Module binary_quantize 

Binary Quantization for Vector Embeddings

Compresses fp32 vectors to binary (1 bit per dimension) for ultra-fast approximate nearest neighbor search using Hamming distance.

Compression Ratio

• fp32: 4 bytes per dimension
• binary: 1 bit per dimension = 32x compression

Example: 1024-dim vector

• fp32: 4096 bytes
• binary: 128 bytes

Algorithm

Simple sign-based quantization:

• positive values → 1
• negative/zero values → 0

For normalized embeddings (e.g., from sentence transformers), this preserves ~95-97% of retrieval quality.

Usage

// Quantize a vector
let binary = BinaryVector::from_f32(&embedding);

// Compute Hamming distance (number of differing bits)
let distance = binary.hamming_distance(&other);

// For retrieval: lower Hamming distance = more similar

References

• “Embedding Quantization” - HuggingFace Blog
• Binary embedding with Matryoshka representation learning

Structs

BinaryIndex

Index of binary vectors for fast batch search

BinarySearchResult

Result from binary search with rescoring capability

BinaryVector

Binary quantized vector stored as packed u64 words

Functions

hamming_distance_simd

Compute Hamming distance between two packed binary vectors using SIMD