rvf-quant 0.2.0

RuVector Format temperature-tiered vector quantization (f32/f16/u8/binary)
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38

# rvf-quant


Temperature-tiered vector quantization for RuVector Format.

## Overview


`rvf-quant` provides quantization codecs that reduce vector storage size based on access temperature:

- **f32** -- full precision for hot vectors
- **f16** -- half precision for warm vectors
- **u8** -- scalar quantization for cool vectors
- **binary** -- 1-bit quantization for cold/archive vectors
- **RaBitQ** -- randomized 1-bit codes (32x code compression) with unbiased
  distance estimation, used by the runtime's opt-in two-stage query path
  (recall@10 0.972 after exact rescore at 100k x 64-dim, Windows x64
  criterion release)
- **Automatic tiering** -- promote/demote vectors based on access patterns

## Usage


```toml
[dependencies]
rvf-quant = "0.2"
```

## Breaking Changes in 0.2.0


- `decode_sketch_seg` now returns `Result` instead of panicking on
  malformed input (crafted-file DoS hardening).

## Features


- `std` (default) -- enable `std` support
- `simd` -- enable SIMD-accelerated quantization

## License


MIT OR Apache-2.0