Redisearch New Vector Quantization
4 days ago
- #Redis
- #compression
- #vector-search
- Redis Query Engine now supports Quantization and Dimensionality Reduction for vector search, reducing memory footprint by 26–37%.
- Intel SVS-VAMANA combines Vamana graph-based search with LVQ and LeanVec compression for efficient vector search.
- LVQ (Locally-adaptive Vector Quantization) adapts per-vector normalization and scalar quantization, optimizing memory usage.
- LeanVec applies dimensionality reduction before LVQ, further cutting memory and compute costs for high-dimensional vectors.
- Benchmarks show SVS-VAMANA reduces memory usage by 51–74% for vector indexes while maintaining search accuracy.
- Performance improvements include up to 144% higher QPS and 60% lower latency on Intel platforms.
- ARM platforms currently perform better with HNSW, as SVS-VAMANA optimizations are primarily for x86.
- Ingestion times are slower with SVS-VAMANA, especially on ARM, where it can be up to 9× slower than HNSW.
- Developers can enable compression via Redis commands, with LVQ recommended for <768 dimensions and LeanVec for ≥768 dimensions.
- Future enhancements include multi-vector support and further optimizations for ARM and other platforms.