Hybrid search runs keyword and semantic retrieval in parallel and merges their results, combining exact-match precision with meaning-based recall.
Fusion is usually done with reciprocal rank fusion, which combines the two orderings by position rather than by raw scores — a practical choice, since BM25 scores and vector distances are not on comparable scales.
The reason hybrid wins is that the two retrievers fail in complementary ways. Keyword search misses paraphrase; vector search misses identifiers. Their union misses far less than either alone, and a re-ranker over the merged candidate set restores precision at the top.