Hybrid Search

Hybrid search combines dense (embedding-based) retrieval and sparse (keyword-based, e.g. BM25) retrieval — getting the semantic understanding of one and the exact-match precision of the other.

Embedding search is great for meaning ('refund policy' matches 'how do I get my money back') but weak for exact terms (product names, error codes, IDs). Keyword search is the opposite. Hybrid search runs both in parallel, fuses scores, and returns a set that beats either alone.

Example

User asks about 'error code E-1042'. Pure embedding misses the exact code; pure keyword misses 'this error'. Hybrid retrieves both signals and finds the right doc.

How OpenLabor uses it

OpenLabor uses hybrid search whenever your data includes structured identifiers (codes, SKUs, IDs).

Is hybrid search always better?

Almost always for production RAG. The exception is purely conceptual content (philosophy, pure prose) where exact-match adds noise.

Related: rag, vector-store, reranker.

AI Labor Glossary