vicinity 0.7.0

Approximate nearest-neighbor search
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

#![cfg(all(feature = "hnsw", feature = "vamana"))]
#![allow(clippy::unwrap_used, clippy::expect_used, clippy::needless_update)]

//! Property tests for Wolverine++ deletion and seeded RNG reproducibility.
//!
//! Properties tested:
//! 1. delete+search never returns deleted IDs (single and batch)
//! 2. Seeded builds are deterministic (same seed = same graph)
//! 3. Graph integrity after deletion: no edges to deleted nodes
//! 4. Recall doesn't collapse after moderate deletion (< 50%)

#![allow(dead_code, unused_imports)]

#[path = "common/mod.rs"]
mod common;
use common::*;

use proptest::prelude::*;
use std::collections::HashSet;

/// Build a seeded HNSW index from normalized random vectors.
#[cfg(feature = "hnsw")]
fn build_hnsw(
    n: usize,
    dim: usize,
    m: usize,
    seed: u64,
) -> (vicinity::hnsw::HNSWIndex, Vec<Vec<f32>>) {
    let params = vicinity::hnsw::HNSWParams {
        m,
        m_max: m * 2,
        ef_construction: 100,
        seed: Some(seed),
        ..vicinity::hnsw::HNSWParams::default()
    };
    let mut idx = vicinity::hnsw::HNSWIndex::with_params(dim, params).unwrap();

    let vectors: Vec<Vec<f32>> = random_vectors(n, dim, seed)
        .into_iter()
        .map(|v| normalize(&v))
        .collect();

    for (i, v) in vectors.iter().enumerate() {
        idx.add(i as u32, v.clone()).unwrap();
    }
    idx.build().unwrap();
    (idx, vectors)
}

// ---------------------------------------------------------------------------
// Property 1: Deleted IDs never appear in search results
// ---------------------------------------------------------------------------

#[cfg(feature = "hnsw")]
proptest! {
    #![proptest_config(ProptestConfig::with_cases(20))]

    #[test]
    fn prop_deleted_ids_absent_from_results(
        n in 50usize..200,
        dim in prop::sample::select(vec![8, 16, 32]),
        num_delete in 1usize..30,
        seed in 1u64..10000,
    ) {
        let (mut idx, _) = build_hnsw(n, dim, 8, seed);
        let num_delete = num_delete.min(n / 2);

        // Delete first `num_delete` IDs
        let deleted_ids: Vec<u32> = (0..num_delete as u32).collect();
        for &id in &deleted_ids {
            idx.delete_with_repair(id).unwrap();
        }

        let deleted_set: HashSet<u32> = deleted_ids.iter().copied().collect();

        // Search with multiple queries
        let queries = random_vectors(5, dim, seed.wrapping_add(9999));
        for q in &queries {
            let q_norm = normalize(q);
            let results = idx.search(&q_norm, 10, 64).unwrap();
            for (id, _) in &results {
                prop_assert!(
                    !deleted_set.contains(id),
                    "deleted id {} found in results after delete_with_repair",
                    id
                );
            }
        }
    }
}

// ---------------------------------------------------------------------------
// Property 2: Batch deletion -- same invariant
// ---------------------------------------------------------------------------

#[cfg(feature = "hnsw")]
proptest! {
    #![proptest_config(ProptestConfig::with_cases(15))]

    #[test]
    fn prop_batch_deleted_ids_absent(
        n in 80usize..200,
        dim in prop::sample::select(vec![8, 16, 32]),
        num_delete in 5usize..40,
        seed in 1u64..10000,
    ) {
        let (mut idx, _) = build_hnsw(n, dim, 8, seed);
        let num_delete = num_delete.min(n / 2);

        let delete_ids: Vec<u32> = (0..num_delete as u32).collect();
        idx.delete_batch_with_repair(&delete_ids).unwrap();

        let deleted_set: HashSet<u32> = delete_ids.iter().copied().collect();

        let queries = random_vectors(5, dim, seed.wrapping_add(7777));
        for q in &queries {
            let q_norm = normalize(q);
            let results = idx.search(&q_norm, 10, 64).unwrap();
            for (id, _) in &results {
                prop_assert!(
                    !deleted_set.contains(id),
                    "deleted id {} in batch results",
                    id
                );
            }
        }
    }
}

// ---------------------------------------------------------------------------
// Property 3: Deleted nodes are truly excluded (high-ef exhaustive check)
// ---------------------------------------------------------------------------

#[cfg(feature = "hnsw")]
proptest! {
    #![proptest_config(ProptestConfig::with_cases(15))]

    #[test]
    fn prop_deleted_exhaustive_check(
        n in 50usize..150,
        dim in prop::sample::select(vec![8, 16, 32]),
        num_delete in 1usize..20,
        seed in 1u64..10000,
    ) {
        let (mut idx, vectors) = build_hnsw(n, dim, 8, seed);
        let num_delete = num_delete.min(n / 3);

        for doc_id in 0..num_delete as u32 {
            idx.delete_with_repair(doc_id).unwrap();
        }

        // Search with very high ef to maximize coverage -- use each deleted
        // vector as query to check that it doesn't find itself
        for doc_id in 0..num_delete as u32 {
            let q = &vectors[doc_id as usize];
            let results = idx.search(q, n, n).unwrap();
            for (id, _) in &results {
                prop_assert!(
                    *id >= num_delete as u32,
                    "deleted id {} found when querying with deleted vector {}",
                    id,
                    doc_id
                );
            }
        }
    }
}

// ---------------------------------------------------------------------------
// Property 4: Seeded HNSW builds are deterministic
// ---------------------------------------------------------------------------

#[cfg(feature = "hnsw")]
proptest! {
    #![proptest_config(ProptestConfig::with_cases(10))]

    #[test]
    fn prop_seeded_hnsw_deterministic(
        n in 20usize..100,
        dim in prop::sample::select(vec![8, 16, 32]),
        seed in 1u64..10000,
    ) {
        let (idx1, _) = build_hnsw(n, dim, 8, seed);
        let (idx2, _) = build_hnsw(n, dim, 8, seed);

        // Same seed -> same search results for multiple queries
        let queries = random_vectors(5, dim, seed + 777);
        for q in &queries {
            let q_norm = normalize(q);
            let r1 = idx1.search(&q_norm, 10, 64).unwrap();
            let r2 = idx2.search(&q_norm, 10, 64).unwrap();
            prop_assert_eq!(r1.len(), r2.len(), "different result count with same seed");
            for (a, b) in r1.iter().zip(r2.iter()) {
                prop_assert_eq!(a.0, b.0, "different result IDs with same seed");
            }
        }
    }
}

// ---------------------------------------------------------------------------
// Property 5: Seeded Vamana builds are deterministic
// ---------------------------------------------------------------------------

#[cfg(feature = "vamana")]
proptest! {
    #![proptest_config(ProptestConfig::with_cases(8))]

    #[test]
    fn prop_seeded_vamana_deterministic(
        n in 20usize..60,
        dim in prop::sample::select(vec![8, 16]),
        seed in 1u64..10000,
    ) {
        let vectors: Vec<Vec<f32>> = random_vectors(n, dim, seed)
            .into_iter()
            .map(|v| normalize(&v))
            .collect();

        let build = |s: u64| {
            let params = vicinity::vamana::VamanaParams {
                max_degree: 16,
                alpha: 1.3,
                ef_construction: 50,
                ef_search: 20,
                seed: Some(s),
                ..vicinity::vamana::VamanaParams::default()
            };
            let mut idx = vicinity::vamana::VamanaIndex::new(dim, params).unwrap();
            for (i, v) in vectors.iter().enumerate() {
                idx.add(i as u32, v.clone()).unwrap();
            }
            idx.build().unwrap();
            idx
        };

        let idx1 = build(seed);
        let idx2 = build(seed);

        // Same query should return same results
        let q = normalize(&random_vectors(1, dim, seed + 999)[0]);
        let r1 = idx1.search(&q, 10, 50).unwrap();
        let r2 = idx2.search(&q, 10, 50).unwrap();

        prop_assert_eq!(r1.len(), r2.len(), "different result count with same seed");
        for (a, b) in r1.iter().zip(r2.iter()) {
            prop_assert_eq!(a.0, b.0, "different result IDs with same seed");
        }
    }
}

// ---------------------------------------------------------------------------
// Property 6: Recall doesn't collapse after moderate deletion
// ---------------------------------------------------------------------------

#[cfg(feature = "hnsw")]
proptest! {
    #![proptest_config(ProptestConfig::with_cases(10))]

    #[test]
    fn prop_recall_survives_deletion(
        n in 100usize..300,
        dim in prop::sample::select(vec![16, 32]),
        delete_frac in 10usize..40, // percent to delete
        seed in 1u64..10000,
    ) {
        let (mut idx, _vectors) = build_hnsw(n, dim, 12, seed);
        let num_delete = (n * delete_frac) / 100;

        for i in 0..num_delete as u32 {
            idx.delete_with_repair(i).unwrap();
        }

        let remaining = n - num_delete;
        let k = 10.min(remaining);
        if k == 0 {
            return Ok(());
        }

        let queries = random_vectors(5, dim, seed + 5555);
        let mut total_results = 0usize;

        for q in &queries {
            let q_norm = normalize(q);
            let results = idx.search(&q_norm, k, 100).unwrap();
            total_results += results.len();

            // All results should be non-deleted
            for (id, _) in &results {
                prop_assert!(*id >= num_delete as u32, "deleted id {} in results", id);
            }
        }

        // Should return k results per query (graph stays connected enough)
        let expected_min = queries.len() * k;
        let recall_fraction = total_results as f64 / expected_min as f64;
        prop_assert!(
            recall_fraction >= 0.3,
            "only {:.0}% of expected results returned after deleting {}% of {} nodes \
             ({} of {} expected)",
            recall_fraction * 100.0,
            delete_frac,
            n,
            total_results,
            expected_min,
        );
    }
}

// ─── SQ8 property tests ─────────────────────────────────────────────────────

// SQ8 quantization preserves ranking: reranked search recall is high.
#[cfg(all(feature = "hnsw", feature = "sq8"))]
proptest! {
    #![proptest_config(ProptestConfig::with_cases(10))]

    #[test]
    fn prop_sq8_reranked_recall(
        n in 100usize..500,
        dim in prop::sample::select(vec![16, 32, 64, 128]),
        seed in 1u64..10000,
    ) {
        use vicinity::distance::DistanceMetric;
        use vicinity::hnsw::{HNSWParams, sq8u::HNSWSq8Index};

        let vecs = random_vectors(n, dim, seed);

        let params = HNSWParams {
            metric: DistanceMetric::L2,
            ..Default::default()
        };
        let mut index = HNSWSq8Index::with_params(dim, params).unwrap();
        for (i, v) in vecs.iter().enumerate() {
            index.add_slice(i as u32, v).unwrap();
        }
        index.build().unwrap();

        let k = 10.min(n);
        let query = &vecs[0];

        // Brute-force ground truth
        let mut gt: Vec<(u32, f32)> = vecs
            .iter()
            .enumerate()
            .map(|(i, v)| (i as u32, vicinity::distance::l2_distance(query, v)))
            .collect();
        gt.sort_by(|a, b| a.1.total_cmp(&b.1));
        let gt_ids: HashSet<u32> = gt.iter().take(k).map(|(id, _)| *id).collect();

        let results = index.search_reranked(query, k, 64, 50).unwrap();
        let result_ids: HashSet<u32> = results.iter().map(|(id, _)| *id).collect();

        let recall = gt_ids.intersection(&result_ids).count() as f64 / k as f64;
        prop_assert!(
            recall >= 0.5,
            "SQ8 reranked recall@{k} = {:.1}%, expected >= 50% (n={n}, dim={dim}, seed={seed})",
            recall * 100.0,
        );
    }
}