MUVERA
lastmod 2025-07-02created 2025-07-021 edits
Making multi-vector retrieval as fast as single-vector search
About
- Retrieval based on single-vector embeddings is fast and efficient, but recent multi-vector models such as ColBERT use many vectors per token to provide richer meaning and higher accuracy
- Multi-vector approaches require complex similarity computations such as Chamfer similarity, which greatly increases computation and retrieval cost and becomes an obstacle to large-scale real-time retrieval
- MUVERA, proposed by a Google research team, compresses multi-vector information into fixed-length vectors (FDE, Fixed Dimensional Encoding), performs ultra-fast retrieval with single-vector MIPS (maximum inner product search), and then reranks
- The approach is data-independent and has a theoretical basis (guaranteed approximation error for Chamfer similarity), achieving over 90% lower latency and over 10% higher recall compared with the existing PLAID
- FDE also supports compression (32x memory savings), and an open-source implementation and paper have been released, making it suitable for adoption in production search, recommendation, and NLP services