What is hybrid search?
A pure vector search can return the wrong policy with total confidence, because it never saw that the policy number was the whole question.
Hybrid search
Hybrid retrieval
Hybrid search runs a keyword search and a vector search on the same query, then combines the two result sets into one ranking. Keyword matching finds exact terms; vector matching finds passages with related meaning, even when the wording differs.
The two methods fail in opposite ways. Keyword search, usually an algorithm called BM25, matches the exact words in a query. It finds a part number or a policy code every time, but misses a question phrased in different words.
Vector search matches meaning instead of words. It compares embeddings, which are numbers that stand in for text, so a paraphrased question can still find the right passage. What it loses is precision on exact strings: model numbers, order IDs, clause references.
Hybrid search runs both, then merges the two ranked lists into one. The merge step usually uses an algorithm called reciprocal rank fusion. A result that scores well on either method rises. One that scores well on both rises further.
Vector-only search is a bet you don't need to make
Most retrieval-augmented generation projects start with vector search alone, because it is the part that feels new. Then a support query arrives with an order number in it. The embedding blurs the digits into something close but wrong, and the assistant answers about the wrong order.
This is not a rare edge case. Product catalogs, contracts and medical records are full of identifiers that carry the meaning: a clause number, a drug code, an account reference. Embeddings represent these poorly. Change one digit in a code and the embedding barely moves.
We set hybrid search as the default on retrieval systems we build, rather than adding it after launch. Retrofitting means re-indexing the corpus and rewriting the retrieval and ranking code, not flipping a setting.
- 01Keyword search alone misses paraphrased questions and synonyms.
- 02Vector search alone misses exact identifiers, codes and short names.
- 03The two failure modes rarely show up in the same test query, so a demo can hide the gap.
- QueryThe user's question, unchanged.
- KeywordBM25 matches exact terms.
- VectorEmbeddings match meaning.
- FuseRank fusion merges both lists.
- AnswerModel sees the combined top results.
Both branches run on every query. Neither is a fallback for the other.
Common questions
01Is hybrid search slower than vector search alone?
It adds a second query and a merge step, so latency is higher than vector search alone. The two searches usually run in parallel, and the merge itself is cheap. Total response time rises only slightly, not by double.
02Do I need a special database for hybrid search?
Most vector databases now support a keyword index and a vector index on the same data, so a separate system is rarely needed. Check whether your database offers built-in rank fusion, or whether you have to merge the two result sets yourself.
03How do you weight keyword results against vector results?
Reciprocal rank fusion is the common default, and it needs no manual weighting. Each result's score comes from its rank position in each list, not a tunable blend. Start from that default and measure before adding your own weights.
04What is BM25 and do I need to know it to use hybrid search?
BM25 is the standard keyword-ranking algorithm behind the keyword half of hybrid search, and most databases implement it for you. You do not need to implement it yourself, only decide that your retrieval pipeline includes both a keyword index and a vector index.
