lattice-embed 0.16.0

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
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
//! Native, pure-Rust embedding service backed by the lattice model-family crates.
//!
//! Model loading is lazy and cancellation-safe; BERT-family and Qwen models take different
//! loading and batching paths. See `docs/service.md` for lifecycle and persistence details.

use super::{
    EmbeddingRole, EmbeddingService, EmbeddingsWithReport, MAX_TEXT_BYTES, TokenCount,
    ValidatedTextBatch,
};
use crate::error::{EmbedError, Result};
use crate::model::{EmbeddingModel, ModelConfig};
use async_trait::async_trait;
use lattice_model_bert::BertModel;
use lattice_model_embeddinggemma2::model::EmbeddingGemma2Model;
use lattice_model_qwen3_embedding::QwenModel;
use lattice_tokenizer::EmbeddingTokenCounts;
use sha2::{Digest, Sha256};
use std::sync::atomic::{AtomicBool, Ordering};
use std::sync::{Arc, OnceLock};
use tracing::{info, warn};

mod identity;
pub use identity::{EmbeddingBackend, EmbeddingSpaceIdentity, IdentifiedEmbeddings, IdentityMode};

/// Loaded model: BERT-family (encoder), Qwen (decoder) or EmbeddingGemma 2 (bidirectional Gemma encoder).
enum LoadedModel {
    Bert {
        model: Arc<BertModel>,
        identity: std::result::Result<EmbeddingSpaceIdentity, String>,
    },
    Qwen {
        model: Arc<QwenModel>,
        identity: std::result::Result<EmbeddingSpaceIdentity, String>,
    },
    EmbeddingGemma2 {
        model: Arc<EmbeddingGemma2Model>,
        output_dim: Option<usize>,
        identity: std::result::Result<EmbeddingSpaceIdentity, String>,
    },
}

// Wrapped model types provide the required thread-safety — see docs/service.md.

impl LoadedModel {
    fn encode_batch(&self, texts: &[&str]) -> std::result::Result<Vec<Vec<f32>>, String> {
        match self {
            LoadedModel::Bert { model, .. } => model.encode_batch(texts).map_err(|e| e.to_string()),
            // Cache-aware; on CPU the cache misses run as packed forwards.
            LoadedModel::Qwen { model: m, .. } => m.encode_batch(texts).map_err(|e| e.to_string()),
            LoadedModel::EmbeddingGemma2 {
                model, output_dim, ..
            } => {
                let ids = texts
                    .iter()
                    .map(|text| model.tokenize(text))
                    .collect::<std::result::Result<Vec<_>, _>>()
                    .map_err(|e| e.to_string())?;
                model
                    .encode_ids_batch(&ids, output_dim.unwrap_or_else(|| model.dimensions()))
                    .map_err(|e| e.to_string())
            }
        }
    }

    /// Token counts per text, from the tokenizer's own pre-truncation length.
    fn count_batch(&self, texts: &[&str]) -> std::result::Result<Vec<TokenCount>, String> {
        match self {
            LoadedModel::Bert { model, .. } => Ok(texts
                .iter()
                .map(|text| token_count(model.embedding_token_counts(text)))
                .collect()),
            LoadedModel::Qwen { model: m, .. } => Ok(texts
                .iter()
                .map(|text| token_count(m.embedding_token_counts(text)))
                .collect()),
            LoadedModel::EmbeddingGemma2 { model, .. } => texts
                .iter()
                .map(|text| {
                    model
                        .tokenize_with_counts(text)
                        .map(|(_, counts)| token_count(counts))
                        .map_err(|e| e.to_string())
                })
                .collect(),
        }
    }

    /// Embeddings from the same computation as [`LoadedModel::encode_batch`], plus counts.
    ///
    /// EmbeddingGemma 2 tokenizes once and embeds those ids, which is exactly what its
    /// `encode` does. BERT and Qwen tokenize for the counts and again inside their
    /// encode entry points, which take text rather than ids.
    fn encode_batch_with_report(
        &self,
        texts: &[&str],
    ) -> std::result::Result<(Vec<Vec<f32>>, Vec<TokenCount>), String> {
        match self {
            LoadedModel::Bert { .. } | LoadedModel::Qwen { .. } => {
                let counts = self.count_batch(texts)?;
                Ok((self.encode_batch(texts)?, counts))
            }
            LoadedModel::EmbeddingGemma2 {
                model, output_dim, ..
            } => {
                let mut batch = Vec::with_capacity(texts.len());
                let mut counts = Vec::with_capacity(texts.len());
                for text in texts {
                    let (ids, text_counts) = model
                        .tokenize_with_counts(text)
                        .map_err(|e| e.to_string())?;
                    batch.push(ids);
                    counts.push(token_count(text_counts));
                }
                let embeddings = model
                    .encode_ids_batch(&batch, output_dim.unwrap_or_else(|| model.dimensions()))
                    .map_err(|e| e.to_string())?;
                Ok((embeddings, counts))
            }
        }
    }

    fn cache_size(&self) -> usize {
        match self {
            LoadedModel::Qwen { model: m, .. } => m.cache_size(),
            _ => 0,
        }
    }
}

fn token_count(counts: EmbeddingTokenCounts) -> TokenCount {
    TokenCount::new(counts.before_truncation, counts.embedded)
}

/// Borrowed views of the prepared texts: the instruction-prefixed copies when a role added
/// one, otherwise the caller's own text.
fn prepared_views<'a>(
    texts: &'a ValidatedTextBatch<'_>,
    prefixed: Option<&'a [String]>,
) -> Vec<&'a str> {
    match prefixed {
        Some(prefixed) => prefixed.iter().map(String::as_str).collect(),
        None => (0..texts.len()).map(|index| texts.get(index)).collect(),
    }
}

/// **Unstable**: model-loading API still evolving; signature may change as lattice-inference matures.
///
/// Pure-Rust local embedding service for one BERT-family or Qwen model configuration.
///
/// It memoizes the load result independently of cancellation.
/// See [`docs/service.md`](../../docs/service.md#nativeembeddingservice-implementation-notes) for loading and architecture details.
pub struct NativeEmbeddingService {
    model: Arc<OnceLock<std::result::Result<LoadedModel, String>>>,
    model_config: ModelConfig,
    identity_mode: IdentityMode,
    initialization_started: AtomicBool,
    model_directory: Option<std::path::PathBuf>,
}

impl Default for NativeEmbeddingService {
    fn default() -> Self {
        Self::new()
    }
}

const LATTICE_EMBED_DIM: &str = "LATTICE_EMBED_DIM";

fn model_config_from_env(model: EmbeddingModel) -> Result<ModelConfig> {
    let output_dim = match std::env::var(LATTICE_EMBED_DIM) {
        Ok(raw) if raw.trim().is_empty() => None,
        Ok(raw) => {
            let dim = raw.trim().parse::<usize>().map_err(|e| {
                EmbedError::InvalidInput(format!("invalid {LATTICE_EMBED_DIM}={raw:?}: {e}"))
            })?;
            Some(dim)
        }
        Err(std::env::VarError::NotPresent) => None,
        Err(e) => {
            return Err(EmbedError::InvalidInput(format!(
                "invalid {LATTICE_EMBED_DIM}: {e}"
            )));
        }
    };
    ModelConfig::try_new(model, output_dim)
}

impl NativeEmbeddingService {
    /// **Unstable**: constructor signature may change; use `EmbeddingService` trait for stable API.
    pub fn new() -> Self {
        Self {
            model: Arc::new(OnceLock::new()),
            model_config: ModelConfig::new(EmbeddingModel::default()),
            identity_mode: IdentityMode::Full,
            initialization_started: AtomicBool::new(false),
            model_directory: None,
        }
    }

    /// **Unstable**: constructor signature may change; use `EmbeddingService` trait for stable API.
    pub fn with_model(model_type: EmbeddingModel) -> Self {
        Self {
            model: Arc::new(OnceLock::new()),
            model_config: ModelConfig::new(model_type),
            identity_mode: IdentityMode::Full,
            initialization_started: AtomicBool::new(false),
            model_directory: None,
        }
    }

    /// **Unstable**: create with explicit model config (model + optional MRL truncation dim).
    pub fn with_model_config(model_config: ModelConfig) -> Result<Self> {
        model_config.validate()?;
        Ok(Self {
            model: Arc::new(OnceLock::new()),
            model_config,
            identity_mode: IdentityMode::Full,
            initialization_started: AtomicBool::new(false),
            model_directory: None,
        })
    }

    /// **Unstable**: create with model config read from `LATTICE_EMBED_DIM` env var.
    pub fn with_model_from_env(model_type: EmbeddingModel) -> Result<Self> {
        let config = model_config_from_env(model_type)?;
        Ok(Self {
            model: Arc::new(OnceLock::new()),
            model_config: config,
            identity_mode: IdentityMode::Full,
            initialization_started: AtomicBool::new(false),
            model_directory: None,
        })
    }

    /// **Unstable**: persistence API may be moved to a separate manager type.
    pub fn save_cache(&self) -> Result<usize> {
        let Some(Ok(model)) = self.model.get() else {
            return Ok(0);
        };
        match model {
            LoadedModel::Qwen { model: m, .. } => {
                let model_name = self.model_config.model.to_string();
                let config = m.effective_inference_config();
                let path = match qwen_persistent_cache_path(
                    &model_name,
                    m.dimensions(),
                    &config.base_model_rev,
                    &config.tokenizer_rev,
                ) {
                    Ok(path) => path,
                    Err(reason) => {
                        warn!("persistent Qwen cache not saved: {reason}");
                        return Ok(0);
                    }
                };
                m.cache_save(&path)
                    .map_err(|e| EmbedError::InferenceFailed(e.to_string()))
            }
            _ => Ok(0),
        }
    }

    /// **Unstable**: internal diagnostic; may be removed or moved to metrics.
    pub fn cache_size(&self) -> usize {
        self.model
            .get()
            .and_then(|r| r.as_ref().ok())
            .map(LoadedModel::cache_size)
            .unwrap_or(0)
    }

    /// **Unstable**: preload the model without performing an encode pass.
    ///
    /// Returns the same loading errors as the first `embed` call.
    /// See [`docs/service.md`](../../docs/service.md#nativeembeddingservice-implementation-notes) for preload behavior.
    pub async fn ensure_loaded(&self) -> Result<()> {
        self.ensure_model().await.map(|_| ())
    }

    /// Ensures the shared model load has completed.
    ///
    /// See [`docs/service.md`](../../docs/service.md#nativeembeddingservice-implementation-notes) for the cancellation invariant.
    #[expect(
        clippy::expect_used,
        reason = "the OnceCell is initialized by the spawn_blocking call that precedes this read"
    )]
    async fn ensure_model(&self) -> Result<&LoadedModel> {
        // Fast path: already loaded.
        if let Some(result) = self.model.get() {
            return result
                .as_ref()
                .map_err(|e| EmbedError::ModelInitialization(e.clone()));
        }

        // The shared lock outlives a cancelled caller — see docs/service.md.
        self.initialization_started.store(true, Ordering::Release);
        let model_lock = self.model.clone();
        let model_config = self.model_config;
        let identity_mode = self.identity_mode;
        let model_directory = self.model_directory.clone();

        tokio::task::spawn_blocking(move || {
            // Serialize initialization on the blocking thread pool.
            model_lock.get_or_init(|| {
                load_model_sync(model_config, identity_mode, model_directory.as_deref())
            });
        })
        .await
        .map_err(|e| EmbedError::ModelInitialization(e.to_string()))?;

        self.model
            .get()
            .expect("set by spawn_blocking")
            .as_ref()
            .map_err(|e| EmbedError::ModelInitialization(e.clone()))
    }

    /// Encodes text that has already been prepared, meaning any retrieval
    /// instruction is applied and the caller-facing contract is checked.
    ///
    /// The length bound here is a memory backstop, not the published cap: it
    /// admits the caller's allowance plus whatever instruction this model
    /// prepends, which is zero bytes for symmetric models.
    async fn encode_prepared(
        &self,
        texts: &[&str],
        model: EmbeddingModel,
    ) -> Result<Vec<Vec<f32>>> {
        if model != self.model_config.model {
            return Err(EmbedError::InvalidInput(format!(
                "requested model {:?} but this service is loaded with {:?}",
                model, self.model_config.model
            )));
        }
        super::validate_texts_bounded(
            texts,
            MAX_TEXT_BYTES.saturating_add(model.max_instruction_bytes()),
        )?;

        let loaded = self.ensure_model().await?;
        loaded
            .encode_batch(texts)
            .map_err(EmbedError::InferenceFailed)
    }

    /// The checks `encode_prepared` makes before it touches the model, for the count and
    /// report paths.
    fn check_prepared(&self, texts: &[&str], model: EmbeddingModel) -> Result<()> {
        if model != self.model_config.model {
            return Err(EmbedError::InvalidInput(format!(
                "requested model {:?} but this service is loaded with {:?}",
                model, self.model_config.model
            )));
        }
        super::validate_texts_bounded(
            texts,
            MAX_TEXT_BYTES.saturating_add(model.max_instruction_bytes()),
        )
    }

    async fn count_prevalidated_with_role(
        &self,
        texts: ValidatedTextBatch<'_>,
        model: EmbeddingModel,
        role: EmbeddingRole,
    ) -> Result<Vec<TokenCount>> {
        let prefixed = role
            .instruction(model)
            .map(|prefix| texts.to_owned_with_prefix(Some(prefix)));
        let views = prepared_views(&texts, prefixed.as_deref());
        self.check_prepared(&views, model)?;
        let loaded = self.ensure_model().await?;
        loaded
            .count_batch(&views)
            .map_err(EmbedError::InferenceFailed)
    }

    async fn report_prevalidated_with_role(
        &self,
        texts: ValidatedTextBatch<'_>,
        model: EmbeddingModel,
        role: EmbeddingRole,
    ) -> Result<EmbeddingsWithReport> {
        let prefixed = role
            .instruction(model)
            .map(|prefix| texts.to_owned_with_prefix(Some(prefix)));
        let views = prepared_views(&texts, prefixed.as_deref());
        self.check_prepared(&views, model)?;
        let loaded = self.ensure_model().await?;
        let (embeddings, token_counts) = loaded
            .encode_batch_with_report(&views)
            .map_err(EmbedError::InferenceFailed)?;
        Ok(EmbeddingsWithReport::new(embeddings, token_counts))
    }

    async fn encode_prevalidated_with_role(
        &self,
        texts: ValidatedTextBatch<'_>,
        model: EmbeddingModel,
        role: EmbeddingRole,
    ) -> Result<Vec<Vec<f32>>> {
        let prefix = role.instruction(model);
        if prefix.is_none()
            && let Some(borrowed) = texts.borrowed()
        {
            return self.encode_prepared(borrowed, model).await;
        }

        if prefix.is_none() {
            let borrowed = (0..texts.len())
                .map(|index| texts.get(index))
                .collect::<Vec<_>>();
            return self.encode_prepared(&borrowed, model).await;
        }

        let prepared = texts.to_owned_with_prefix(prefix);
        let borrowed = prepared.iter().map(String::as_str).collect::<Vec<_>>();
        self.encode_prepared(&borrowed, model).await
    }
}

/// Synchronous model loading (runs on blocking thread pool).
fn load_model_sync(
    model_config: ModelConfig,
    mode: IdentityMode,
    directory: Option<&std::path::Path>,
) -> std::result::Result<LoadedModel, String> {
    load_model_sync_with_bert_acquisition(
        model_config,
        mode,
        directory,
        |model_name| {
            if identity::identity_family(model_config.model) == identity::IdentityFamily::BertSmall
            {
                let mode = identity::weight_mode(mode);
                match directory {
                    Some(dir) => BertModel::from_directory_with_digest_mode(dir, mode, &mut |_| {}),
                    None => BertModel::from_pretrained_with_digest_mode(model_name, mode),
                }
                .map_err(|e| e.to_string())
            } else if identity::identity_family(model_config.model)
                == identity::IdentityFamily::BertFull
            {
                match (mode, directory) {
                    (IdentityMode::Full, Some(dir)) => BertModel::from_directory_with_digest_mode(
                        dir,
                        identity::weight_mode(mode),
                        &mut |_| {},
                    ),
                    (IdentityMode::Full, None) => BertModel::from_pretrained_with_digest_mode(
                        model_name,
                        identity::weight_mode(mode),
                    ),
                    (IdentityMode::Header, Some(dir)) => BertModel::from_directory(dir),
                    (IdentityMode::Header, None) => BertModel::from_pretrained(model_name),
                }
                .map_err(|e| e.to_string())
            } else {
                BertModel::from_pretrained(model_name).map_err(|e| e.to_string())
            }
        },
        identity::identify,
    )
}

fn load_model_sync_with_bert_acquisition(
    model_config: ModelConfig,
    mode: IdentityMode,
    directory: Option<&std::path::Path>,
    acquire_bert: impl FnOnce(&str) -> std::result::Result<BertModel, String>,
    identify: impl FnOnce(
        &BertModel,
        IdentityMode,
    ) -> std::result::Result<EmbeddingSpaceIdentity, String>,
) -> std::result::Result<LoadedModel, String> {
    match model_config.model {
        EmbeddingModel::BgeSmallEnV15
        | EmbeddingModel::BgeBaseEnV15
        | EmbeddingModel::BgeLargeEnV15
        | EmbeddingModel::MultilingualE5Small
        | EmbeddingModel::MultilingualE5Base
        | EmbeddingModel::AllMiniLmL6V2
        | EmbeddingModel::ParaphraseMultilingualMiniLmL12V2 => {
            let model_name = match model_config.model {
                EmbeddingModel::BgeSmallEnV15 => "bge-small-en-v1.5",
                EmbeddingModel::BgeBaseEnV15 => "bge-base-en-v1.5",
                EmbeddingModel::BgeLargeEnV15 => "bge-large-en-v1.5",
                EmbeddingModel::MultilingualE5Small => "multilingual-e5-small",
                EmbeddingModel::MultilingualE5Base => "multilingual-e5-base",
                EmbeddingModel::AllMiniLmL6V2 => "all-minilm-l6-v2",
                EmbeddingModel::ParaphraseMultilingualMiniLmL12V2 => {
                    "paraphrase-multilingual-minilm-l12-v2"
                }
                _ => unreachable!(),
            };
            info!(model = model_name, "loading native BERT embedding model");
            let mut bert = acquire_bert(model_name)?;
            // Route each model family through its correct pooling strategy.
            // BGE uses CLS pooling; E5 and MiniLM use mean pooling.
            if let Some(pooling) = model_config.model.bert_pooling() {
                bert.set_pooling(pooling);
            }
            let identity = match model_config.model {
                EmbeddingModel::BgeSmallEnV15 => identify(&bert, mode),
                EmbeddingModel::BgeBaseEnV15 | EmbeddingModel::BgeLargeEnV15 => {
                    identity::identify_bge_variant(&bert, mode, model_config.model)
                }
                EmbeddingModel::MultilingualE5Small | EmbeddingModel::MultilingualE5Base => {
                    identity::identify_e5(&bert, mode, model_config.model)
                }
                EmbeddingModel::AllMiniLmL6V2
                | EmbeddingModel::ParaphraseMultilingualMiniLmL12V2 => {
                    identity::identify_minilm(&bert, mode, model_config.model)
                }
                _ => Ok(EmbeddingSpaceIdentity::Unavailable),
            };
            Ok(LoadedModel::Bert {
                model: Arc::new(bert),
                identity,
            })
        }
        EmbeddingModel::Qwen3Embedding0_6B | EmbeddingModel::Qwen3Embedding4B => {
            load_qwen_model(model_config, mode, directory)
        }
        EmbeddingModel::EmbeddingGemma2 => {
            load_embeddinggemma2_model(model_config, mode, directory)
        }
        other => Err(format!("unsupported model: {other:?}")),
    }
}

fn load_qwen_model(
    model_config: ModelConfig,
    mode: IdentityMode,
    directory: Option<&std::path::Path>,
) -> std::result::Result<LoadedModel, String> {
    model_config.validate().map_err(|e| e.to_string())?;
    let model_type = model_config.model;
    let model_name = model_type.to_string();
    info!(
        model = %model_name,
        output_dim = ?model_config.output_dim,
        "loading Qwen embedding model"
    );
    let (model_dir, env_override) = match directory {
        Some(directory) => (directory.to_path_buf(), false),
        None => qwen_model_dir(model_type).map_err(|e| e.to_string())?,
    };
    let mut model = if mode == IdentityMode::Full {
        QwenModel::from_directory_with_full_report(&model_dir)
    } else {
        QwenModel::from_directory(&model_dir)
    }
    .map_err(|e| e.to_string())?;
    model.set_output_dim(model_config.output_dim);
    let config = model.effective_inference_config();
    match qwen_persistent_cache_path(
        &model_name,
        model.dimensions(),
        &config.base_model_rev,
        &config.tokenizer_rev,
    ) {
        Ok(cache_path) => {
            let managed = allow_legacy_cache(directory, env_override);
            let legacy =
                managed.then(|| legacy_embedding_cache_path(&model_name, model.dimensions()));
            match restore_embedding_cache(&cache_path, legacy.as_deref(), |path| {
                model.cache_load(path)
            }) {
                Ok((n, source)) => {
                    if n > 0 {
                        info!(entries = n, path = %source.display(), "loaded embedding cache");
                    }
                    if source != cache_path
                        && let Err(e) = model.cache_save(&cache_path)
                    {
                        warn!(path = %cache_path.display(), error = %e, "embedding cache migration could not be saved");
                    }
                }
                Err((path, e)) => {
                    warn!(path = %path.display(), error = %e, "embedding cache failed integrity check, ignoring (will regenerate on next save)");
                }
            }
        }
        Err(reason) => warn!("persistent Qwen cache not restored: {reason}"),
    }
    let identity = identity::identify_qwen(&model, model_type, mode);
    Ok(LoadedModel::Qwen {
        model: Arc::new(model),
        identity,
    })
}

fn load_embeddinggemma2_model(
    model_config: ModelConfig,
    mode: IdentityMode,
    directory: Option<&std::path::Path>,
) -> std::result::Result<LoadedModel, String> {
    model_config.validate().map_err(|e| e.to_string())?;
    info!(
        output_dim = ?model_config.output_dim,
        "loading EmbeddingGemma 2 embedding model"
    );
    let model_dir = match directory {
        Some(directory) => directory.to_path_buf(),
        None => embeddinggemma2_model_dir().map_err(|e| e.to_string())?,
    };
    let model = if mode == IdentityMode::Full {
        EmbeddingGemma2Model::from_model_dir_with_full_report(&model_dir, &mut |_| {})
    } else {
        EmbeddingGemma2Model::from_model_dir(&model_dir)
    }
    .map_err(|e| e.to_string())?;
    let identity = identity::identify_gemma(&model, model_config.output_dim, mode);
    Ok(LoadedModel::EmbeddingGemma2 {
        model: Arc::new(model),
        output_dim: model_config.output_dim,
        identity,
    })
}

/// Locate the EmbeddingGemma 2 model directory: `LATTICE_EMBEDDINGGEMMA2_MODEL_DIR`, else
/// `~/.lattice/models/embeddinggemma-2`.
fn embeddinggemma2_model_dir() -> Result<std::path::PathBuf> {
    if let Ok(dir) = std::env::var("LATTICE_EMBEDDINGGEMMA2_MODEL_DIR") {
        return Ok(std::path::PathBuf::from(dir));
    }
    let home = std::env::var("HOME")
        .map_err(|_| EmbedError::ModelInitialization("HOME not set".into()))?;
    let dir = std::path::PathBuf::from(home)
        .join(".lattice")
        .join("models")
        .join("embeddinggemma-2");
    if dir.join("model.safetensors").exists() {
        Ok(dir)
    } else {
        Err(EmbedError::ModelInitialization(format!(
            "EmbeddingGemma 2 model not found at {dir}. Download it with:\n  huggingface-cli download {repo} --local-dir {dir}",
            dir = dir.display(),
            repo = EmbeddingModel::EmbeddingGemma2.model_id()
        )))
    }
}

fn allow_legacy_cache(directory: Option<&std::path::Path>, env_override: bool) -> bool {
    directory.is_none() && !env_override
}

fn cache_revisions_known(base: &str, tokenizer: &str) -> bool {
    [base, tokenizer]
        .iter()
        .all(|revision| !revision.is_empty() && *revision != "none")
}

fn qwen_persistent_cache_path(
    model: &str,
    dim: usize,
    base_revision: &str,
    tokenizer_revision: &str,
) -> std::result::Result<std::path::PathBuf, &'static str> {
    if !cache_revisions_known(base_revision, tokenizer_revision) {
        return Err("model or tokenizer revision is unknown");
    }
    Ok(embedding_cache_path(model, dim, base_revision))
}

#[cfg(test)]
thread_local! {
    static TEST_CACHE_DIRECTORY: std::cell::RefCell<Option<std::path::PathBuf>> = const { std::cell::RefCell::new(None) };
}

fn legacy_embedding_cache_path(model: &str, dim: usize) -> std::path::PathBuf {
    #[cfg(test)]
    if let Some(directory) = TEST_CACHE_DIRECTORY.with(|value| value.borrow().clone()) {
        return directory.join(format!("embed_{model}_{dim}d.bin"));
    }
    let home = std::env::var("HOME").unwrap_or_else(|_| "/tmp".into());
    std::path::PathBuf::from(home)
        .join(".lattice")
        .join("cache")
        .join(format!("embed_{model}_{dim}d.bin"))
}

fn embedding_cache_path(model: &str, dim: usize, base_revision: &str) -> std::path::PathBuf {
    let digest = Sha256::digest(base_revision.as_bytes());
    legacy_embedding_cache_path(model, dim)
        .with_file_name(format!("embed_{model}_{dim}d_{digest:x}.bin"))
}

fn restore_embedding_cache<F>(
    path: &std::path::Path,
    legacy: Option<&std::path::Path>,
    mut load: F,
) -> std::result::Result<
    (usize, std::path::PathBuf),
    (std::path::PathBuf, lattice_inference_core::InferenceError),
>
where
    F: FnMut(
        &std::path::Path,
    ) -> std::result::Result<usize, lattice_inference_core::InferenceError>,
{
    let count = load(path).map_err(|error| (path.to_path_buf(), error))?;
    if count == 0
        && matches!(path.try_exists(), Ok(false))
        && let Some(legacy) = legacy
        && matches!(legacy.try_exists(), Ok(true))
    {
        return load(legacy)
            .map(|count| (count, legacy.to_path_buf()))
            .map_err(|error| (legacy.to_path_buf(), error));
    }
    Ok((count, path.to_path_buf()))
}

/// Locate Qwen3-Embedding model directory for the given model variant, and
/// report whether `LATTICE_QWEN_MODEL_DIR` chose it.
fn qwen_model_dir(model_type: EmbeddingModel) -> Result<(std::path::PathBuf, bool)> {
    // Check env override first — applies to whichever Qwen model is loaded.
    if let Ok(dir) = std::env::var("LATTICE_QWEN_MODEL_DIR") {
        return Ok((std::path::PathBuf::from(dir), true));
    }

    let slug = match model_type {
        EmbeddingModel::Qwen3Embedding0_6B => "qwen3-embedding-0.6b",
        EmbeddingModel::Qwen3Embedding4B => "qwen3-embedding-4b",
        other => {
            return Err(EmbedError::ModelInitialization(format!(
                "not a Qwen model: {other}"
            )));
        }
    };

    let home = std::env::var("HOME")
        .map_err(|_| EmbedError::ModelInitialization("HOME not set".into()))?;
    let dir = std::path::PathBuf::from(home)
        .join(".lattice")
        .join("models")
        .join(slug);

    if dir.join("model.safetensors").exists() || dir.join("model.safetensors.index.json").exists() {
        Ok((dir, false))
    } else {
        Err(EmbedError::ModelInitialization(format!(
            "Qwen3 model not found at {dir}. Download it with:\n  huggingface-cli download {repo} --local-dir {dir}",
            dir = dir.display(),
            repo = model_type.model_id()
        )))
    }
}

#[async_trait]
impl EmbeddingService for NativeEmbeddingService {
    async fn embed(&self, texts: &[String], model: EmbeddingModel) -> Result<Vec<Vec<f32>>> {
        let texts = ValidatedTextBatch::new(texts)?;
        self.encode_prevalidated_with_role(texts, model, EmbeddingRole::Generic)
            .await
    }

    async fn embed_with_role(
        &self,
        texts: &[String],
        model: EmbeddingModel,
        role: EmbeddingRole,
    ) -> Result<Vec<Vec<f32>>> {
        let texts = ValidatedTextBatch::new(texts)?;
        self.encode_prevalidated_with_role(texts, model, role).await
    }

    async fn embed_with_role_prevalidated(
        &self,
        texts: ValidatedTextBatch<'_>,
        model: EmbeddingModel,
        role: EmbeddingRole,
    ) -> Result<Vec<Vec<f32>>> {
        self.encode_prevalidated_with_role(texts, model, role).await
    }

    async fn count_tokens(
        &self,
        texts: &[String],
        model: EmbeddingModel,
        role: EmbeddingRole,
    ) -> Result<Vec<TokenCount>> {
        let texts = ValidatedTextBatch::new(texts)?;
        self.count_prevalidated_with_role(texts, model, role).await
    }

    async fn embed_with_report(
        &self,
        texts: &[String],
        model: EmbeddingModel,
        role: EmbeddingRole,
    ) -> Result<EmbeddingsWithReport> {
        let texts = ValidatedTextBatch::new(texts)?;
        self.report_prevalidated_with_role(texts, model, role).await
    }

    fn model_config(&self, model: EmbeddingModel) -> ModelConfig {
        if model == self.model_config.model {
            self.model_config
        } else {
            ModelConfig::new(model)
        }
    }

    fn supports_model(&self, model: EmbeddingModel) -> bool {
        model == self.model_config.model
    }

    fn name(&self) -> &'static str {
        "native-bert"
    }
}

#[cfg(test)]
#[path = "native_pooling_tests.rs"]
mod pooling_tests;

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

    #[test]
    fn test_cache_path_contains_dim_in_filename() {
        let path = embedding_cache_path("qwen3-embedding-4b", 1024, "revision-4b");
        let filename = path.file_name().unwrap().to_str().unwrap();
        assert!(filename.starts_with("embed_qwen3-embedding-4b_1024d_"));
        assert!(filename.ends_with(".bin"));
        let odd_revision = embedding_cache_path("qwen3-embedding-4b", 1024, "../odd/revision");
        assert_eq!(odd_revision.parent(), path.parent());
    }

    #[test]
    fn test_cache_path_different_dims_produce_different_paths() {
        let path_1024 = embedding_cache_path("qwen3-embedding-4b", 1024, "revision-4b");
        let path_2560 = embedding_cache_path("qwen3-embedding-4b", 2560, "revision-4b");
        assert_ne!(path_1024, path_2560);
        assert!(path_1024.to_string_lossy().contains("1024d"));
        assert!(path_2560.to_string_lossy().contains("2560d"));
    }

    #[test]
    fn test_cache_path_model_slug_differentiates_variants() {
        let path_4b = embedding_cache_path("qwen3-embedding-4b", 2560, "revision-4b");
        let path_06b = embedding_cache_path("qwen3-embedding-0.6b", 1024, "revision-06b");
        assert_ne!(path_4b, path_06b);
        assert!(path_4b.to_string_lossy().contains("qwen3-embedding-4b"));
        assert!(path_06b.to_string_lossy().contains("qwen3-embedding-0.6b"));
    }

    #[test]
    fn test_cache_path_same_model_same_dim_same_path() {
        let p1 = embedding_cache_path("qwen3-embedding-4b", 1024, "revision-4b");
        let p2 = embedding_cache_path("qwen3-embedding-4b", 1024, "revision-4b");
        assert_eq!(p1, p2);
    }

    #[test]
    fn persistent_cache_requires_known_revisions_and_managed_legacy() {
        assert!(cache_revisions_known("base", "tokenizer"));
        for (base, tokenizer) in [
            ("none", "tokenizer"),
            ("", "tokenizer"),
            ("base", "none"),
            ("base", ""),
        ] {
            assert!(!cache_revisions_known(base, tokenizer));
            assert_eq!(
                qwen_persistent_cache_path("fixture", 32, base, tokenizer).unwrap_err(),
                "model or tokenizer revision is unknown"
            );
        }
        assert!(allow_legacy_cache(None, false));
        assert!(!allow_legacy_cache(None, true));
        assert!(!allow_legacy_cache(
            Some(std::path::Path::new("/explicit")),
            false
        ));
    }

    #[test]
    fn legacy_cache_fallback_requires_absent_new_file_and_propagates_refusals() {
        let dir = tempfile::tempdir().unwrap();
        let new_path = dir.path().join("new.bin");
        let legacy = dir.path().join("old.bin");
        std::fs::write(&legacy, b"fixture").unwrap();
        let mut calls = Vec::new();
        let (count, source) = restore_embedding_cache(&new_path, Some(&legacy), |path| {
            calls.push(path.to_path_buf());
            Ok(if path == legacy.as_path() { 3 } else { 0 })
        })
        .unwrap();
        assert_eq!(count, 3);
        assert_eq!(source, legacy);
        assert_eq!(calls, [new_path.clone(), legacy.clone()]);
        std::fs::write(&new_path, b"fixture").unwrap();
        calls.clear();
        let (count, source) = restore_embedding_cache(&new_path, Some(&legacy), |path| {
            calls.push(path.to_path_buf());
            Ok(0)
        })
        .unwrap();
        assert_eq!(count, 0);
        assert_eq!(source, new_path);
        assert_eq!(calls, std::slice::from_ref(&new_path));
        calls.clear();
        let (failed_path, _) = restore_embedding_cache(&new_path, Some(&legacy), |path| {
            calls.push(path.to_path_buf());
            Err(lattice_inference_core::InferenceError::Inference(
                "refused".into(),
            ))
        })
        .unwrap_err();
        assert_eq!(failed_path, new_path);
        assert_eq!(calls, [new_path]);
    }
}