benchmarks: ready for hybrid revised

retrieval-pipeline/src/pipeline/config.rs

@@ -63,6 +63,7 @@ pub struct RetrievalTuning {
     pub rerank_blend_weight: f32,
     pub rerank_scores_only: bool,
     pub rerank_keep_top: usize,
+    pub chunk_result_cap: usize,
 }
 
 impl Default for RetrievalTuning {
@@ -86,6 +87,7 @@ impl Default for RetrievalTuning {
             rerank_blend_weight: 0.65,
             rerank_scores_only: false,
             rerank_keep_top: 8,
+            chunk_result_cap: 5,
         }
     }
 }

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

@@ -675,7 +675,13 @@ pub fn assemble_chunks(ctx: &mut PipelineContext<'_>) -> Result<(), AppError> {
         ctx.config.tuning.lexical_match_weight,
     );
 
-    let limit = ctx.config.tuning.chunk_vector_take.max(1);
+    // Limit how many chunks we return to keep context size reasonable.
+    let limit = ctx
+        .config
+        .tuning
+        .chunk_result_cap
+        .max(1)
+        .min(ctx.config.tuning.chunk_vector_take.max(1));
     if chunk_values.len() > limit {
         chunk_values.truncate(limit);
     }

retrieval-pipeline/src/reranking/mod.rs

@@ -29,7 +29,10 @@ impl RerankerPool {
     /// Build the pool at startup.
     /// `pool_size` controls max parallel reranks.
     pub fn new(pool_size: usize) -> Result<Arc<Self>, AppError> {
-        Self::new_with_options(pool_size, RerankInitOptions::default())
+        Self::new_with_options(
+            pool_size,
+            RerankInitOptions::new(fastembed::RerankerModel::JINARerankerV1TurboEn),
+        )
     }
 
     fn new_with_options(