Hybrid search combines two different retrieval methods — dense vector search and sparse keyword search — into a single query, so results reflect both semantic meaning and exact term matches. It is widely regarded as the gold standard for production retrieval because it captures the strengths of both approaches while covering each one’s weaknesses.
Vector search excels at understanding meaning, matching paraphrases and synonyms regardless of exact wording. Keyword search, typically using BM25, excels at exact matches: product codes, names, error strings, and rare jargon that embeddings may not represent well. Many real queries need both at once — a search for a specific model number described in natural language combines a precise term with a semantic intent.
The technical heart of hybrid search is fusing the two result sets, since vector and keyword scores live on different scales. Methods like reciprocal rank fusion or weighted score combination merge them into one ranking. Done well, hybrid search consistently outperforms either method alone, which is why most mature vector databases now support it natively.