MUVERA: Making multi-vector retrieval as fast as single-vector search
Google Research introduces MUVERA, a novel algorithm that transforms complex multi-vector retrieval into efficient single-vector maximum inner product search. The approach uses Fixed Dimensional Encodings (FDEs) to compress multi-vector sets into single vectors while preserving similarity relationships. MUVERA achieves 10% higher recall than existing methods with 90% reduced latency, making multi-vector retrieval practical for large-scale applications like search engines and recommendation systems.
>>Click here to continue<<
