innr 0.2.2

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
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228

//! ColBERT-style MaxSim Late Interaction Scoring
//!
//! MaxSim computes the sum of per-query-token maximum similarities across
//! document tokens. This is the scoring function used by ColBERT and its
//! descendants (ColBERTv2, PLAID, ColPali).
//!
//! The key trade-off: MaxSim retains token-level expressiveness (like
//! cross-encoders) but uses simple aggregation (like bi-encoders), enabling
//! precomputation of document token embeddings.
//!
//! ```bash
//! cargo run --example maxsim_colbert --release --features maxsim
//! ```

use innr::{dot, maxsim};
use std::time::Instant;

fn main() {
    println!("ColBERT MaxSim Late Interaction Scoring");
    println!("=======================================\n");

    // 1. Basic correctness demonstration
    demo_basic();

    // 2. Non-commutativity (a common source of bugs)
    demo_non_commutative();

    // 3. Realistic-scale scoring with timing
    demo_realistic_scale();

    println!("Done!");
}

// =============================================================================
// Demos
// =============================================================================

fn demo_basic() {
    println!("1. Basic MaxSim Scoring");
    println!("   --------------------\n");

    println!("   MaxSim(Q, D) = sum_i max_j (q_i . d_j)\n");

    // Small example: 2 query tokens, 3 doc tokens, dim=4
    let q0 = [1.0f32, 0.0, 0.0, 0.0]; // "selects" docs with high dim-0
    let q1 = [0.0f32, 1.0, 0.0, 0.0]; // "selects" docs with high dim-1

    let d0 = [0.9f32, 0.1, 0.0, 0.0]; // high dim-0 -> matches q0
    let d1 = [0.1f32, 0.8, 0.0, 0.0]; // high dim-1 -> matches q1
    let d2 = [0.5f32, 0.5, 0.0, 0.0]; // moderate on both

    let query: Vec<&[f32]> = vec![&q0, &q1];
    let doc: Vec<&[f32]> = vec![&d0, &d1, &d2];

    let score = maxsim(&query, &doc);

    // Manual computation:
    //   q0 best match: max(0.9, 0.1, 0.5) = 0.9 (d0)
    //   q1 best match: max(0.1, 0.8, 0.5) = 0.8 (d1)
    //   total = 0.9 + 0.8 = 1.7
    println!("   Query tokens: 2, Doc tokens: 3, Dim: 4");
    println!("   MaxSim score: {:.4}", score);
    println!("   Expected:     1.7000");
    assert!((score - 1.7).abs() < 1e-5, "basic MaxSim mismatch");
    println!();
}

fn demo_non_commutative() {
    println!("2. MaxSim Is Not Commutative");
    println!("   -------------------------\n");

    println!("   maxsim(Q, D) != maxsim(D, Q) in general.");
    println!("   The first argument is always the \"query\" side.\n");

    // 1 query token vs 3 doc tokens
    let q0 = [1.0f32, 0.0, 0.0, 0.0];

    let d0 = [0.5f32, 0.5, 0.0, 0.0];
    let d1 = [0.3f32, 0.7, 0.0, 0.0];
    let d2 = [0.8f32, 0.2, 0.0, 0.0];

    let query: Vec<&[f32]> = vec![&q0];
    let doc: Vec<&[f32]> = vec![&d0, &d1, &d2];

    let score_qd = maxsim(&query, &doc);
    let score_dq = maxsim(&doc, &query);

    // maxsim(Q, D): 1 query token, picks best of 3 docs
    //   q0 best: max(0.5, 0.3, 0.8) = 0.8
    //   total = 0.8
    //
    // maxsim(D, Q): 3 "query" tokens, each picks best of 1 doc
    //   d0 best: 0.5
    //   d1 best: 0.3
    //   d2 best: 0.8
    //   total = 0.5 + 0.3 + 0.8 = 1.6
    println!("   maxsim(Q[1], D[3]) = {:.4}", score_qd);
    println!("   maxsim(D[3], Q[1]) = {:.4}", score_dq);
    println!(
        "   Difference:          {:.4}\n",
        (score_qd - score_dq).abs()
    );

    assert!((score_qd - 0.8).abs() < 1e-5);
    assert!((score_dq - 1.6).abs() < 1e-5);
}

fn demo_realistic_scale() {
    println!("3. Realistic Scale: 32 Query Tokens x 128 Doc Tokens x 128d");
    println!("   ----------------------------------------------------------\n");

    let dim = 128;
    let n_query_tokens = 32;
    let n_doc_tokens = 128;

    // Generate synthetic normalized embeddings
    let query_vecs: Vec<Vec<f32>> = (0..n_query_tokens)
        .map(|i| generate_normalized(dim, i as u64))
        .collect();
    let doc_vecs: Vec<Vec<f32>> = (0..n_doc_tokens)
        .map(|i| generate_normalized(dim, (i + 1000) as u64))
        .collect();

    let query_refs: Vec<&[f32]> = query_vecs.iter().map(|v| v.as_slice()).collect();
    let doc_refs: Vec<&[f32]> = doc_vecs.iter().map(|v| v.as_slice()).collect();

    // -- innr::maxsim --
    let start = Instant::now();
    let score = maxsim(&query_refs, &doc_refs);
    let innr_time = start.elapsed();

    // -- Naive loop for verification --
    let start = Instant::now();
    let naive_score = naive_maxsim(&query_refs, &doc_refs);
    let naive_time = start.elapsed();

    println!("   innr::maxsim score:  {:.6}", score);
    println!("   Naive loop score:    {:.6}", naive_score);
    println!(
        "   Match:               {}",
        if (score - naive_score).abs() < 1e-3 {
            "yes"
        } else {
            "NO -- divergence"
        }
    );
    assert!(
        (score - naive_score).abs() < 1e-3,
        "maxsim vs naive diverged: {} vs {}",
        score,
        naive_score
    );
    println!();
    println!("   Single-pair timing:");
    println!("     innr::maxsim: {:?}", innr_time);
    println!("     Naive loop:   {:?}", naive_time);
    println!();

    // -- Batch timing: score query against many documents --
    let n_docs = 1000;
    let all_docs: Vec<Vec<Vec<f32>>> = (0..n_docs)
        .map(|doc_id| {
            (0..n_doc_tokens)
                .map(|tok| generate_normalized(dim, (doc_id * n_doc_tokens + tok + 5000) as u64))
                .collect()
        })
        .collect();

    let start = Instant::now();
    let mut scores: Vec<f32> = Vec::with_capacity(n_docs);
    for doc in &all_docs {
        let refs: Vec<&[f32]> = doc.iter().map(|v| v.as_slice()).collect();
        scores.push(maxsim(&query_refs, &refs));
    }
    let batch_time = start.elapsed();
    std::hint::black_box(&scores);

    println!(
        "   Batch: {} docs scored in {:?} ({:.1} us/doc)",
        n_docs,
        batch_time,
        batch_time.as_micros() as f64 / n_docs as f64
    );

    // Show top-5
    let mut indexed: Vec<(usize, f32)> = scores.iter().copied().enumerate().collect();
    indexed.sort_by(|a, b| b.1.total_cmp(&a.1));
    println!("   Top-5 scores:");
    for (rank, &(idx, score)) in indexed.iter().take(5).enumerate() {
        println!("     #{}: doc {} = {:.4}", rank + 1, idx, score);
    }
    println!();
}

// =============================================================================
// Helpers
// =============================================================================

/// Naive O(|Q| * |D| * dim) MaxSim using innr::dot.
fn naive_maxsim(query: &[&[f32]], doc: &[&[f32]]) -> f32 {
    query
        .iter()
        .map(|q| {
            doc.iter()
                .map(|d| dot(q, d))
                .fold(f32::NEG_INFINITY, f32::max)
        })
        .sum()
}

/// Deterministic pseudo-random normalized embedding.
fn generate_normalized(dim: usize, seed: u64) -> Vec<f32> {
    let mut v: Vec<f32> = (0..dim)
        .map(|i| {
            let x = seed
                .wrapping_mul(6364136223846793005)
                .wrapping_add(i as u64 * 1442695040888963407);
            ((x >> 33) as f32 / (1u64 << 31) as f32) * 2.0 - 1.0
        })
        .collect();
    let norm: f32 = v.iter().map(|x| x * x).sum::<f32>().sqrt();
    if norm > f32::EPSILON {
        for x in &mut v {
            *x /= norm;
        }
    }
    v
}