Module quantizer

Distance acceleration through vector quantization.

The Quantizer trait abstracts over different strategies for accelerating distance computations in high-dimensional spaces. The HNSW graph stores both full-precision vectors (for exact operations like trajectories and signatures) and compact codes (for fast approximate distance during graph construction/search).

Two-Phase Distance Computation

1. Candidate selection: Use fast approximate distance on compact codes
2. Final ranking: Use exact distance on full vectors

This mirrors production systems like Qdrant (Scalar Quantization), Faiss (Product Quantization), and Weaviate (Binary Quantization).

Available Strategies

Strategy	Code size (D=768)	Speedup	Recall impact
NoQuantizer	0 bytes	1×	None
Scalar (SQ8)	768 bytes	~4×	< 1%
Product (PQ96)	96 bytes	~8×	1-3%
Binary (BQ)	96 bytes	~32×	5-10%

Example

use cvx_core::traits::quantizer::{Quantizer, NoQuantizer, L2Fn};

// No acceleration (default, exact distances)
let q = NoQuantizer::new(L2Fn);
let code = q.encode(&[1.0, 2.0, 3.0]);

Structs

L2Fn

Simple L2 distance function for use with quantizers.

NoQuantizer

Identity quantizer — no acceleration, exact distances only.

ScalarQuantizer

Scalar Quantization: compress each float32 dimension to uint8.

Traits

Quantizer

Acceleration strategy for distance computations.