Hybrid search

Hybrid search combines two signals: semantic similarity from embeddings and lexical relevance from matching words or terms.

Tiny example

For “ERR-1047 session reset,” keyword search protects the exact error code while semantic search finds passages about interrupted sessions. A fusion step merges the two ranked lists.

query → keyword results ─┐
                         ├→ rank fusion → candidates
query → vector results  ─┘

FDE note

The combination method matters. Score scales from two retrievers may not be directly comparable, so rank-based fusion is often a practical baseline.¹ Evaluate hybrid search by query type; it should earn its added complexity.

Footnotes

1. Elastic documents hybrid search and reciprocal rank fusion as a way to combine independently ranked result sets. ↩