Hybrid ranking

Expand diagram

Hybrid retrieval and ranking

Dense and BM25 rankings feed weighted fusion and optional reranking. Graph retrieval is a separate path; its matches count as evidence only after resolving to source text.

Use the zoom controls, drag to pan, or pinch on a touch screen. Press Escape to close.

Weighted hybridCopy Anchor Link

The weighted hybrid retriever asks each active member for up to 30 candidates. Each member's scores are min-max normalized within that query, then the fitted member weights are applied to the combined candidate pool.

The current fitted weights are dense 0.60 and BM25 0.40. TF-IDF and Word2Vec have zero weight in this configuration. A fused score is query-local and should not be compared as a confidence score across different questions.

Cross-encoder rerankingCopy Anchor Link

hybrid_rr sends up to RERANK_TOPN candidates (default 20) to bge-reranker-v2-m3. The cross-encoder evaluates the query and passage together and returns a revised order.

If the reranker is unavailable or fails, the system returns the weighted fusion order instead of failing the request. Set RERANK_ENABLED=0 to opt out.

For the measured benchmark results and their limits, see Evaluation and benchmarks.