Module rabitq 

RaBitQ Quantization for OmenDB

RaBitQ (SIGMOD 2025) provides flexible vector compression with arbitrary bit rates (2-8 bits per dimension) using optimal rescaling per vector.

Tiered Compression Strategy

• L0-L2 (hot): Full precision f32 (no compression)
• L3-L4 (warm): RaBitQ 4-bit (8× compression, 98% recall)
• L5-L6 (cold): RaBitQ 2-bit (16× compression, 95% recall)

Key Features

• Flexible compression (2, 3, 4, 5, 7, 8 bits/dimension)
• Optimal rescaling for each vector
• SIMD-accelerated distance (AVX2/NEON)
• Same query speed as scalar quantization
• Better accuracy than binary quantization

Structs

ADCTable

Asymmetric Distance Computation (ADC) lookup table for fast quantized search

QuantizedVector

A quantized vector with optimal rescaling

RaBitQ

RaBitQ quantizer

RaBitQParams

Configuration for RaBitQ quantization

TrainedParams

Trained quantization parameters computed from data

Enums

QuantizationBits

Number of bits per dimension for quantization