What is hybrid search in RAG?

Hybrid search combines two retrieval methods that fail in opposite ways: dense vector search, which matches meaning, and lexical keyword search like BM25, which matches exact words. You run both against the query and fuse the two ranked lists — often with reciprocal rank fusion — into one result set. You get semantic understanding and exact-term precision instead of having to choose between them.

Why does hybrid beat either method alone?

Vector search is great at paraphrase — ask about a car and it happily finds automobile — but it's oddly weak at exact tokens: product SKUs, error codes, function names, acronyms, people's names. It will return something semantically close and wrong. Keyword search is the mirror image: it nails ERR_1042 or GDPR Article 17 exactly, but whiffs the moment the user phrases the idea differently from the document.

Real questions need both at once, which is why hybrid is the default in most serious RAG stacks.

How are the two result lists combined?

You don't average the raw scores — a cosine similarity and a BM25 score aren't on the same scale, so adding them directly is meaningless. The common trick is reciprocal rank fusion: each result earns points based on its rank position in each list, and you sum those points. A chunk that ranks decently in both lists beats one that ranks first in only one.

It's simple, needs no tuning, and doesn't care that the two systems score on different scales.

Dense vs sparse vs hybrid

The cost is modest — you maintain two indexes and add one fusion step — but for anything with a mix of natural-language questions and exact-term lookups, it's almost always worth it. If your users search by both concepts and specific identifiers, hybrid is the safe default.