jin 0.1.0

Approximate Nearest Neighbor Search: HNSW, DiskANN, IVF-PQ, ScaNN, 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
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382

//! Edge case tests for jin.
//!
//! Tests unusual inputs and boundary conditions that could cause failures.

use jin::hnsw::HNSWIndex;
use std::collections::HashSet;

fn normalize(v: &[f32]) -> Vec<f32> {
    let norm: f32 = v.iter().map(|x| x * x).sum::<f32>().sqrt();
    if norm < 1e-10 {
        v.to_vec()
    } else {
        v.iter().map(|x| x / norm).collect()
    }
}

// =============================================================================
// Dimension edge cases
// =============================================================================

#[test]
fn very_small_dimension() {
    let dim = 2; // Minimum practical dimension
    let mut hnsw = HNSWIndex::new(dim, 16, 16).expect("Failed to create");

    let vectors: Vec<Vec<f32>> = (0..50)
        .map(|i| {
            let angle = (i as f32) * 0.1;
            normalize(&[angle.cos(), angle.sin()])
        })
        .collect();

    for (i, v) in vectors.iter().enumerate() {
        hnsw.add(i as u32, v.clone()).expect("Failed to add");
    }
    hnsw.build().expect("Failed to build");

    let results = hnsw.search(&vectors[0], 5, 50).expect("Search failed");
    assert_eq!(results.len(), 5);
    assert_eq!(results[0].0, 0); // Should find itself
}

#[test]
fn high_dimension() {
    let dim = 1024; // Higher than typical BERT (768)
    let mut hnsw = HNSWIndex::new(dim, 16, 16).expect("Failed to create");

    let vectors: Vec<Vec<f32>> = (0..20)
        .map(|i| {
            let v: Vec<f32> = (0..dim).map(|d| ((i * d) as f32).sin()).collect();
            normalize(&v)
        })
        .collect();

    for (i, v) in vectors.iter().enumerate() {
        hnsw.add(i as u32, v.clone()).expect("Failed to add");
    }
    hnsw.build().expect("Failed to build");

    let results = hnsw.search(&vectors[10], 5, 50).expect("Search failed");
    assert!(!results.is_empty());
}

// =============================================================================
// Vector count edge cases
// =============================================================================

#[test]
fn small_index() {
    let dim = 32;
    let mut hnsw = HNSWIndex::new(dim, 16, 16).expect("Failed to create");

    // Only 3 vectors (less than M)
    let vectors: Vec<Vec<f32>> = (0..3).map(|i| normalize(&vec![i as f32; dim])).collect();

    for (i, v) in vectors.iter().enumerate() {
        hnsw.add(i as u32, v.clone()).expect("Failed to add");
    }
    hnsw.build().expect("Failed to build");

    let results = hnsw.search(&vectors[0], 10, 50).expect("Search failed");
    assert_eq!(results.len(), 3, "Should return all 3 vectors");
}

#[test]
fn index_with_m_vectors() {
    // Exactly M vectors - boundary case for neighbor lists
    let dim = 32;
    let m = 16;
    let mut hnsw = HNSWIndex::new(dim, m, m).expect("Failed to create");

    let vectors: Vec<Vec<f32>> = (0..m)
        .map(|i| {
            let v: Vec<f32> = (0..dim).map(|d| ((i + d) as f32 * 0.1).sin()).collect();
            normalize(&v)
        })
        .collect();

    for (i, v) in vectors.iter().enumerate() {
        hnsw.add(i as u32, v.clone()).expect("Failed to add");
    }
    hnsw.build().expect("Failed to build");

    let results = hnsw.search(&vectors[0], m, 50).expect("Search failed");
    assert_eq!(results.len(), m);
}

// =============================================================================
// Special vector patterns
// =============================================================================

#[test]
fn identical_vectors() {
    let dim = 32;
    let mut hnsw = HNSWIndex::new(dim, 16, 16).expect("Failed to create");

    // All vectors are identical
    let base = normalize(&vec![1.0; dim]);
    for i in 0..10 {
        hnsw.add(i as u32, base.clone()).expect("Failed to add");
    }
    hnsw.build().expect("Failed to build");

    let results = hnsw.search(&base, 5, 50).expect("Search failed");

    // All results should have distance ~0
    for (_, dist) in &results {
        assert!(*dist < 0.01, "Identical vectors should have ~0 distance");
    }
}

#[test]
fn nearly_identical_vectors() {
    let dim = 64;
    let mut hnsw = HNSWIndex::new(dim, 16, 16).expect("Failed to create");

    let base: Vec<f32> = (0..dim).map(|i| i as f32 * 0.01).collect();

    // Add slightly perturbed versions
    for i in 0..50 {
        let mut v = base.clone();
        v[i % dim] += 1e-5 * (i as f32);
        hnsw.add(i as u32, normalize(&v)).expect("Failed to add");
    }
    hnsw.build().expect("Failed to build");

    let results = hnsw
        .search(&normalize(&base), 10, 100)
        .expect("Search failed");
    assert_eq!(results.len(), 10);
}

#[test]
fn well_clustered_vectors() {
    // Create two distinct clusters
    let dim = 32;
    let mut hnsw = HNSWIndex::new(dim, 16, 16).expect("Failed to create");

    // Cluster 1: centered around [1, 0, 0, ...]
    for i in 0..25 {
        let mut v = vec![0.0; dim];
        v[0] = 1.0;
        v[(i % (dim - 1)) + 1] = 0.1;
        hnsw.add(i as u32, normalize(&v)).expect("Failed to add");
    }

    // Cluster 2: centered around [-1, 0, 0, ...]
    for i in 25..50 {
        let mut v = vec![0.0; dim];
        v[0] = -1.0;
        v[(i % (dim - 1)) + 1] = 0.1;
        hnsw.add(i as u32, normalize(&v)).expect("Failed to add");
    }

    hnsw.build().expect("Failed to build");

    // Query from cluster 1
    let mut query = vec![0.0; dim];
    query[0] = 1.0;
    let results = hnsw
        .search(&normalize(&query), 10, 100)
        .expect("Search failed");

    // Should mostly find cluster 1 vectors (indices 0-24)
    let cluster1_count = results.iter().filter(|(i, _)| *i < 25).count();
    assert!(
        cluster1_count >= 8,
        "Should find mostly cluster 1 vectors, got {}/10",
        cluster1_count
    );
}

// =============================================================================
// Query edge cases
// =============================================================================

#[test]
fn query_not_in_index() {
    let dim = 32;
    let mut hnsw = HNSWIndex::new(dim, 16, 16).expect("Failed to create");

    // Add only positive vectors
    for i in 0..30 {
        let v: Vec<f32> = (0..dim).map(|d| ((i + d) as f32).abs()).collect();
        hnsw.add(i as u32, normalize(&v)).expect("Failed to add");
    }
    hnsw.build().expect("Failed to build");

    // Query with a negative vector (opposite direction)
    let query: Vec<f32> = (0..dim).map(|d| -1.0 * (d as f32 + 1.0)).collect();
    let results = hnsw
        .search(&normalize(&query), 5, 50)
        .expect("Search failed");

    assert_eq!(results.len(), 5);
    // Distances should be high (opposite direction)
    assert!(
        results[0].1 > 0.5,
        "Query in opposite direction should have high distance"
    );
}

#[test]
fn multiple_queries_returns_results() {
    let dim = 32;
    let mut hnsw = HNSWIndex::new(dim, 32, 32).expect("Failed to create");

    // Use normalized distinct vectors
    let vectors: Vec<Vec<f32>> = (0..50)
        .map(|i| {
            let angle = i as f32 * 0.2;
            let mut v = vec![0.0; dim];
            v[0] = angle.cos();
            v[1] = angle.sin();
            // Add small variation in other dimensions
            for d in 2..dim {
                v[d] = (d as f32 * 0.01) * (i as f32 * 0.1).sin();
            }
            normalize(&v)
        })
        .collect();

    for (i, v) in vectors.iter().enumerate() {
        hnsw.add(i as u32, v.clone()).expect("Failed to add");
    }
    hnsw.build().expect("Failed to build");

    // Run queries and verify we get results
    for i in 0..10 {
        let results = hnsw.search(&vectors[i], 5, 100).expect("Search failed");
        assert_eq!(results.len(), 5, "Should return 5 results for query {}", i);

        // Results should be sorted by distance
        for j in 1..results.len() {
            assert!(
                results[j].1 >= results[j - 1].1 - 1e-5,
                "Results not sorted at query {}: {} > {}",
                i,
                results[j - 1].1,
                results[j].1
            );
        }
    }
}

// =============================================================================
// Parameter edge cases
// =============================================================================

#[test]
fn small_ef_search() {
    let dim = 32;
    let mut hnsw = HNSWIndex::new(dim, 16, 16).expect("Failed to create");

    let vectors: Vec<Vec<f32>> = (0..50)
        .map(|i| normalize(&vec![i as f32 * 0.1; dim]))
        .collect();

    for (i, v) in vectors.iter().enumerate() {
        hnsw.add(i as u32, v.clone()).expect("Failed to add");
    }
    hnsw.build().expect("Failed to build");

    // Very small ef_search
    let results = hnsw.search(&vectors[25], 5, 5).expect("Search failed");

    // Should still return 5 results
    assert_eq!(results.len(), 5);
}

#[test]
fn large_ef_search() {
    let dim = 32;
    let n = 100;
    let mut hnsw = HNSWIndex::new(dim, 16, 16).expect("Failed to create");

    let vectors: Vec<Vec<f32>> = (0..n)
        .map(|i| normalize(&vec![i as f32 * 0.1; dim]))
        .collect();

    for (i, v) in vectors.iter().enumerate() {
        hnsw.add(i as u32, v.clone()).expect("Failed to add");
    }
    hnsw.build().expect("Failed to build");

    // ef_search larger than index size
    let results = hnsw.search(&vectors[50], 10, 500).expect("Search failed");
    assert_eq!(results.len(), 10);
}

#[test]
fn k_equals_n() {
    let dim = 32;
    let n = 50;
    let mut hnsw = HNSWIndex::new(dim, 16, 16).expect("Failed to create");

    // Use distinct vectors
    let vectors: Vec<Vec<f32>> = (0..n)
        .map(|i| {
            let mut v = vec![0.0; dim];
            v[i % dim] = 1.0;
            v[(i + 7) % dim] = 0.5;
            normalize(&v)
        })
        .collect();

    for (i, v) in vectors.iter().enumerate() {
        hnsw.add(i as u32, v.clone()).expect("Failed to add");
    }
    hnsw.build().expect("Failed to build");

    // Request all vectors with high ef
    let results = hnsw.search(&vectors[0], n, 200).expect("Search failed");

    // Should return at least most vectors (HNSW may miss some with low connectivity)
    assert!(
        results.len() >= n - 5,
        "Should return most vectors, got {}/{}",
        results.len(),
        n
    );
}

// =============================================================================
// Stability tests
// =============================================================================

#[test]
fn deterministic_single_query() {
    let dim = 32;
    let mut hnsw = HNSWIndex::new(dim, 16, 16).expect("Failed to create");

    // Use distinct vectors
    let vectors: Vec<Vec<f32>> = (0..50)
        .map(|i| {
            let mut v = vec![0.0; dim];
            v[i % dim] = 1.0;
            v[(i * 3) % dim] = 0.3;
            normalize(&v)
        })
        .collect();

    for (i, v) in vectors.iter().enumerate() {
        hnsw.add(i as u32, v.clone()).expect("Failed to add");
    }
    hnsw.build().expect("Failed to build");

    let query = &vectors[25];

    // Same query should give same results
    let results1 = hnsw.search(query, 10, 50).expect("Search failed");
    let results2 = hnsw.search(query, 10, 50).expect("Search failed");
    let results3 = hnsw.search(query, 10, 50).expect("Search failed");

    // Same indices (order might vary for tied distances)
    let ids1: HashSet<u32> = results1.iter().map(|(i, _)| *i).collect();
    let ids2: HashSet<u32> = results2.iter().map(|(i, _)| *i).collect();
    let ids3: HashSet<u32> = results3.iter().map(|(i, _)| *i).collect();

    assert_eq!(ids1, ids2, "Same query should find same vectors");
    assert_eq!(ids2, ids3, "Same query should find same vectors");
}