Weaviate 1.31: MUVERA Encoding for Efficient Multi-Vector Embeddings

AI Impact Summary

Weaviate 1.31 introduces MUVERA, an encoding algorithm designed to dramatically reduce the memory footprint and improve indexing speeds of multi-vector embeddings. MUVERA converts multi-vector embeddings into single fixed-size vectors by leveraging fixed dimensional encodings (FDEs), resulting in a roughly 70% reduction in memory usage and a significant decrease in import times (from 20+ minutes to 3-6 minutes). While this offers performance gains, it introduces a potential trade-off in recall quality that can be mitigated with increased HNSW ef values.