Hybrid Search

Hybrid search runs dense and sparse engines in parallel and merges the results: dense understands semantics but misses keywords; sparse does exact matching but doesn’t understand synonyms. Merging is normalized weighted sums or RRF (Reciprocal Rank Fusion, uses only ranks). Then neural cross-encoder re-ranking: query and document are fed together, the model compares word-by-word – slower but much more accurate than bi-encoder. No single strategy is reliable across all scenarios.

Related: [Dense Embedding], [Sparse BM25 Search], RAG