hybrid-search

Hybrid search combines BM25 keyword matching with vector-based semantic similarity to retrieve results that match both exact terms and conceptual meaning. This fused ranking eliminates the trade-off between precision and recall, improving retrieval for domains like product catalogs, legal documents, and knowledge bases where both exact terminology and conceptual relevance matter.

Hybrid search requires Milvus as the vector search engine, combined with an embedding model for dense vectors and a BM25-based sparse indexing mechanism. Milvus handles the vector storage and retrieval while BM25 processes keyword matching; together they enable fused ranking of results.

Use hybrid search when your domain requires both exact term matching and conceptual understanding—product catalogs with model numbers, technical documentation with domain-specific terminology, or knowledge bases where precision on keywords coexists with relevance through meaning. Hybrid search outperforms single-method approaches in these scenarios.

Yes, hybrid search is specifically designed for technical and legal documents where domain terms must match exactly while conceptual understanding is equally important. The fused BM25 and semantic matching handles both precise terminology lookup and contextual relevance simultaneously.

Hybrid search fuses BM25 sparse keyword matching with dense vector embeddings through unified ranking. BM25 scores keywords based on term frequency and document length, while vectors capture semantic meaning; the combined ranking surfaces results that satisfy both criteria.