lattice-embed 0.13.0

SIMD-accelerated vector operations and embedding generation
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
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
use super::*;
use crate::error::EmbedError;
use crate::service::{
    MAX_TEXT_BYTES, NativeEmbeddingService, reset_validate_texts_calls, validate_texts_calls,
};
use std::sync::atomic::{AtomicBool, AtomicUsize, Ordering};
use std::sync::{Arc, Mutex};

#[derive(Default)]
struct ProbeService {
    calls: AtomicUsize,
    requests: Mutex<Vec<Vec<String>>>,
    fail: bool,
}

impl ProbeService {
    fn failing() -> Self {
        Self {
            fail: true,
            ..Self::default()
        }
    }

    fn calls(&self) -> usize {
        self.calls.load(Ordering::Relaxed)
    }

    fn requests(&self) -> Vec<Vec<String>> {
        self.requests.lock().unwrap().clone()
    }

    fn record(&self, texts: &[String]) -> Result<Vec<Vec<f32>>> {
        self.calls.fetch_add(1, Ordering::Relaxed);
        self.requests.lock().unwrap().push(texts.to_vec());
        if self.fail {
            return Err(EmbedError::InferenceFailed("probe failure".into()));
        }
        Ok(texts.iter().map(|text| vec![text.len() as f32]).collect())
    }
}

#[async_trait]
impl EmbeddingService for ProbeService {
    async fn embed(&self, texts: &[String], _model: EmbeddingModel) -> Result<Vec<Vec<f32>>> {
        self.record(texts)
    }

    async fn embed_with_role_prevalidated(
        &self,
        texts: ValidatedTextBatch<'_>,
        model: EmbeddingModel,
        role: EmbeddingRole,
    ) -> Result<Vec<Vec<f32>>> {
        let prepared = texts.to_owned_with_prefix(role.instruction(model));
        self.record(&prepared)
    }

    fn supports_model(&self, _model: EmbeddingModel) -> bool {
        true
    }

    fn name(&self) -> &'static str {
        "cache-probe"
    }
}

#[derive(Default)]
struct BorrowProbe {
    saw_borrowed: AtomicBool,
}

#[async_trait]
impl EmbeddingService for BorrowProbe {
    async fn embed(&self, _texts: &[String], _model: EmbeddingModel) -> Result<Vec<Vec<f32>>> {
        panic!("cache delegation must use the prevalidated hook")
    }

    async fn embed_with_role_prevalidated(
        &self,
        texts: ValidatedTextBatch<'_>,
        _model: EmbeddingModel,
        _role: EmbeddingRole,
    ) -> Result<Vec<Vec<f32>>> {
        self.saw_borrowed
            .store(texts.borrowed().is_some(), Ordering::Relaxed);
        Ok((0..texts.len())
            .map(|index| vec![texts.get(index).len() as f32])
            .collect())
    }

    fn supports_model(&self, _model: EmbeddingModel) -> bool {
        true
    }

    fn name(&self) -> &'static str {
        "borrow-probe"
    }
}

/// Stands in for a native service: every byte of the prepared text is one token and the
/// model keeps the first `COUNT_PROBE_LIMIT`.
const COUNT_PROBE_LIMIT: usize = 8;

#[derive(Default)]
struct CountProbe {
    embed_calls: AtomicUsize,
    count_calls: AtomicUsize,
    drop_a_count: bool,
}

#[async_trait]
impl EmbeddingService for CountProbe {
    async fn embed(&self, texts: &[String], _model: EmbeddingModel) -> Result<Vec<Vec<f32>>> {
        self.embed_calls.fetch_add(1, Ordering::Relaxed);
        Ok(texts.iter().map(|text| vec![text.len() as f32]).collect())
    }

    async fn embed_with_role_prevalidated(
        &self,
        texts: ValidatedTextBatch<'_>,
        model: EmbeddingModel,
        role: EmbeddingRole,
    ) -> Result<Vec<Vec<f32>>> {
        self.embed_calls.fetch_add(1, Ordering::Relaxed);
        let prepared = texts.to_owned_with_prefix(role.instruction(model));
        Ok(prepared
            .iter()
            .map(|text| vec![text.len() as f32])
            .collect())
    }

    async fn count_tokens(
        &self,
        texts: &[String],
        model: EmbeddingModel,
        role: EmbeddingRole,
    ) -> Result<Vec<crate::service::TokenCount>> {
        self.count_calls.fetch_add(1, Ordering::Relaxed);
        let prefix = role.instruction(model).unwrap_or("");
        let mut counts: Vec<_> = texts
            .iter()
            .map(|text| {
                let visible = prefix.len() + text.len();
                crate::service::TokenCount::new(visible, visible.min(COUNT_PROBE_LIMIT))
            })
            .collect();
        if self.drop_a_count {
            counts.pop();
        }
        Ok(counts)
    }

    fn supports_model(&self, _model: EmbeddingModel) -> bool {
        true
    }

    fn name(&self) -> &'static str {
        "count-probe"
    }
}

#[tokio::test(flavor = "current_thread")]
async fn count_tokens_forwards_to_the_wrapped_service_and_counts_the_role_instruction() {
    let inner = Arc::new(CountProbe::default());
    let service = CachedEmbeddingService::new(inner.clone(), 16);
    let model = EmbeddingModel::MultilingualE5Small;
    let texts = vec!["ab".to_string(), "abcdefghij".to_string()];

    let generic = service
        .count_tokens(&texts, model, EmbeddingRole::Generic)
        .await
        .expect("the cached wrapper must not fall through to the default error");
    assert_eq!(
        generic,
        vec![
            crate::service::TokenCount::new(2, 2),
            crate::service::TokenCount::new(10, 8)
        ]
    );
    assert!(!generic[0].truncated() && generic[1].truncated());

    let query = service
        .count_tokens(&texts, model, EmbeddingRole::Query)
        .await
        .expect("count_tokens");
    let prefix = "query: ".len();
    assert_eq!(query[0].before_truncation, 2 + prefix);
    assert_eq!(query[0].embedded, COUNT_PROBE_LIMIT);
    assert_eq!(inner.count_calls.load(Ordering::Relaxed), 2);
    assert_eq!(inner.embed_calls.load(Ordering::Relaxed), 0);
}

#[tokio::test(flavor = "current_thread")]
async fn embed_with_report_matches_embed_with_role_and_keeps_the_vector_cache() {
    let inner = Arc::new(CountProbe::default());
    let service = CachedEmbeddingService::new(inner.clone(), 16);
    let model = EmbeddingModel::MultilingualE5Small;
    let texts = vec!["ab".to_string(), "abcdefghij".to_string()];

    let report = service
        .embed_with_report(&texts, model, EmbeddingRole::Query)
        .await
        .expect("embed_with_report");
    let plain = service
        .embed_with_role(&texts, model, EmbeddingRole::Query)
        .await
        .expect("embed_with_role");
    assert_eq!(report.embeddings, plain);
    assert_eq!(report.token_counts.len(), texts.len());
    assert_eq!(
        report.token_counts[1].before_truncation,
        10 + "query: ".len()
    );
    assert!(report.token_counts[1].truncated());
    // Vectors were computed once and the second call was served from the cache; the
    // counts are not cached, so the wrapped service was asked for them each time it was.
    assert_eq!(inner.embed_calls.load(Ordering::Relaxed), 1);
    assert_eq!(inner.count_calls.load(Ordering::Relaxed), 1);

    let again = service
        .embed_with_report(&texts, model, EmbeddingRole::Query)
        .await
        .expect("embed_with_report again");
    assert_eq!(again.embeddings, plain);
    assert_eq!(again.token_counts, report.token_counts);
    assert_eq!(inner.embed_calls.load(Ordering::Relaxed), 1);
    assert_eq!(inner.count_calls.load(Ordering::Relaxed), 2);
}

#[tokio::test(flavor = "current_thread")]
async fn embed_with_report_rejects_a_count_per_input_mismatch() {
    let inner = Arc::new(CountProbe {
        drop_a_count: true,
        ..CountProbe::default()
    });
    let service = CachedEmbeddingService::new(inner, 16);
    let texts = vec!["ab".to_string(), "cd".to_string()];
    let err = service
        .embed_with_report(
            &texts,
            EmbeddingModel::BgeSmallEnV15,
            EmbeddingRole::Generic,
        )
        .await
        .expect_err("a short count list must not be returned as a report");
    assert!(matches!(err, EmbedError::InferenceFailed(_)), "got: {err}");
}

#[tokio::test(flavor = "current_thread")]
async fn token_methods_over_a_service_without_overrides_stay_unsupported() {
    let service = CachedEmbeddingService::new(Arc::new(ProbeService::default()), 16);
    let texts = vec!["ab".to_string()];
    let model = EmbeddingModel::BgeSmallEnV15;
    let err = service
        .count_tokens(&texts, model, EmbeddingRole::Generic)
        .await
        .expect_err("the wrapped service has no counts to forward");
    assert!(
        matches!(&err, EmbedError::Unsupported(m) if m.contains("cache-probe")),
        "got: {err}"
    );
    let err = service
        .embed_with_report(&texts, model, EmbeddingRole::Generic)
        .await
        .expect_err("no counts, no report");
    assert!(matches!(err, EmbedError::Unsupported(_)), "got: {err}");
}

#[tokio::test(flavor = "current_thread")]
async fn disabled_delegate_validates_once_and_preserves_role_preparation() {
    let inner = Arc::new(ProbeService::default());
    let service = CachedEmbeddingService::new(inner.clone(), 0);
    let texts = vec!["hello".to_string()];
    let model = EmbeddingModel::BgeSmallEnV15;

    reset_validate_texts_calls();
    let result = service.embed_query(&texts, model).await.unwrap();

    assert_eq!(validate_texts_calls(), 1);
    assert_eq!(inner.calls(), 1);
    let requests = inner.requests();
    assert_eq!(requests.len(), 1);
    assert_eq!(
        requests[0][0],
        format!("{}hello", model.query_instruction().unwrap())
    );
    assert_eq!(result, vec![vec![requests[0][0].len() as f32]]);
}

#[tokio::test(flavor = "current_thread")]
async fn miss_validates_once_and_delegates_only_uncached_texts() {
    let inner = Arc::new(ProbeService::default());
    let service = CachedEmbeddingService::new(inner.clone(), 128);
    let model = EmbeddingModel::AllMiniLmL6V2;

    reset_validate_texts_calls();
    service.embed(&["cached".to_string()], model).await.unwrap();
    assert_eq!(validate_texts_calls(), 1);

    let calls_before = inner.calls();
    reset_validate_texts_calls();
    let result = service
        .embed(&["cached".to_string(), "missing".to_string()], model)
        .await
        .unwrap();

    assert_eq!(validate_texts_calls(), 1);
    assert_eq!(inner.calls(), calls_before + 1);
    assert_eq!(
        inner.requests().last().unwrap(),
        &vec!["missing".to_string()]
    );
    assert_eq!(result, vec![vec![6.0], vec![7.0]]);
}

#[tokio::test(flavor = "current_thread")]
async fn miss_delegates_borrowed_text_views_after_one_validation() {
    let inner = Arc::new(BorrowProbe::default());
    let service = CachedEmbeddingService::new(inner.clone(), 128);

    reset_validate_texts_calls();
    let result = service
        .embed(&["uncached".to_string()], EmbeddingModel::AllMiniLmL6V2)
        .await
        .unwrap();

    assert_eq!(validate_texts_calls(), 1);
    assert!(inner.saw_borrowed.load(Ordering::Relaxed));
    assert_eq!(result, vec![vec![8.0]]);
}

#[tokio::test(flavor = "current_thread")]
async fn all_hit_validates_once_without_delegating() {
    let inner = Arc::new(ProbeService::default());
    let service = CachedEmbeddingService::new(inner.clone(), 128);
    let texts = vec!["cached".to_string()];
    let model = EmbeddingModel::AllMiniLmL6V2;

    service.embed(&texts, model).await.unwrap();
    let calls_before = inner.calls();

    reset_validate_texts_calls();
    let result = service.embed(&texts, model).await.unwrap();

    assert_eq!(validate_texts_calls(), 1);
    assert_eq!(inner.calls(), calls_before);
    assert_eq!(result, vec![vec![6.0]]);
}

#[tokio::test(flavor = "current_thread")]
async fn invalid_input_validates_once_without_delegating() {
    let inner = Arc::new(ProbeService::default());
    let service = CachedEmbeddingService::new(inner.clone(), 128);
    let texts = vec!["x".repeat(MAX_TEXT_BYTES + 1)];

    reset_validate_texts_calls();
    let error = service
        .embed(&texts, EmbeddingModel::AllMiniLmL6V2)
        .await
        .unwrap_err();

    assert_eq!(validate_texts_calls(), 1);
    assert!(matches!(
        error,
        EmbedError::TextTooLong {
            max: MAX_TEXT_BYTES,
            ..
        }
    ));
    assert_eq!(inner.calls(), 0);
}

#[tokio::test(flavor = "current_thread")]
async fn delegate_error_still_validates_once() {
    let inner = Arc::new(ProbeService::failing());
    let service = CachedEmbeddingService::new(inner.clone(), 0);

    reset_validate_texts_calls();
    let error = service
        .embed(&["valid".to_string()], EmbeddingModel::AllMiniLmL6V2)
        .await
        .unwrap_err();

    assert_eq!(validate_texts_calls(), 1);
    assert_eq!(inner.calls(), 1);
    assert!(matches!(error, EmbedError::InferenceFailed(_)));
}

#[tokio::test(flavor = "current_thread")]
async fn nested_cache_delegate_preserves_single_validation() {
    let probe = Arc::new(ProbeService::default());
    let inner = Arc::new(CachedEmbeddingService::new(probe.clone(), 0));
    let service = CachedEmbeddingService::new(inner, 0);

    reset_validate_texts_calls();
    let result = service
        .embed_passage(&["nested".to_string()], EmbeddingModel::AllMiniLmL6V2)
        .await
        .unwrap();

    assert_eq!(validate_texts_calls(), 1);
    assert_eq!(probe.calls(), 1);
    assert_eq!(result, vec![vec![6.0]]);
}

#[tokio::test(flavor = "current_thread")]
async fn native_delegate_preserves_model_validation_without_rescanning_caller_text() {
    let inner = Arc::new(NativeEmbeddingService::default());
    let service = CachedEmbeddingService::new(inner, 0);

    reset_validate_texts_calls();
    let error = service
        .embed(&["valid".to_string()], EmbeddingModel::BgeBaseEnV15)
        .await
        .unwrap_err();

    assert_eq!(validate_texts_calls(), 1);
    assert!(matches!(error, EmbedError::InvalidInput(_)));
}