Two retrievers, different strengths
Keyword search, such as BM25, excels at exact terms, rare names, and codes. Vector search excels at meaning and paraphrase. Each fails where the other shines, so blending them, called hybrid search, often beats either alone.
How fusion works
You run both retrievers, then merge their ranked lists into one. The merge cannot simply add raw scores, since keyword and vector scores live on different scales.
Reciprocal rank fusion
A robust method is reciprocal rank fusion. Instead of using scores, it uses each item rank in each list, giving more weight to items that rank high. Because it ignores raw scores, it sidesteps the scale mismatch entirely.
When hybrid wins
- A query has a rare keyword that the embedding blurs away.
- A query is paraphrased so keywords miss but meaning matches.
Hybrid covers both cases, which is why it is a strong default for retrieval.
Key idea
Hybrid search runs keyword and vector retrievers together and fuses their rankings, often with reciprocal rank fusion, so each method covers the other blind spots.