Expand description
Binary (1-bit) quantization — a second optional in-memory index mode.
Each dimension is thresholded at its learned per-dimension mean and stored as
a single bit (1 if above the mean, else 0), packed into u64 words. Distance
is Hamming distance (popcount of XORed codes), which approximates angular
similarity. This is an extreme compressor: D bits/vector = D/8 bytes vs
D·4 for float32 → ~32× smaller.
Binary codes are coarse: on their own they give a fast, tiny filter with
modest recall. The intended pipeline is binary-Hamming for candidate
generation followed by a full-precision rerank of the top candidates — see
examples/quant_bench.rs, which measures both raw and reranked recall.
Structs§
- Binary
Quantized Index - Binary
Quantizer - Per-dimension mean-threshold binary quantizer.