Summary
Hybrid search combines vector similarity with keyword search (BM25). The two find different things: vectors capture semantic meaning; BM25 captures exact terms, product codes, unusual proper nouns. Fusion of both consistently outperforms either alone.
Objectives
- 01State what BM25 does that vector search does not.
- 02Describe reciprocal rank fusion (RRF).
- 03Explain when to weight one signal higher than the other.
The Lesson
The gap
Vector search misses when the query contains a rare exact term (a SKU, a person's name, a legal citation). BM25 excels at exact term matching but misses paraphrases. Hybrid = both, fused.
Fusion
Reciprocal Rank Fusion: for each candidate, sum 1/(rank_in_vector + k) + 1/(rank_in_bm25 + k). Simple, hyperparameter-light, robust. Alternatives (learned rerankers, weighted linear combinations) exist but RRF is the sensible default.
Key Ideas
- BM25 + vectors beats either alone.
- Reciprocal Rank Fusion is the default fusion method.