Module quantization 

Scalar Quantization (SQ8) and Binary Quantization for memory-efficient vector storage.

This module implements quantization strategies to reduce memory usage:

Benefits

Metric	f32	SQ8	Binary
RAM/vector (768d)	3 KB	770 bytes	96 bytes
Cache efficiency	Baseline	~4x better	~32x better
Recall loss	0%	~0.5-1%	~5-10%

Engine integration status

The figures above describe the quantization primitives themselves. In the collection query path: RaBitQ (binary traversal backend), SQ8 (int8 traversal backend, Euclidean/Cosine) and PQ (ADC rescoring) are wired end-to-end. Persistence across reopens covers TRAIN-QUANTIZER-produced artifacts (rabitq.idx, sq8.idx, codebook.pq) plus lazily-trained RaBitQ/SQ8 quantizers (persisted by the full flush); a PQ quantizer trained lazily from inserts (no TRAIN statement) is in-memory only and retrains after a restart. The Binary collection mode stays full-precision f32 in the search path. See docs/guides/QUANTIZATION.md.

Structs

BinaryQuantizedVector

A binary quantized vector using 1-bit per dimension.

PQCodebook

Per-subspace centroid tables learned with k-means.

PQVector

Compressed representation of a vector: one centroid id per subspace.

PreparedQuery

Preprocessed query data for RaBitQ distance computation.

ProductQuantizer

Product quantizer model and helpers for train/encode/decode.

QuantizedVector

A quantized vector using 8-bit scalar quantization.

RaBitQCorrection

Scalar correction factors for a RaBitQ-encoded vector.

RaBitQIndex

RaBitQ index holding the random rotation matrix and dataset centroid.

RaBitQVector

Binary-quantized vector with scalar correction factors.

Enums

StorageMode

Storage mode for vectors.

Constants

STORAGE_MODE_NAMES

Canonical names of every StorageMode variant, in declaration order.

Traits

QuantizationCodec

Trait for serializing and deserializing quantized vectors to/from bytes.

Functions

cosine_similarity_quantized

Computes approximate cosine similarity between a query (f32) and quantized vector.

cosine_similarity_quantized_simd

SIMD-optimized cosine similarity between f32 query and SQ8 vector.

dot_product_quantized

Computes the approximate dot product between a query vector (f32) and a quantized vector.

dot_product_quantized_simd

Dot product between f32 query and SQ8 quantized vector with 8-wide unrolling.

euclidean_squared_quantized

Computes the approximate squared Euclidean distance between a query (f32) and quantized vector.

euclidean_squared_quantized_simd

SIMD-optimized squared Euclidean distance between f32 query and SQ8 vector.

train_opq

Train a PQ codebook with optional PCA pre-rotation.