fiqa => "FiQA-2018 (BEIR)",
+            Self::HotpotQa => "HotpotQA (BEIR)",
+            Self::Nfcorpus => "NFCorpus (BEIR)",
+            Self::Quora => "Quora (IR)",
+            Self::TrecCovid => "TREC-COVID (BEIR)",
+            Self::Scifact => "SciFact (BEIR)",
+            Self::NqBeir => "Natural Questions (BEIR)",
+        }
+    }
+
+    pub fn category(self) -> &'static str {
+        match self {
+            Self::SquadV2 => "SQuAD v2.0",
+            Self::NaturalQuestions => "Natural Questions",
+            Self::Beir => "BEIR",
+            Self::Fever => "FEVER",
+            Self::Fiqa => "FiQA-2018",
+            Self::HotpotQa => "HotpotQA",
+            Self::Nfcorpus => "NFCorpus",
+            Self::Quora => "Quora",
+            Self::TrecCovid => "TREC-COVID",
+            Self::Scifact => "SciFact",
+            Self::NqBeir => "Natural Questions",
+        }
+    }
+
+    pub fn entity_suffix(self) -> &'static str {
+        match self {
+            Self::SquadV2 => "SQuAD",
+            Self::NaturalQuestions => "Natural Questions",
+            Self::Beir => "BEIR",
+            Self::Fever => "FEVER",
+            Self::Fiqa => "FiQA",
+            Self::HotpotQa => "HotpotQA",
+            Self::Nfcorpus => "NFCorpus",
+            Self::Quora => "Quora",
+            Self::TrecCovid => "TREC-COVID",
+            Self::Scifact => "SciFact",
+            Self::NqBeir => "Natural Questions",
+        }
+    }
+
+    pub fn source_prefix(self) -> &'static str {
+        match self {
+            Self::SquadV2 => "squad",
+            Self::NaturalQuestions => "nq",
+            Self::Beir => "beir",
+            Self::Fever => "fever",
+            Self::Fiqa => "fiqa",
+            Self::HotpotQa => "hotpotqa",
+            Self::Nfcorpus => "nfcorpus",
+            Self::Quora => "quora",
+            Self::TrecCovid => "trec-covid",
+            Self::Scifact => "scifact",
+            Self::NqBeir => "nq-beir",
+        }
+    }
+
+    pub fn default_raw_path(self) -> PathBuf {
+        dataset_entry_for_kind(self)
+            .map(|entry| entry.raw_path.clone())
+            .unwrap_or_else(|err| panic!("dataset manifest missing entry for {:?}: {err}", self))
+    }
+
+    pub fn default_converted_path(self) -> PathBuf {
+        dataset_entry_for_kind(self)
+            .map(|entry| entry.converted_path.clone())
+            .unwrap_or_else(|err| panic!("dataset manifest missing entry for {:?}: {err}", self))
+    }
+}
+
+impl std::fmt::Display for DatasetKind {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        write!(f, "{}", self.id())
+    }
+}
+
+impl FromStr for DatasetKind {
+    type Err = anyhow::Error;
+
+    fn from_str(s: &str) -> Result<Self, Self::Err> {
+        match s.to_ascii_lowercase().as_str() {
+            "squad" | "squad-v2" | "squad_v2" => Ok(Self::SquadV2),
+            "nq" | "natural-questions" | "natural_questions" | "natural-questions-dev" => {
+                Ok(Self::NaturalQuestions)
+            }
+            "beir" => Ok(Self::Beir),
+            "fever" => Ok(Self::Fever),
+            "fiqa" | "fiqa-2018" => Ok(Self::Fiqa),
+            "hotpotqa" | "hotpot-qa" => Ok(Self::HotpotQa),
+            "nfcorpus" | "nf-corpus" => Ok(Self::Nfcorpus),
+            "quora" => Ok(Self::Quora),
+            "trec-covid" | "treccovid" | "trec_covid" => Ok(Self::TrecCovid),
+            "scifact" => Ok(Self::Scifact),
+            "nq-beir" | "natural-questions-beir" => Ok(Self::NqBeir),
+            other => {
+                anyhow::bail!("unknown dataset '{other}'. Expected one of: squad, natural-questions, beir, fever, fiqa, hotpotqa, nfcorpus, quora, trec-covid, scifact, nq-beir.")
+            }
+        }
+    }
+}
+
+pub const BEIR_DATASETS: [DatasetKind; 8] = [
+    DatasetKind::Fever,
+    DatasetKind::Fiqa,
+    DatasetKind::HotpotQa,
+    DatasetKind::Nfcorpus,
+    DatasetKind::Quora,
+    DatasetKind::TrecCovid,
+    DatasetKind::Scifact,
+    DatasetKind::NqBeir,
+];
+
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct DatasetMetadata {
+    pub id: String,
+    pub label: String,
+    pub category: String,
+    pub entity_suffix: String,
+    pub source_prefix: String,
+    #[serde(default)]
+    pub include_unanswerable: bool,
+    #[serde(default)]
+    pub context_token_limit: Option<usize>,
+}
+
+impl DatasetMetadata {
+    pub fn for_kind(
+        kind: DatasetKind,
+        include_unanswerable: bool,
+        context_token_limit: Option<usize>,
+    ) -> Self {
+        if let Ok(entry) = dataset_entry_for_kind(kind) {
+            return Self {
+                id: entry.metadata.id.clone(),
+                label: entry.metadata.label.clone(),
+                category: entry.metadata.category.clone(),
+                entity_suffix: entry.metadata.entity_suffix.clone(),
+                source_prefix: entry.metadata.source_prefix.clone(),
+                include_unanswerable,
+                context_token_limit,
+            };
+        }
+
+        Self {
+            id: kind.id().to_string(),
+            label: kind.label().to_string(),
+            category: kind.category().to_string(),
+            entity_suffix: kind.entity_suffix().to_string(),
+            source_prefix: kind.source_prefix().to_string(),
+            include_unanswerable,
+            context_token_limit,
+        }
+    }
+}
+
+fn default_metadata() -> DatasetMetadata {
+    DatasetMetadata::for_kind(DatasetKind::default(), false, None)
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct ConvertedDataset {
+    pub generated_at: DateTime<Utc>,
+    #[serde(default = "default_metadata")]
+    pub metadata: DatasetMetadata,
+    pub source: String,
+    pub paragraphs: Vec<ConvertedParagraph>,
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct ConvertedParagraph {
+    pub id: String,
+    pub title: String,
+    pub context: String,
+    pub questions: Vec<ConvertedQuestion>,
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct ConvertedQuestion {
+    pub id: String,
+    pub question: String,
+    pub answers: Vec<String>,
+    pub is_impossible: bool,
+}
+
+pub fn convert(
+    raw_path: &Path,
+    dataset: DatasetKind,
+    include_unanswerable: bool,
+    context_token_limit: Option<usize>,
+) -> Result<ConvertedDataset> {
+    let paragraphs = match dataset {
+        DatasetKind::SquadV2 => squad::convert_squad(raw_path)?,
+        DatasetKind::NaturalQuestions => {
+            nq::convert_nq(raw_path, include_unanswerable, context_token_limit)?
+        }
+        DatasetKind::Beir => convert_beir_mix(include_unanswerable, context_token_limit)?,
+        DatasetKind::Fever
+        | DatasetKind::Fiqa
+        | DatasetKind::HotpotQa
+        | DatasetKind::Nfcorpus
+        | DatasetKind::Quora
+        | DatasetKind::TrecCovid
+        | DatasetKind::Scifact
+        | DatasetKind::NqBeir => beir::convert_beir(raw_path, dataset)?,
+    };
+
+    let metadata_limit = match dataset {
+        DatasetKind::NaturalQuestions => None,
+        _ => context_token_limit,
+    };
+
+    let generated_at = match dataset {
+        DatasetKind::Beir
+        | DatasetKind::Fever
+        | DatasetKind::Fiqa
+        | DatasetKind::HotpotQa
+        | DatasetKind::Nfcorpus
+        | DatasetKind::Quora
+        | DatasetKind::TrecCovid
+        | DatasetKind::Scifact
+        | DatasetKind::NqBeir => base_timestamp(),
+        _ => Utc::now(),
+    };
+
+    let source_label = match dataset {
+        DatasetKind::Beir => "beir-mix".to_string(),
+        _ => raw_path.display().to_string(),
+    };
+
+    Ok(ConvertedDataset {
+        generated_at,
+        metadata: DatasetMetadata::for_kind(dataset, include_unanswerable, metadata_limit),
+        source: source_label,
+        paragraphs,
+    })
+}
+
+fn convert_beir_mix(
+    include_unanswerable: bool,
+    _context_token_limit: Option<usize>,
+) -> Result<Vec<ConvertedParagraph>> {
+    if include_unanswerable {
+        warn!("BEIR mix ignores include_unanswerable flag; all questions are answerable");
+    }
+
+    let mut paragraphs = Vec::new();
+    for subset in BEIR_DATASETS {
+        let entry = dataset_entry_for_kind(subset)?;
+        let subset_paragraphs = beir::convert_beir(&entry.raw_path, subset)?;
+        paragraphs.extend(subset_paragraphs);
+    }
+
+    Ok(paragraphs)
+}
+
+fn ensure_parent(path: &Path) -> Result<()> {
+    if let Some(parent) = path.parent() {
+        fs::create_dir_all(parent)
+            .with_context(|| format!("creating parent directory for {}", path.display()))?;
+    }
+    Ok(())
+}
+
+pub fn write_converted(dataset: &ConvertedDataset, converted_path: &Path) -> Result<()> {
+    ensure_parent(converted_path)?;
+    let json =
+        serde_json::to_string_pretty(dataset).context("serialising converted dataset to JSON")?;
+    fs::write(converted_path, json)
+        .with_context(|| format!("writing converted dataset to {}", converted_path.display()))
+}
+
+pub fn read_converted(converted_path: &Path) -> Result<ConvertedDataset> {
+    let raw = fs::read_to_string(converted_path)
+        .with_context(|| format!("reading converted dataset at {}", converted_path.display()))?;
+    let mut dataset: ConvertedDataset = serde_json::from_str(&raw)
+        .with_context(|| format!("parsing converted dataset at {}", converted_path.display()))?;
+    if dataset.metadata.id.trim().is_empty() {
+        dataset.metadata = default_metadata();
+    }
+    if dataset.source.is_empty() {
+        dataset.source = converted_path.display().to_string();
+    }
+    Ok(dataset)
+}
+
+pub fn ensure_converted(
+    dataset_kind: DatasetKind,
+    raw_path: &Path,
+    converted_path: &Path,
+    force: bool,
+    include_unanswerable: bool,
+    context_token_limit: Option<usize>,
+) -> Result<ConvertedDataset> {
+    if force || !converted_path.exists() {
+        let dataset = convert(
+            raw_path,
+            dataset_kind,
+            include_unanswerable,
+            context_token_limit,
+        )?;
+        write_converted(&dataset, converted_path)?;
+        return Ok(dataset);
+    }
+
+    match read_converted(converted_path) {
+        Ok(dataset)
+            if dataset.metadata.id == dataset_kind.id()
+                && dataset.metadata.include_unanswerable == include_unanswerable
+                && dataset.metadata.context_token_limit == context_token_limit =>
+        {
+            Ok(dataset)
+        }
+        _ => {
+            let dataset = convert(
+                raw_path,
+                dataset_kind,
+                include_unanswerable,
+                context_token_limit,
+            )?;
+            write_converted(&dataset, converted_path)?;
+            Ok(dataset)
+        }
+    }
+}
+
+pub fn base_timestamp() -> DateTime<Utc> {
+    Utc.with_ymd_and_hms(2023, 1, 1, 0, 0, 0).unwrap()
+}

evaluations/src/datasets/nq.rs

@@ -0,0 +1,234 @@
+use std::{
+    collections::BTreeSet,
+    fs::File,
+    io::{BufRead, BufReader},
+    path::Path,
+};
+
+use anyhow::{Context, Result};
+use serde::Deserialize;
+use tracing::warn;
+
+use super::{ConvertedParagraph, ConvertedQuestion};
+
+pub fn convert_nq(
+    raw_path: &Path,
+    include_unanswerable: bool,
+    _context_token_limit: Option<usize>,
+) -> Result<Vec<ConvertedParagraph>> {
+    #[allow(dead_code)]
+    #[derive(Debug, Deserialize)]
+    struct NqExample {
+        question_text: String,
+        document_title: String,
+        example_id: i64,
+        document_tokens: Vec<NqToken>,
+        long_answer_candidates: Vec<NqLongAnswerCandidate>,
+        annotations: Vec<NqAnnotation>,
+    }
+
+    #[derive(Debug, Deserialize)]
+    struct NqToken {
+        token: String,
+        #[serde(default)]
+        html_token: bool,
+    }
+
+    #[allow(dead_code)]
+    #[derive(Debug, Deserialize)]
+    struct NqLongAnswerCandidate {
+        start_token: i32,
+        end_token: i32,
+    }
+
+    #[allow(dead_code)]
+    #[derive(Debug, Deserialize)]
+    struct NqAnnotation {
+        short_answers: Vec<NqShortAnswer>,
+        #[serde(default)]
+        yes_no_answer: String,
+        long_answer: NqLongAnswer,
+    }
+
+    #[derive(Debug, Deserialize)]
+    struct NqShortAnswer {
+        start_token: i32,
+        end_token: i32,
+    }
+
+    #[allow(dead_code)]
+    #[derive(Debug, Deserialize)]
+    struct NqLongAnswer {
+        candidate_index: i32,
+    }
+
+    fn join_tokens(tokens: &[NqToken], start: usize, end: usize) -> String {
+        let mut buffer = String::new();
+        let end = end.min(tokens.len());
+        for token in tokens.iter().skip(start).take(end.saturating_sub(start)) {
+            if token.html_token {
+                continue;
+            }
+            let text = token.token.trim();
+            if text.is_empty() {
+                continue;
+            }
+            let attach = matches!(
+                text,
+                "," | "." | "!" | "?" | ";" | ":" | ")" | "]" | "}" | "%" | "…" | "..."
+            ) || text.starts_with('\'')
+                || text == "n't"
+                || text == "'s"
+                || text == "'re"
+                || text == "'ve"
+                || text == "'d"
+                || text == "'ll";
+
+            if buffer.is_empty() || attach {
+                buffer.push_str(text);
+            } else {
+                buffer.push(' ');
+                buffer.push_str(text);
+            }
+        }
+
+        buffer.trim().to_string()
+    }
+
+    let file = File::open(raw_path).with_context(|| {
+        format!(
+            "opening Natural Questions dataset at {}",
+            raw_path.display()
+        )
+    })?;
+    let reader = BufReader::new(file);
+
+    let mut paragraphs = Vec::new();
+    for (line_idx, line) in reader.lines().enumerate() {
+        let line = line.with_context(|| {
+            format!(
+                "reading Natural Questions line {} from {}",
+                line_idx + 1,
+                raw_path.display()
+            )
+        })?;
+        if line.trim().is_empty() {
+            continue;
+        }
+        let example: NqExample = serde_json::from_str(&line).with_context(|| {
+            format!(
+                "parsing Natural Questions JSON (line {}) at {}",
+                line_idx + 1,
+                raw_path.display()
+            )
+        })?;
+
+        let mut answer_texts: Vec<String> = Vec::new();
+        let mut short_answer_texts: Vec<String> = Vec::new();
+        let mut has_short_or_yesno = false;
+        let mut has_short_answer = false;
+        for annotation in &example.annotations {
+            for short in &annotation.short_answers {
+                if short.start_token < 0 || short.end_token <= short.start_token {
+                    continue;
+                }
+                let start = short.start_token as usize;
+                let end = short.end_token as usize;
+                if start >= example.document_tokens.len() || end > example.document_tokens.len() {
+                    continue;
+                }
+                let text = join_tokens(&example.document_tokens, start, end);
+                if !text.is_empty() {
+                    answer_texts.push(text.clone());
+                    short_answer_texts.push(text);
+                    has_short_or_yesno = true;
+                    has_short_answer = true;
+                }
+            }
+
+            match annotation
+                .yes_no_answer
+                .trim()
+                .to_ascii_lowercase()
+                .as_str()
+            {
+                "yes" => {
+                    answer_texts.push("yes".to_string());
+                    has_short_or_yesno = true;
+                }
+                "no" => {
+                    answer_texts.push("no".to_string());
+                    has_short_or_yesno = true;
+                }
+                _ => {}
+            }
+        }
+
+        let mut answers = dedupe_strings(answer_texts);
+        let is_unanswerable = !has_short_or_yesno || answers.is_empty();
+        if is_unanswerable {
+            if !include_unanswerable {
+                continue;
+            }
+            answers.clear();
+        }
+
+        let paragraph_id = format!("nq-{}", example.example_id);
+        let question_id = format!("nq-{}", example.example_id);
+
+        let context = join_tokens(&example.document_tokens, 0, example.document_tokens.len());
+        if context.is_empty() {
+            continue;
+        }
+
+        if has_short_answer && !short_answer_texts.is_empty() {
+            let normalized_context = context.to_ascii_lowercase();
+            let missing_answer = short_answer_texts.iter().any(|answer| {
+                let needle = answer.trim().to_ascii_lowercase();
+                !needle.is_empty() && !normalized_context.contains(&needle)
+            });
+            if missing_answer {
+                warn!(
+                    question_id = %question_id,
+                    "Skipping Natural Questions example because answers were not found in the assembled context"
+                );
+                continue;
+            }
+        }
+
+        if !include_unanswerable && (!has_short_answer || short_answer_texts.is_empty()) {
+            // yes/no-only questions are excluded by default unless --llm-mode is used
+            continue;
+        }
+
+        let question = ConvertedQuestion {
+            id: question_id,
+            question: example.question_text.trim().to_string(),
+            answers,
+            is_impossible: is_unanswerable,
+        };
+
+        paragraphs.push(ConvertedParagraph {
+            id: paragraph_id,
+            title: example.document_title.trim().to_string(),
+            context,
+            questions: vec![question],
+        });
+    }
+
+    Ok(paragraphs)
+}
+
+fn dedupe_strings<I>(values: I) -> Vec<String>
+where
+    I: IntoIterator<Item = String>,
+{
+    let mut set = BTreeSet::new();
+    for value in values {
+        let trimmed = value.trim();
+        if !trimmed.is_empty() {
+            set.insert(trimmed.to_string());
+        }
+    }
+    set.into_iter().collect()
+}

evaluations/src/datasets/squad.rs

@@ -0,0 +1,107 @@
+use std::{collections::BTreeSet, fs, path::Path};
+
+use anyhow::{Context, Result};
+use serde::Deserialize;
+
+use super::{ConvertedParagraph, ConvertedQuestion};
+
+pub fn convert_squad(raw_path: &Path) -> Result<Vec<ConvertedParagraph>> {
+    #[derive(Debug, Deserialize)]
+    struct SquadDataset {
+        data: Vec<SquadArticle>,
+    }
+
+    #[derive(Debug, Deserialize)]
+    struct SquadArticle {
+        title: String,
+        paragraphs: Vec<SquadParagraph>,
+    }
+
+    #[derive(Debug, Deserialize)]
+    struct SquadParagraph {
+        context: String,
+        qas: Vec<SquadQuestion>,
+    }
+
+    #[derive(Debug, Deserialize)]
+    struct SquadQuestion {
+        id: String,
+        question: String,
+        answers: Vec<SquadAnswer>,
+        #[serde(default)]
+        is_impossible: bool,
+    }
+
+    #[derive(Debug, Deserialize)]
+    struct SquadAnswer {
+        text: String,
+    }
+
+    let raw = fs::read_to_string(raw_path)
+        .with_context(|| format!("reading raw SQuAD dataset at {}", raw_path.display()))?;
+    let parsed: SquadDataset = serde_json::from_str(&raw)
+        .with_context(|| format!("parsing SQuAD dataset at {}", raw_path.display()))?;
+
+    let mut paragraphs = Vec::new();
+    for (article_idx, article) in parsed.data.into_iter().enumerate() {
+        for (paragraph_idx, paragraph) in article.paragraphs.into_iter().enumerate() {
+            let mut questions = Vec::new();
+            for qa in paragraph.qas {
+                let answers = dedupe_strings(qa.answers.into_iter().map(|answer| answer.text));
+                questions.push(ConvertedQuestion {
+                    id: qa.id,
+                    question: qa.question.trim().to_string(),
+                    answers,
+                    is_impossible: qa.is_impossible,
+                });
+            }
+
+            let paragraph_id =
+                format!("{}-{}", slugify(&article.title, article_idx), paragraph_idx);
+
+            paragraphs.push(ConvertedParagraph {
+                id: paragraph_id,
+                title: article.title.trim().to_string(),
+                context: paragraph.context.trim().to_string(),
+                questions,
+            });
+        }
+    }
+
+    Ok(paragraphs)
+}
+
+fn dedupe_strings<I>(values: I) -> Vec<String>
+where
+    I: IntoIterator<Item = String>,
+{
+    let mut set = BTreeSet::new();
+    for value in values {
+        let trimmed = value.trim();
+        if !trimmed.is_empty() {
+            set.insert(trimmed.to_string());
+        }
+    }
+    set.into_iter().collect()
+}
+
+fn slugify(input: &str, fallback_idx: usize) -> String {
+    let mut slug = String::new();
+    let mut last_dash = false;
+    for ch in input.chars() {
+        let c = ch.to_ascii_lowercase();
+        if c.is_ascii_alphanumeric() {
+            slug.push(c);
+            last_dash = false;
+        } else if !last_dash {
+            slug.push('-');
+            last_dash = true;
+        }
+    }
+
+    slug = slug.trim_matches('-').to_string();
+    if slug.is_empty() {
+        slug = format!("article-{fallback_idx}");
+    }
+    slug
+}

evaluations/src/db_helpers.rs

@@ -0,0 +1,110 @@
+use anyhow::{Context, Result};
+use common::storage::{db::SurrealDbClient, indexes::ensure_runtime_indexes};
+use tracing::info;
+
+// Helper functions for index management during namespace reseed
+pub async fn remove_all_indexes(db: &SurrealDbClient) -> Result<()> {
+    let _ = db;
+    info!("Removing ALL indexes before namespace reseed (no-op placeholder)");
+    Ok(())
+}
+
+pub async fn recreate_indexes(db: &SurrealDbClient, dimension: usize) -> Result<()> {
+    info!("Recreating ALL indexes after namespace reseed via shared runtime helper");
+    ensure_runtime_indexes(db, dimension)
+        .await
+        .context("creating runtime indexes")
+}
+
+pub async fn reset_namespace(db: &SurrealDbClient, namespace: &str, database: &str) -> Result<()> {
+    let query = format!(
+        "REMOVE NAMESPACE {ns};
+         DEFINE NAMESPACE {ns};
+         DEFINE DATABASE {db};",
+        ns = namespace,
+        db = database
+    );
+    db.client
+        .query(query)
+        .await
+        .context("resetting SurrealDB namespace")?;
+    db.client
+        .use_ns(namespace)
+        .use_db(database)
+        .await
+        .context("selecting namespace/database after reset")?;
+    Ok(())
+}
+
+// // Test helper to force index dimension change
+// #[allow(dead_code)]
+// pub async fn change_embedding_length_in_hnsw_indexes(
+//     db: &SurrealDbClient,
+//     dimension: usize,
+// ) -> Result<()> {
+//     recreate_indexes(db, dimension).await
+// }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use serde::Deserialize;
+    use uuid::Uuid;
+
+    #[derive(Debug, Deserialize)]
+    struct FooRow {
+        label: String,
+    }
+
+    #[tokio::test]
+    async fn reset_namespace_drops_existing_rows() {
+        let namespace = format!("reset_ns_{}", Uuid::new_v4().simple());
+        let database = format!("reset_db_{}", Uuid::new_v4().simple());
+        let db = SurrealDbClient::memory(&namespace, &database)
+            .await
+            .expect("in-memory db");
+
+        db.client
+            .query(
+                "DEFINE TABLE foo SCHEMALESS;
+                 CREATE foo:foo SET label = 'before';",
+            )
+            .await
+            .expect("seed namespace")
+            .check()
+            .expect("seed response");
+
+        let mut before = db
+            .client
+            .query("SELECT * FROM foo")
+            .await
+            .expect("select before reset");
+        let existing: Vec<FooRow> = before.take(0).expect("rows before reset");
+        assert_eq!(existing.len(), 1);
+        assert_eq!(existing[0].label, "before");
+
+        reset_namespace(&db, &namespace, &database)
+            .await
+            .expect("namespace reset");
+
+        match db.client.query("SELECT * FROM foo").await {
+            Ok(mut response) => {
+                let rows: Vec<FooRow> = response.take(0).unwrap_or_default();
+                assert!(
+                    rows.is_empty(),
+                    "reset namespace should drop rows, found {:?}",
+                    rows
+                );
+            }
+            Err(error) => {
+                let message = error.to_string();
+                assert!(
+                    message.to_ascii_lowercase().contains("table")
+                        || message.to_ascii_lowercase().contains("namespace")
+                        || message.to_ascii_lowercase().contains("foo"),
+                    "unexpected error after namespace reset: {message}"
+                );
+            }
+        }
+    }
+}

evaluations/src/eval.rs

@@ -0,0 +1,128 @@
+//! Evaluation utilities module - re-exports from focused submodules.
+
+// Re-export types from the root types module
+pub use crate::types::*;
+
+// Re-export from focused modules at crate root (crate-internal only)
+pub(crate) use crate::cases::{cases_from_manifest, SeededCase};
+pub(crate) use crate::namespace::{
+    can_reuse_namespace, connect_eval_db, default_database, default_namespace, ensure_eval_user,
+    record_namespace_state,
+};
+pub(crate) use crate::settings::{enforce_system_settings, load_or_init_system_settings};
+
+use std::path::Path;
+
+use anyhow::{Context, Result};
+use common::storage::db::SurrealDbClient;
+use tokio::io::AsyncWriteExt;
+use tracing::info;
+
+use crate::{
+    args::{self, Config},
+    datasets::ConvertedDataset,
+    slice::{self},
+};
+
+/// Grow the slice ledger to contain the target number of cases.
+pub async fn grow_slice(dataset: &ConvertedDataset, config: &Config) -> Result<()> {
+    let ledger_limit = ledger_target(config);
+    let slice_settings = slice::slice_config_with_limit(config, ledger_limit);
+    let slice =
+        slice::resolve_slice(dataset, &slice_settings).context("resolving dataset slice")?;
+    info!(
+        slice = slice.manifest.slice_id.as_str(),
+        cases = slice.manifest.case_count,
+        positives = slice.manifest.positive_paragraphs,
+        negatives = slice.manifest.negative_paragraphs,
+        total_paragraphs = slice.manifest.total_paragraphs,
+        "Slice ledger ready"
+    );
+    println!(
+        "Slice `{}` now contains {} questions ({} positives, {} negatives)",
+        slice.manifest.slice_id,
+        slice.manifest.case_count,
+        slice.manifest.positive_paragraphs,
+        slice.manifest.negative_paragraphs
+    );
+    Ok(())
+}
+
+pub(crate) fn ledger_target(config: &Config) -> Option<usize> {
+    match (config.slice_grow, config.limit) {
+        (Some(grow), Some(limit)) => Some(limit.max(grow)),
+        (Some(grow), None) => Some(grow),
+        (None, limit) => limit,
+    }
+}
+
+pub(crate) async fn write_chunk_diagnostics(path: &Path, cases: &[CaseDiagnostics]) -> Result<()> {
+    args::ensure_parent(path)?;
+    let mut file = tokio::fs::File::create(path)
+        .await
+        .with_context(|| format!("creating diagnostics file {}", path.display()))?;
+    for case in cases {
+        let line = serde_json::to_vec(case).context("serialising chunk diagnostics entry")?;
+        file.write_all(&line).await?;
+        file.write_all(b"\n").await?;
+    }
+    file.flush().await?;
+    Ok(())
+}
+
+pub(crate) async fn warm_hnsw_cache(db: &SurrealDbClient, dimension: usize) -> Result<()> {
+    // Create a dummy embedding for cache warming
+    let dummy_embedding: Vec<f32> = (0..dimension).map(|i| (i as f32).sin()).collect();
+
+    info!("Warming HNSW caches with sample queries");
+
+    // Warm up chunk embedding index - just query the embedding table to load HNSW index
+    let _ = db
+        .client
+        .query(
+            r#"SELECT chunk_id
+               FROM text_chunk_embedding
+               WHERE embedding <|1,1|> $embedding
+               LIMIT 5"#,
+        )
+        .bind(("embedding", dummy_embedding.clone()))
+        .await
+        .context("warming text chunk HNSW cache")?;
+
+    // Warm up entity embedding index
+    let _ = db
+        .client
+        .query(
+            r#"SELECT entity_id
+               FROM knowledge_entity_embedding
+               WHERE embedding <|1,1|> $embedding
+               LIMIT 5"#,
+        )
+        .bind(("embedding", dummy_embedding))
+        .await
+        .context("warming knowledge entity HNSW cache")?;
+
+    info!("HNSW cache warming completed");
+    Ok(())
+}
+
+use chrono::{DateTime, SecondsFormat, Utc};
+
+pub fn format_timestamp(timestamp: &DateTime<Utc>) -> String {
+    timestamp.to_rfc3339_opts(SecondsFormat::Secs, true)
+}
+
+pub(crate) fn sanitize_model_code(code: &str) -> String {
+    code.chars()
+        .map(|ch| {
+            if ch.is_ascii_alphanumeric() {
+                ch.to_ascii_lowercase()
+            } else {
+                '_'
+            }
+        })
+        .collect()
+}
+
+// Re-export run_evaluation from the pipeline module at crate root
+pub use crate::pipeline::run_evaluation;

evaluations/src/inspection.rs

@@ -0,0 +1,184 @@
+use std::{
+    collections::HashMap,
+    fs,
+    path::{Path, PathBuf},
+};
+
+use anyhow::{anyhow, Context, Result};
+use common::storage::{db::SurrealDbClient, types::text_chunk::TextChunk};
+
+use crate::{args::Config, corpus, eval::connect_eval_db, snapshot::DbSnapshotState};
+
+pub async fn inspect_question(config: &Config) -> Result<()> {
+    let question_id = config
+        .inspect_question
+        .as_ref()
+        .ok_or_else(|| anyhow!("--inspect-question is required for inspection mode"))?;
+    let manifest_path = config
+        .inspect_manifest
+        .as_ref()
+        .ok_or_else(|| anyhow!("--inspect-manifest must be provided for inspection mode"))?;
+
+    let manifest = load_manifest(manifest_path)?;
+    let chunk_lookup = build_chunk_lookup(&manifest);
+
+    let question = manifest
+        .questions
+        .iter()
+        .find(|q| q.question_id == *question_id)
+        .ok_or_else(|| {
+            anyhow!(
+                "question '{}' not found in manifest {}",
+                question_id,
+                manifest_path.display()
+            )
+        })?;
+
+    println!("Question: {}", question.question_text);
+    println!("Answers: {:?}", question.answers);
+    println!(
+        "matching_chunk_ids ({}):",
+        question.matching_chunk_ids.len()
+    );
+
+    let mut missing_in_manifest = Vec::new();
+    for chunk_id in &question.matching_chunk_ids {
+        if let Some(entry) = chunk_lookup.get(chunk_id) {
+            println!(
+                "  - {} (paragraph: {})\n    snippet: {}",
+                chunk_id, entry.paragraph_title, entry.snippet
+            );
+        } else {
+            println!("  - {} (missing from manifest)", chunk_id);
+            missing_in_manifest.push(chunk_id.clone());
+        }
+    }
+
+    if missing_in_manifest.is_empty() {
+        println!("All matching_chunk_ids are present in the ingestion manifest");
+    } else {
+        println!(
+            "Missing chunk IDs in manifest {}: {:?}",
+            manifest_path.display(),
+            missing_in_manifest
+        );
+    }
+
+    let db_state_path = config
+        .database
+        .inspect_db_state
+        .clone()
+        .unwrap_or_else(|| default_state_path(config, &manifest));
+    if let Some(state) = load_db_state(&db_state_path)? {
+        if let (Some(ns), Some(db_name)) = (state.namespace.as_deref(), state.database.as_deref()) {
+            match connect_eval_db(config, ns, db_name).await {
+                Ok(db) => match verify_chunks_in_db(&db, &question.matching_chunk_ids).await? {
+                    MissingChunks::None => println!(
+                        "All matching_chunk_ids exist in namespace '{}', database '{}'",
+                        ns, db_name
+                    ),
+                    MissingChunks::Missing(list) => println!(
+                        "Missing chunks in namespace '{}', database '{}': {:?}",
+                        ns, db_name, list
+                    ),
+                },
+                Err(err) => {
+                    println!(
+                        "Failed to connect to SurrealDB namespace '{}' / database '{}': {err}",
+                        ns, db_name
+                    );
+                }
+            }
+        } else {
+            println!(
+                "State file {} is missing namespace/database fields; skipping live DB validation",
+                db_state_path.display()
+            );
+        }
+    } else {
+        println!(
+            "State file {} not found; skipping live DB validation",
+            db_state_path.display()
+        );
+    }
+
+    Ok(())
+}
+
+struct ChunkEntry {
+    paragraph_title: String,
+    snippet: String,
+}
+
+fn load_manifest(path: &Path) -> Result<corpus::CorpusManifest> {
+    let bytes =
+        fs::read(path).with_context(|| format!("reading ingestion manifest {}", path.display()))?;
+    serde_json::from_slice(&bytes)
+        .with_context(|| format!("parsing ingestion manifest {}", path.display()))
+}
+
+fn build_chunk_lookup(manifest: &corpus::CorpusManifest) -> HashMap<String, ChunkEntry> {
+    let mut lookup = HashMap::new();
+    for paragraph in &manifest.paragraphs {
+        for chunk in &paragraph.chunks {
+            let snippet = chunk
+                .chunk
+                .chunk
+                .chars()
+                .take(160)
+                .collect::<String>()
+                .replace('\n', " ");
+            lookup.insert(
+                chunk.chunk.id.clone(),
+                ChunkEntry {
+                    paragraph_title: paragraph.title.clone(),
+                    snippet,
+                },
+            );
+        }
+    }
+    lookup
+}
+
+fn default_state_path(config: &Config, manifest: &corpus::CorpusManifest) -> PathBuf {
+    config
+        .cache_dir
+        .join("snapshots")
+        .join(&manifest.metadata.dataset_id)
+        .join(&manifest.metadata.slice_id)
+        .join("db/state.json")
+}
+
+fn load_db_state(path: &Path) -> Result<Option<DbSnapshotState>> {
+    if !path.exists() {
+        return Ok(None);
+    }
+    let bytes = fs::read(path).with_context(|| format!("reading db state {}", path.display()))?;
+    let state = serde_json::from_slice(&bytes)
+        .with_context(|| format!("parsing db state {}", path.display()))?;
+    Ok(Some(state))
+}
+
+enum MissingChunks {
+    None,
+    Missing(Vec<String>),
+}
+
+async fn verify_chunks_in_db(db: &SurrealDbClient, chunk_ids: &[String]) -> Result<MissingChunks> {
+    let mut missing = Vec::new();
+    for chunk_id in chunk_ids {
+        let exists = db
+            .get_item::<TextChunk>(chunk_id)
+            .await
+            .with_context(|| format!("fetching text_chunk {}", chunk_id))?
+            .is_some();
+        if !exists {
+            missing.push(chunk_id.clone());
+        }
+    }
+    if missing.is_empty() {
+        Ok(MissingChunks::None)
+    } else {
+        Ok(MissingChunks::Missing(missing))
+    }
+}

evaluations/src/main.rs

@@ -0,0 +1,247 @@
+mod args;
+mod cache;
+mod cases;
+mod corpus;
+mod datasets;
+mod db_helpers;
+mod eval;
+mod inspection;
+mod namespace;
+mod openai;
+mod perf;
+mod pipeline;
+mod report;
+mod settings;
+mod slice;
+mod snapshot;
+mod types;
+
+use anyhow::Context;
+use tokio::runtime::Builder;
+use tracing::info;
+use tracing_subscriber::{fmt, EnvFilter};
+
+/// Configure SurrealDB environment variables for optimal performance
+fn configure_surrealdb_performance(cpu_count: usize) {
+    // Set environment variables only if they're not already set
+    let indexing_batch_size = std::env::var("SURREAL_INDEXING_BATCH_SIZE")
+        .unwrap_or_else(|_| (cpu_count * 2).to_string());
+    std::env::set_var("SURREAL_INDEXING_BATCH_SIZE", indexing_batch_size);
+
+    let max_order_queue = std::env::var("SURREAL_MAX_ORDER_LIMIT_PRIORITY_QUEUE_SIZE")
+        .unwrap_or_else(|_| (cpu_count * 4).to_string());
+    std::env::set_var(
+        "SURREAL_MAX_ORDER_LIMIT_PRIORITY_QUEUE_SIZE",
+        max_order_queue,
+    );
+
+    let websocket_concurrent = std::env::var("SURREAL_WEBSOCKET_MAX_CONCURRENT_REQUESTS")
+        .unwrap_or_else(|_| cpu_count.to_string());
+    std::env::set_var(
+        "SURREAL_WEBSOCKET_MAX_CONCURRENT_REQUESTS",
+        websocket_concurrent,
+    );
+
+    let websocket_buffer = std::env::var("SURREAL_WEBSOCKET_RESPONSE_BUFFER_SIZE")
+        .unwrap_or_else(|_| (cpu_count * 8).to_string());
+    std::env::set_var("SURREAL_WEBSOCKET_RESPONSE_BUFFER_SIZE", websocket_buffer);
+
+    let transaction_cache = std::env::var("SURREAL_TRANSACTION_CACHE_SIZE")
+        .unwrap_or_else(|_| (cpu_count * 16).to_string());
+    std::env::set_var("SURREAL_TRANSACTION_CACHE_SIZE", transaction_cache);
+
+    info!(
+        indexing_batch_size = %std::env::var("SURREAL_INDEXING_BATCH_SIZE").unwrap(),
+        max_order_queue = %std::env::var("SURREAL_MAX_ORDER_LIMIT_PRIORITY_QUEUE_SIZE").unwrap(),
+        websocket_concurrent = %std::env::var("SURREAL_WEBSOCKET_MAX_CONCURRENT_REQUESTS").unwrap(),
+        websocket_buffer = %std::env::var("SURREAL_WEBSOCKET_RESPONSE_BUFFER_SIZE").unwrap(),
+        transaction_cache = %std::env::var("SURREAL_TRANSACTION_CACHE_SIZE").unwrap(),
+        "Configured SurrealDB performance variables"
+    );
+}
+
+fn main() -> anyhow::Result<()> {
+    // Create an explicit multi-threaded runtime with optimized configuration
+    let runtime = Builder::new_multi_thread()
+        .enable_all()
+        .worker_threads(std::thread::available_parallelism()?.get())
+        .max_blocking_threads(std::thread::available_parallelism()?.get())
+        .thread_stack_size(10 * 1024 * 1024) // 10MiB stack size
+        .thread_name("eval-retrieval-worker")
+        .build()
+        .context("failed to create tokio runtime")?;
+
+    runtime.block_on(async_main())
+}
+
+async fn async_main() -> anyhow::Result<()> {
+    // Log runtime configuration
+    let cpu_count = std::thread::available_parallelism()?.get();
+    info!(
+        cpu_cores = cpu_count,
+        worker_threads = cpu_count,
+        blocking_threads = cpu_count,
+        thread_stack_size = "10MiB",
+        "Started multi-threaded tokio runtime"
+    );
+
+    // Configure SurrealDB environment variables for better performance
+    configure_surrealdb_performance(cpu_count);
+
+    let filter = std::env::var("RUST_LOG").unwrap_or_else(|_| "info".to_string());
+    let _ = fmt()
+        .with_env_filter(EnvFilter::try_new(&filter).unwrap_or_else(|_| EnvFilter::new("info")))
+        .try_init();
+
+    let parsed = args::parse()?;
+
+    // Clap handles help automatically, so we don't need to check for it manually
+
+    if parsed.config.inspect_question.is_some() {
+        inspection::inspect_question(&parsed.config).await?;
+        return Ok(());
+    }
+
+    let dataset_kind = parsed.config.dataset;
+
+    if parsed.config.convert_only {
+        info!(
+            dataset = dataset_kind.id(),
+            "Starting dataset conversion only run"
+        );
+        let dataset = crate::datasets::convert(
+            parsed.config.raw_dataset_path.as_path(),
+            dataset_kind,
+            parsed.config.llm_mode,
+            parsed.config.context_token_limit(),
+        )
+        .with_context(|| {
+            format!(
+                "converting {} dataset at {}",
+                dataset_kind.label(),
+                parsed.config.raw_dataset_path.display()
+            )
+        })?;
+        crate::datasets::write_converted(&dataset, parsed.config.converted_dataset_path.as_path())
+            .with_context(|| {
+                format!(
+                    "writing converted dataset to {}",
+                    parsed.config.converted_dataset_path.display()
+                )
+            })?;
+        println!(
+            "Converted dataset written to {}",
+            parsed.config.converted_dataset_path.display()
+        );
+        return Ok(());
+    }
+
+    info!(dataset = dataset_kind.id(), "Preparing converted dataset");
+    let dataset = crate::datasets::ensure_converted(
+        dataset_kind,
+        parsed.config.raw_dataset_path.as_path(),
+        parsed.config.converted_dataset_path.as_path(),
+        parsed.config.force_convert,
+        parsed.config.llm_mode,
+        parsed.config.context_token_limit(),
+    )
+    .with_context(|| {
+        format!(
+            "preparing converted dataset at {}",
+            parsed.config.converted_dataset_path.display()
+        )
+    })?;
+
+    info!(
+        questions = dataset
+            .paragraphs
+            .iter()
+            .map(|p| p.questions.len())
+            .sum::<usize>(),
+        paragraphs = dataset.paragraphs.len(),
+        dataset = dataset.metadata.id.as_str(),
+        "Dataset ready"
+    );
+
+    if parsed.config.slice_grow.is_some() {
+        eval::grow_slice(&dataset, &parsed.config)
+            .await
+            .context("growing slice ledger")?;
+        return Ok(());
+    }
+
+    info!("Running retrieval evaluation");
+    let summary = eval::run_evaluation(&dataset, &parsed.config)
+        .await
+        .context("running retrieval evaluation")?;
+
+    let report = report::write_reports(
+        &summary,
+        parsed.config.report_dir.as_path(),
+        parsed.config.summary_sample,
+    )
+    .with_context(|| format!("writing reports to {}", parsed.config.report_dir.display()))?;
+    let perf_mirrors = perf::mirror_perf_outputs(
+        &report.record,
+        &summary,
+        parsed.config.report_dir.as_path(),
+        parsed.config.perf_log_json.as_deref(),
+        parsed.config.perf_log_dir.as_deref(),
+    )
+    .with_context(|| {
+        format!(
+            "writing perf mirrors under {}",
+            parsed.config.report_dir.display()
+        )
+    })?;
+
+    let perf_note = if perf_mirrors.is_empty() {
+        String::new()
+    } else {
+        format!(
+            " | Perf mirrors: {}",
+            perf_mirrors
+                .iter()
+                .map(|path| path.display().to_string())
+                .collect::<Vec<_>>()
+                .join(", ")
+        )
+    };
+
+    if summary.llm_cases > 0 {
+        println!(
+            "[{}] Retrieval Precision@{k}: {precision:.3} ({correct}/{retrieval_total}) + LLM: {llm_answered}/{llm_total} ({llm_precision:.3}) → JSON: {json} | Markdown: {md} | History: {history}{perf_note}",
+            summary.dataset_label,
+            k = summary.k,
+            precision = summary.precision,
+            correct = summary.correct,
+            retrieval_total = summary.retrieval_cases,
+            llm_answered = summary.llm_answered,
+            llm_total = summary.llm_cases,
+            llm_precision = summary.llm_precision,
+            json = report.paths.json.display(),
+            md = report.paths.markdown.display(),
+            history = report.history_path.display(),
+            perf_note = perf_note,
+        );
+    } else {
+        println!(
+            "[{}] Retrieval Precision@{k}: {precision:.3} ({correct}/{retrieval_total}) → JSON: {json} | Markdown: {md} | History: {history}{perf_note}",
+            summary.dataset_label,
+            k = summary.k,
+            precision = summary.precision,
+            correct = summary.correct,
+            retrieval_total = summary.retrieval_cases,
+            json = report.paths.json.display(),
+            md = report.paths.markdown.display(),
+            history = report.history_path.display(),
+            perf_note = perf_note,
+        );
+    }
+
+    if parsed.config.perf_log_console {
+        perf::print_console_summary(&report.record);
+    }
+
+    Ok(())
+}

evaluations/src/namespace.rs

@@ -0,0 +1,230 @@
+//! Database namespace management utilities.
+
+use anyhow::{anyhow, Context, Result};
+use chrono::Utc;
+use common::storage::{
+    db::SurrealDbClient,
+    types::user::{Theme, User},
+    types::StoredObject,
+};
+use serde::Deserialize;
+use tracing::{info, warn};
+
+use crate::{
+    args::Config,
+    datasets,
+    snapshot::{self, DbSnapshotState},
+};
+
+/// Connect to the evaluation database with fallback auth strategies.
+pub(crate) async fn connect_eval_db(
+    config: &Config,
+    namespace: &str,
+    database: &str,
+) -> Result<SurrealDbClient> {
+    match SurrealDbClient::new(
+        &config.database.db_endpoint,
+        &config.database.db_username,
+        &config.database.db_password,
+        namespace,
+        database,
+    )
+    .await
+    {
+        Ok(client) => {
+            info!(
+                endpoint = %config.database.db_endpoint,
+                namespace,
+                database,
+                auth = "root",
+                "Connected to SurrealDB"
+            );
+            Ok(client)
+        }
+        Err(root_err) => {
+            info!(
+                endpoint = %config.database.db_endpoint,
+                namespace,
+                database,
+                "Root authentication failed; trying namespace-level auth"
+            );
+            let namespace_client = SurrealDbClient::new_with_namespace_user(
+                &config.database.db_endpoint,
+                namespace,
+                &config.database.db_username,
+                &config.database.db_password,
+                database,
+            )
+            .await
+            .map_err(|ns_err| {
+                anyhow!(
+                    "failed to connect to SurrealDB via root ({root_err}) or namespace ({ns_err}) credentials"
+                )
+            })?;
+            info!(
+                endpoint = %config.database.db_endpoint,
+                namespace,
+                database,
+                auth = "namespace",
+                "Connected to SurrealDB"
+            );
+            Ok(namespace_client)
+        }
+    }
+}
+
+/// Check if the namespace contains any corpus data.
+pub(crate) async fn namespace_has_corpus(db: &SurrealDbClient) -> Result<bool> {
+    #[derive(Deserialize)]
+    struct CountRow {
+        count: i64,
+    }
+
+    let mut response = db
+        .client
+        .query("SELECT count() AS count FROM text_chunk")
+        .await
+        .context("checking namespace corpus state")?;
+    let rows: Vec<CountRow> = response.take(0).unwrap_or_default();
+    Ok(rows.first().map(|row| row.count).unwrap_or(0) > 0)
+}
+
+/// Determine if we can reuse an existing namespace based on cached state.
+#[allow(clippy::too_many_arguments)]
+pub(crate) async fn can_reuse_namespace(
+    db: &SurrealDbClient,
+    descriptor: &snapshot::Descriptor,
+    namespace: &str,
+    database: &str,
+    dataset_id: &str,
+    slice_id: &str,
+    ingestion_fingerprint: &str,
+    slice_case_count: usize,
+) -> Result<bool> {
+    let state = match descriptor.load_db_state().await? {
+        Some(state) => state,
+        None => {
+            info!("No namespace state recorded; reseeding corpus from cached shards");
+            return Ok(false);
+        }
+    };
+
+    if state.slice_case_count != slice_case_count {
+        info!(
+            requested_cases = slice_case_count,
+            stored_cases = state.slice_case_count,
+            "Skipping live namespace reuse; cached state does not match requested window"
+        );
+        return Ok(false);
+    }
+
+    if state.dataset_id != dataset_id
+        || state.slice_id != slice_id
+        || state.ingestion_fingerprint != ingestion_fingerprint
+        || state.namespace.as_deref() != Some(namespace)
+        || state.database.as_deref() != Some(database)
+    {
+        info!(
+            namespace,
+            database, "Cached namespace metadata mismatch; rebuilding corpus from ingestion cache"
+        );
+        return Ok(false);
+    }
+
+    if namespace_has_corpus(db).await? {
+        Ok(true)
+    } else {
+        info!(
+            namespace,
+            database,
+            "Namespace metadata matches but tables are empty; reseeding from ingestion cache"
+        );
+        Ok(false)
+    }
+}
+
+/// Record the current namespace state to allow future reuse checks.
+pub(crate) async fn record_namespace_state(
+    descriptor: &snapshot::Descriptor,
+    dataset_id: &str,
+    slice_id: &str,
+    ingestion_fingerprint: &str,
+    namespace: &str,
+    database: &str,
+    slice_case_count: usize,
+) {
+    let state = DbSnapshotState {
+        dataset_id: dataset_id.to_string(),
+        slice_id: slice_id.to_string(),
+        ingestion_fingerprint: ingestion_fingerprint.to_string(),
+        snapshot_hash: descriptor.metadata_hash().to_string(),
+        updated_at: Utc::now(),
+        namespace: Some(namespace.to_string()),
+        database: Some(database.to_string()),
+        slice_case_count,
+    };
+    if let Err(err) = descriptor.store_db_state(&state).await {
+        warn!(error = %err, "Failed to record namespace state");
+    }
+}
+
+fn sanitize_identifier(input: &str) -> String {
+    let mut cleaned: String = input
+        .chars()
+        .map(|ch| {
+            if ch.is_ascii_alphanumeric() {
+                ch.to_ascii_lowercase()
+            } else {
+                '_'
+            }
+        })
+        .collect();
+    if cleaned.is_empty() {
+        cleaned.push('x');
+    }
+    if cleaned.len() > 64 {
+        cleaned.truncate(64);
+    }
+    cleaned
+}
+
+/// Generate a default namespace name based on dataset and limit.
+pub(crate) fn default_namespace(dataset_id: &str, limit: Option<usize>) -> String {
+    let dataset_component = sanitize_identifier(dataset_id);
+    let limit_component = match limit {
+        Some(value) if value > 0 => format!("limit{}", value),
+        _ => "all".to_string(),
+    };
+    format!("eval_{}_{}", dataset_component, limit_component)
+}
+
+/// Generate the default database name for evaluations.
+pub(crate) fn default_database() -> String {
+    "retrieval_eval".to_string()
+}
+
+/// Ensure the evaluation user exists in the database.
+pub(crate) async fn ensure_eval_user(db: &SurrealDbClient) -> Result<User> {
+    let timestamp = datasets::base_timestamp();
+    let user = User {
+        id: "eval-user".to_string(),
+        created_at: timestamp,
+        updated_at: timestamp,
+        email: "eval-retrieval@minne.dev".to_string(),
+        password: "not-used".to_string(),
+        anonymous: false,
+        api_key: None,
+        admin: false,
+        timezone: "UTC".to_string(),
+        theme: Theme::System,
+    };
+
+    if let Some(existing) = db.get_item::<User>(user.get_id()).await? {
+        return Ok(existing);
+    }
+
+    db.store_item(user.clone())
+        .await
+        .context("storing evaluation user")?;
+    Ok(user)
+}