RAG
Hybrid search
Hybrid search combines sparse lexical ranking with dense semantic ranking.
How it works
The two retrievers capture complementary signals. Score normalization or rank fusion is needed because their raw scores are not directly comparable.
A concrete example
Illustrative: reciprocal-rank fusion combines dense rank 2 and BM25 rank 5 for one document.
Do not confuse it with Reranking
hybrid search merges retrievers; reranking applies a later, usually more expensive scorer.