Hybrid Search

Hybrid search combines multiple retrieval methods -- most commonly sparse keyword search (BM25) and dense semantic search (embedding similarity) -- to produce better results than either method alone. The key insight is that these methods have complementary strengths: BM25 excels at exact term matching while dense retrieval excels at semantic understanding.

In practice, hybrid search runs both BM25 and dense retrieval in parallel, then combines their results using fusion techniques such as reciprocal rank fusion (RRF) or learned weighted combination. A common formula weights dense scores higher (e.g., 0.7 semantic + 0.3 BM25), though optimal weights depend on the domain and query types.

Hybrid search is the recommended approach for production RAG systems because it handles the full spectrum of user queries -- from exact-match lookups ("invoice #12345") to meaning-based searches ("documents about corporate restructuring"). Systems that rely solely on semantic search often fail on rare terms, product codes, and proper nouns where BM25 provides critical coverage.