innr 0.5.0

SIMD-accelerated vector similarity primitives with binary, ternary, and scalar quantization
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
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197

//! Generic [`Distance`](crate::distance::Distance) trait for using innr's
//! metrics as a pluggable backend.
//!
//! The free functions elsewhere in this crate ([`crate::cosine`], [`crate::dot`],
//! [`crate::l2_distance`], ...) are the fast path when you know the metric at the
//! call site. This module wraps them behind a trait so a generic index (a custom
//! HNSW, a brute-force searcher, a clustering routine) can be parameterized over
//! the metric instead of hard-coding one.
//!
//! The trait shape intentionally mirrors the convention used by `anndists` /
//! `hnsw_rs` (`eval(&self, &[T], &[T]) -> f32`, smaller = closer) so innr's
//! metrics drop into code already written against that ecosystem.
//!
//! # Convention
//!
//! `eval` returns a **distance**: smaller means more similar. Similarity metrics
//! are converted accordingly:
//!
//! - [`DistCosine`](crate::distance::DistCosine) returns `1 - cosine_similarity` (range `[0, 2]`).
//! - [`DistDot`](crate::distance::DistDot) returns `-dot` so that larger dot products sort first.
//! - [`DistL2`](crate::distance::DistL2), [`DistL1`](crate::distance::DistL1) are already distances.
//! - [`DistHamming`](crate::distance::DistHamming) returns the bit-Hamming distance as `f32`.
//! - [`DistSlotU32`](crate::distance::DistSlotU32) returns the integer-slot Hamming distance as `f32`.
//!
//! # Example
//!
//! ```rust
//! use innr::distance::{Distance, DistCosine, DistL2};
//!
//! fn nearest<D: Distance<f32>>(metric: &D, query: &[f32], corpus: &[Vec<f32>]) -> usize {
//!     corpus
//!         .iter()
//!         .enumerate()
//!         .min_by(|(_, a), (_, b)| {
//!             metric
//!                 .eval(query, a)
//!                 .total_cmp(&metric.eval(query, b))
//!         })
//!         .map(|(i, _)| i)
//!         .unwrap()
//! }
//!
//! let corpus = vec![vec![1.0, 0.0], vec![0.0, 1.0], vec![1.0, 1.0]];
//! assert_eq!(nearest(&DistCosine, &[1.0, 0.1], &corpus), 0);
//! assert_eq!(nearest(&DistL2, &[0.9, 0.9], &corpus), 2);
//! ```

use crate::{cosine, dot, hamming_distance, l1_distance, l2_distance, slot::slot_hamming_u32};

/// A distance metric over slices of `T`. `eval` returns a distance: smaller is
/// more similar.
///
/// Mirrors the `anndists` / `hnsw_rs` trait shape so innr's metrics can back a
/// generic index written against that convention.
pub trait Distance<T> {
    /// Distance between `a` and `b`. Smaller means more similar.
    fn eval(&self, a: &[T], b: &[T]) -> f32;
}

/// Cosine distance: `1 - cosine_similarity`. Range `[0, 2]`.
#[derive(Debug, Clone, Copy, Default)]
pub struct DistCosine;

impl Distance<f32> for DistCosine {
    #[inline]
    fn eval(&self, a: &[f32], b: &[f32]) -> f32 {
        1.0 - cosine(a, b)
    }
}

/// Negated dot product, so that larger inner products sort first. Use with
/// normalized vectors for a maximum-inner-product search.
#[derive(Debug, Clone, Copy, Default)]
pub struct DistDot;

impl Distance<f32> for DistDot {
    #[inline]
    fn eval(&self, a: &[f32], b: &[f32]) -> f32 {
        -dot(a, b)
    }
}

/// Euclidean (L2) distance.
#[derive(Debug, Clone, Copy, Default)]
pub struct DistL2;

impl Distance<f32> for DistL2 {
    #[inline]
    fn eval(&self, a: &[f32], b: &[f32]) -> f32 {
        l2_distance(a, b)
    }
}

/// Manhattan (L1) distance.
#[derive(Debug, Clone, Copy, Default)]
pub struct DistL1;

impl Distance<f32> for DistL1 {
    #[inline]
    fn eval(&self, a: &[f32], b: &[f32]) -> f32 {
        l1_distance(a, b)
    }
}

/// Bit-Hamming distance over byte-packed binary vectors (see
/// [`crate::hamming_distance`]).
#[derive(Debug, Clone, Copy, Default)]
pub struct DistHamming;

impl Distance<u8> for DistHamming {
    #[inline]
    fn eval(&self, a: &[u8], b: &[u8]) -> f32 {
        hamming_distance(a, b) as f32
    }
}

/// Integer-slot Hamming distance over `u32` slots (see
/// [`crate::slot_hamming_u32`]). The natural metric for MinHash sketches.
#[derive(Debug, Clone, Copy, Default)]
pub struct DistSlotU32;

impl Distance<u32> for DistSlotU32 {
    #[inline]
    fn eval(&self, a: &[u32], b: &[u32]) -> f32 {
        slot_hamming_u32(a, b) as f32
    }
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn cosine_distance_zero_for_parallel() {
        let d = DistCosine.eval(&[1.0, 2.0, 3.0], &[2.0, 4.0, 6.0]);
        assert!(
            d.abs() < 1e-6,
            "parallel vectors should have cosine distance 0, got {d}"
        );
    }

    #[test]
    fn dot_distance_orders_by_inner_product() {
        // larger dot product -> smaller (more negative) distance
        let near = DistDot.eval(&[1.0, 0.0], &[1.0, 0.0]);
        let far = DistDot.eval(&[1.0, 0.0], &[0.0, 1.0]);
        assert!(near < far);
    }

    #[test]
    fn l2_matches_free_function() {
        let a = [1.0f32, 2.0, 3.0];
        let b = [4.0f32, 0.0, 3.0];
        assert_eq!(DistL2.eval(&a, &b), l2_distance(&a, &b));
    }

    #[test]
    fn l1_matches_free_function() {
        let a = [1.0f32, 2.0, 3.0];
        let b = [4.0f32, 0.0, 3.0];
        assert_eq!(DistL1.eval(&a, &b), l1_distance(&a, &b));
    }

    #[test]
    fn hamming_matches_free_function() {
        let a = [0b1111_0000u8, 0xFF];
        let b = [0b1010_1010u8, 0x00];
        assert_eq!(DistHamming.eval(&a, &b), hamming_distance(&a, &b) as f32);
    }

    #[test]
    fn slot_distance_counts_differing_slots() {
        let a = [1u32, 2, 3, 4];
        let b = [1u32, 0, 3, 9];
        assert_eq!(DistSlotU32.eval(&a, &b), 2.0);
    }

    // A generic consumer compiles and runs against any Distance impl.
    fn closest<T, D: Distance<T>>(metric: &D, q: &[T], corpus: &[Vec<T>]) -> usize {
        corpus
            .iter()
            .enumerate()
            .min_by(|(_, a), (_, b)| metric.eval(q, a).total_cmp(&metric.eval(q, b)))
            .map(|(i, _)| i)
            .unwrap()
    }

    #[test]
    fn generic_index_over_metric() {
        let corpus = vec![vec![1.0f32, 0.0], vec![0.0, 1.0], vec![1.0, 1.0]];
        assert_eq!(closest(&DistCosine, &[1.0, 0.05], &corpus), 0);
        assert_eq!(closest(&DistL2, &[0.95, 0.95], &corpus), 2);

        let sketches = vec![vec![1u32, 2, 3, 4], vec![1, 2, 3, 9], vec![9, 9, 9, 9]];
        assert_eq!(closest(&DistSlotU32, &[1, 2, 3, 4], &sketches), 0);
    }
}