hermes_core/structures/vector/quantization/
mod.rs

1//! Vector quantization methods for IVF indexes
2//!
3//! This module provides different quantization strategies:
4//! - `rabitq` - RaBitQ binary quantization (32x compression)
5//! - `pq` - Product Quantization with OPQ and anisotropic loss (ScaNN-style)
6//!
7//! All quantizers implement the `Quantizer` trait for use with IVF indexes.
8
9mod pq;
10mod rabitq;
11
12pub use pq::{DistanceTable, PQCodebook, PQConfig, PQVector};
13pub use rabitq::{QuantizedQuery, QuantizedVector, RaBitQCodebook, RaBitQConfig};
14
15use super::ivf::cluster::QuantizedCode;
16
17/// Trait for vector quantization methods
18///
19/// Quantizers encode vectors into compact codes and provide
20/// fast approximate distance computation.
21pub trait Quantizer: Clone + Send + Sync {
22    /// The quantized code type
23    type Code: QuantizedCode;
24
25    /// Configuration type
26    type Config: Clone;
27
28    /// Query-specific precomputed data for fast distance computation
29    type QueryData;
30
31    /// Encode a vector (optionally relative to a centroid)
32    fn encode(&self, vector: &[f32], centroid: Option<&[f32]>) -> Self::Code;
33
34    /// Prepare query-specific data for fast distance computation
35    fn prepare_query(&self, query: &[f32], centroid: Option<&[f32]>) -> Self::QueryData;
36
37    /// Compute approximate distance using precomputed query data
38    fn compute_distance(&self, query_data: &Self::QueryData, code: &Self::Code) -> f32;
39
40    /// Decode a code back to approximate vector (if supported)
41    fn decode(&self, _code: &Self::Code) -> Option<Vec<f32>> {
42        None
43    }
44
45    /// Memory usage of the quantizer itself (codebooks, etc.)
46    fn size_bytes(&self) -> usize;
47}
hermes_core/structures/vector/quantization/mod.rs

hermes_core/structures/vector/quantization/
mod.rs
