multi-vector-search

Multi-vector search lets you embed and query separate fields—title, description, reviews, images—simultaneously, then fuse their relevance scores using RRF or weighted rankers. This preserves distinct semantic meaning in each field and returns results ranked by combined relevance across all facets.

Single-vector search treats an entire item as one embedding, losing nuance from distinct parts. Multi-vector search creates separate embeddings for title, description, reviews, and images, then combines their scores to capture how different fields contribute to relevance—critical for catalogs and documents with heterogeneous content.

Yes. Multi-vector search supports both fusion strategies: RRF (Reciprocal Rank Fusion) provides rank-based combination, while weighted rankers let you assign field-specific importance scores. This flexibility tailors ranking to your use case—balanced retrieval or emphasis on title over reviews, for example.

Multi-vector search uses Milvus as the vector database backend and integrates with Python embedding libraries to generate field-specific embeddings. This stack handles indexing separate field vectors and executing fused retrieval queries at scale.

Product catalogs, document repositories, and Q&A systems where items have distinct semantic parts—title, description, reviews, images—benefit most. Queries like 'smartphone with good camera' rely on matching across multiple fields to surface relevant results single-vector search would miss.

The Skill handles fusion of scores from separate field embeddings through RRF or weighted rankers, abstracting normalization details. You index each field with its own embedding independently; the fusion layer combines their rankings without requiring manual normalization.