Glossary
Hybrid search
Hybrid search runs keyword search and vector search on the same query and merges the two result lists into one ranking. Keyword matching catches exact strings such as product codes, names and clause numbers, while vector matching catches paraphrase, so retrieval finds both kinds of evidence.
Each lane misses what the other finds
Semantic search knows that "staff leaving" and "employee attrition" describe the same thing. It cannot reliably find error code E-4012, because embeddings blur exact strings and a code one digit away lands in nearly the same place. Keyword search finds the code at once and has no idea the two phrases share a meaning. Enterprise content such as contracts, tickets and part lists draws both kinds of query.
How two result lists become one
The keyword lane ranks passages by how well their words match the query terms. The vector lane returns the passages nearest the query in meaning, usually from a vector database. The two lanes score on different scales, so their scores cannot simply be added. Common practice merges on rank position instead: a passage near the top of both lists outranks one that tops only one. Some systems then pass the merged candidates to a reranker, a model that reads the question and each candidate together and makes the final cut.
Two indexes, one definition of what exists
The two indexes are separate stores that have to agree. A document deleted from only one keeps appearing through the other lane. A permission filter applied to the keyword lane and missed on the vector lane shows a person a passage they may not read, and the merged list hides which lane it came from. The fix is plumbing: updates and access rules flow through one ingestion pipeline and reach both indexes together.
Where does hybrid search sit in RAG?
It is the retrieval step in most production retrieval-augmented generation. Its inputs are the passages produced by document chunking, and its output is the short list of evidence handed to the model. Missed exact codes and stale passages are frequent causes of RAG hallucination.
Frequently asked questions
Is hybrid search the same as vector search with filters?
No. Filters narrow candidates by attributes such as date or department. Hybrid search adds a second, independent way of finding candidates, based on exact terms.
Is a reranker required for hybrid search?
No. Merging the two lists gives a usable ranking. A reranker refines the final few positions at extra cost per query.