Module int8_quantize 

int8 Quantization for Vector Embeddings

Compresses fp32 vectors to int8 (8 bits per dimension) for efficient storage and fast approximate distance computation.

Compression Ratio

• fp32: 4 bytes per dimension
• int8: 1 byte per dimension = 4x compression

Example: 1024-dim vector

• fp32: 4096 bytes
• int8: 1024 bytes

Quantization Methods

Symmetric Quantization

Maps [-max_abs, +max_abs] → [-127, +127]

• scale = max(|v|) / 127
• quantized = round(v / scale)

Asymmetric Quantization

Maps [min, max] → [0, 255]

• scale = (max - min) / 255
• zero_point = round(-min / scale)
• quantized = round(v / scale) + zero_point

Usage

// Quantize a vector (symmetric)
let int8 = Int8Vector::from_f32(&embedding);

// Compute dot product (SIMD accelerated)
let dot = int8.dot_product(&other);

// Rescore binary search candidates
let rescored = int8.rescore_candidates(&binary_results, &query);

Structs

Int8Index

Index of int8 vectors for batch operations

Int8Vector

int8 quantized vector with scale factor

Functions

dot_product_i8_f32_simd

Compute dot product of i8 vector with f32 query (asymmetric)

dot_product_i8_simd

Compute dot product of two i8 vectors using SIMD

l2_squared_i8_simd

Compute L2 squared distance between two i8 vectors