feat: pool fastembed, batch embeddings, and reconcile embedding config on startup

evaluations/src/pipeline/stages/prepare_db.rs

@@ -11,7 +11,7 @@ use crate::{
     },
     openai,
 };
-use common::utils::embedding::EmbeddingProvider;
+use common::utils::embedding::{default_embedding_pool_size, EmbeddingProvider};
 
 use super::super::{
     context::{EvalStage, EvaluationContext},
@@ -43,9 +43,12 @@ pub(crate) async fn prepare_db(
     // Create embedding provider directly from config (eval only supports FastEmbed and Hashed)
     let embedding_provider = match config.embedding_backend {
         crate::args::EmbeddingBackend::FastEmbed => {
-            EmbeddingProvider::new_fastembed(config.embedding_model.clone())
-                .await
-                .context("creating FastEmbed provider")?
+            EmbeddingProvider::new_fastembed(
+                config.embedding_model.clone(),
+                default_embedding_pool_size(),
+            )
+            .await
+            .context("creating FastEmbed provider")?
         }
         crate::args::EmbeddingBackend::Hashed => {
             EmbeddingProvider::new_hashed(1536).context("creating Hashed provider")?

evaluations/src/pipeline/stages/run_queries.rs

@@ -136,12 +136,10 @@ pub(crate) async fn run_queries(
     let embedding_provider_for_queries = ctx.embedding_provider()?.clone();
     let rerank_pool_for_queries = rerank_pool.clone();
     let db = ctx.db()?.clone();
-    let openai_client = ctx.openai_client()?;
 
     let raw_results = stream::iter(cases_iter)
         .map(move |(idx, case)| {
             let db = db.clone();
-            let openai_client = Arc::clone(&openai_client);
             let user_id = user_id.clone();
             let retrieval_config = Arc::clone(&retrieval_config);
             let embedding_provider = embedding_provider_for_queries.clone();
@@ -180,8 +178,7 @@ pub(crate) async fn run_queries(
 
                 let params = pipeline::RetrievalParams {
                     db_client: &db,
-                    openai_client: &openai_client,
-                    embedding_provider: Some(&embedding_provider),
+                    embedding_provider: &embedding_provider,
                     input_text: &question,
                     user_id: &user_id,
                     config: (*retrieval_config).clone(),