Expand description
Vector quantization codec family — binary, ternary, 4-bit scalar, residual, and unified layout types shared across HNSW hot-path, IVF-PQ, and BBQ.
Modules§
- bbq
- BBQ — Better Binary Quantization (Elasticsearch 9.1, mid-2025).
- codec
- Dual-phase
VectorCodectrait — the seam that makes future quantization algorithms drop-in additions rather than engine rewrites. - hamming
- Shared SIMD Hamming-distance kernel with runtime CPU-feature dispatch.
- layout
- Unified Quantized Vector layout — the cache-aligned superset format that absorbs binary / ternary (BitNet 1.58) / 4-bit scalar / residual codecs without polymorphic indirection in the hot path.
- opq
- Optimized Product Quantization (OPQ) — Non-Para OPQ via iterative SVD-Procrustes rotation that minimizes PQ reconstruction error, yielding 10–20% recall improvement over vanilla PQ at equal memory.
- opq_
kmeans - Lloyd’s k-means for OPQ codebook training.
- opq_
rotation - Shared utilities for OPQ codebook training.
- rabitq
- RaBitQ — 1-bit quantization with O(1/√D) MSE error bound (SIGMOD 2024).
- ternary
- BitNet b1.58 ternary quantization codec.