A search architecture that combines dense vector embeddings (semantic search) with sparse representations like BM25 (lexical search) to achieve better overall search quality. The industry standard approach for production RAG systems in 2026.
Hybrid Search combines dense vector embeddings (semantic search) with sparse representations like BM25 or SPLADE (lexical search) into a single unified search system. This approach provides the best overall search quality by leveraging the complementary strengths of both methods.
Hybrid search executes two parallel searches:
Dense search is good at:
Sparse search is good at:
Combines rankings from both methods:
score(d) = Σ 1/(k + rank_i(d))
where k is typically 60
Combines scores with weights:
score = α × dense_score + (1-α) × sparse_score
Hybrid search is now the industry standard for production RAG systems, with major platforms supporting it natively:
Typical improvements over vector-only search:
Loading more......
Implementation-dependent; some vector databases include hybrid search at no additional cost.