Learning path · Retrieval & Ranking · 44
Hybrid Search
Combining dense vector retrieval with sparse lexical methods (BM25) for robust recall across paraphrase and keyword queries.
Why it matters
- Fixes semantic search blind spots on SKUs, error codes, and names.
- Industry default for production RAG indexes.
- Fusion weight tuning belongs in eval harnesses.
Key ideas
- BM25 plus vectors
- Reciprocal rank fusion
- Weighted blending
Hybrid search runs lexical and semantic retrievers in parallel, then fuses rankings—RRF or learned weights. Lexical wins on exact tokens; semantic wins on conceptual questions. Calibrate on real query logs, not only synthetic paraphrases. Expose fusion parameters to observability so on-call engineers can diagnose sudden recall drops after index rebuilds. A/B test fusion weights on live traffic with guardrail metrics; offline Recall@K alone misses user preference for exact matches. Ship only after eval gates pass on representative production failures.
Updated 2026-08-09 · Full learning path