What it is
Hybrid search blends two retrieval methods: dense vector search that captures meaning, and sparse keyword search like BM25 that matches exact terms. Each covers the other's weak spots.
Why blend them
- Dense search finds semantically related text even when words differ, but it can miss rare exact terms like a product code or a name.
- Sparse search nails exact keywords and rare tokens, but it fails when the query and document use different words for the same idea.
Combining them recovers documents that either method alone would drop.
Fusing the scores
The two methods produce scores on different scales, so you cannot just add them. A common fix is reciprocal rank fusion, which combines results by rank rather than raw score. Each document gets credit based on its position in each list, so a high rank in either method lifts it.
- It needs no score normalization.
- It is robust to one method having odd score magnitudes.
- Weights can tilt the blend toward keywords or semantics.
Key idea
Hybrid search fuses dense semantic and sparse keyword retrieval, often via reciprocal rank fusion, so exact terms and meaning are both covered.