Module binary 

Binary Quantization (BBQ) for OmenDB

1-bit quantization with SIMD-optimized Hamming distance.

Algorithm

• Quantize: bit[d] = 1 if f32[d] > threshold[d] else 0
• Distance: Hamming distance via XOR + popcnt
• Correction: Apply norm-based correction for accurate ranking

Performance

• 32x compression (f32 → 1 bit)
• 2-4x faster search than SQ8 (SIMD Hamming is extremely fast)
• ~85% raw recall, ~95-98% with rescore

When to Use

• Dimensions >= 384 (below this, SQ8 has better recall)
• Large datasets (>100K vectors) where memory matters
• Cost-sensitive deployments

Structs

BinaryParams

Binary quantization parameters

Functions

corrected_distance

Compute corrected distance for binary quantization

hamming_distance

Compute Hamming distance between two binary vectors