qql 0.3.1

QQL runtime and executor for Qdrant REST and gRPC
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
//! Embedding adapters for the runtime.
//!
//! The shared `Embedder` trait and AST resolve live in `qql-embed`.
//! This module re-exports them and provides `HttpEmbedder` (reqwest).

#[cfg(feature = "rest")]
use async_trait::async_trait;
#[cfg(feature = "rest")]
use reqwest::Client;
#[cfg(feature = "rest")]
use serde::{Deserialize, Serialize};

#[cfg(feature = "rest")]
use qql_core::error::QqlError;

// Re-export shared API so existing `qql::embedder::Embedder` paths keep working.
pub use qql_embed::embedder::{Embedder, EmbedderBound, SparseEmbedder};
pub use qql_embed::SparseVector;

#[cfg(feature = "rest")]
#[derive(Debug, Clone, Serialize)]
struct EmbedRequest<'a> {
    model: &'a str,
    input: &'a [String],
}

#[cfg(feature = "rest")]
#[derive(Debug, Clone, Deserialize)]
struct EmbedResponse {
    data: Vec<EmbedData>,
}

/// OpenAI-compatible embedding payload: dense `[f32]` or multi `[[f32],…]`.
#[cfg(feature = "rest")]
#[derive(Debug, Clone, Deserialize)]
#[serde(untagged)]
enum EmbeddingPayload {
    Dense(Vec<f32>),
    Multi(Vec<Vec<f32>>),
}

#[cfg(feature = "rest")]
#[derive(Debug, Clone, Deserialize)]
struct EmbedData {
    index: usize,
    embedding: EmbeddingPayload,
}

/// Options for constructing an [`HttpEmbedder`].
#[cfg(feature = "rest")]
#[derive(Debug, Clone, Default)]
pub struct HttpEmbedderOptions {
    /// OpenAI-compatible embedding endpoint URL (e.g. `https://host/v1/embeddings`).
    pub endpoint: String,
    /// Bearer token sent as `Authorization` when non-empty.
    pub api_key: String,
    /// Default dense model name.
    pub model: String,
    /// Expected dense dimension; must be positive.
    pub dimension: usize,
    /// Optional multi/ColBERT endpoint. Falls back to `endpoint` when empty.
    pub multi_endpoint: Option<String>,
    /// Bearer token for `multi_endpoint`; falls back to `api_key`.
    pub multi_api_key: Option<String>,
    /// Multi/ColBERT model name; falls back to `model` when unset.
    pub multi_model: Option<String>,
    /// Expected per-token dim for multi responses. `0` skips dim checks.
    pub multi_dimension: usize,
    /// Optional image/CLIP vision endpoint. Falls back to `endpoint` when empty.
    pub image_endpoint: Option<String>,
    /// Bearer token for `image_endpoint`; falls back to `api_key`.
    pub image_api_key: Option<String>,
    /// Image/CLIP vision model name; falls back to `model` when unset.
    pub image_model: Option<String>,
    /// Expected dense dim for image responses (CLIP = 512). `0` uses dense dim.
    pub image_dimension: usize,
    /// Cross-encoder pair-rerank endpoint (Cohere-compatible).
    pub rerank_endpoint: Option<String>,
    /// Bearer token for the rerank endpoint; falls back to `api_key`.
    pub rerank_api_key: Option<String>,
    /// Cross-encoder model name (Cohere-style `rerank` API).
    pub rerank_model: Option<String>,
}

/// OpenAI-compatible HTTP embedder (`POST {"model","input":[...]}`).
///
/// Endpoint is **required** — no default URL. Works with OpenAI, Ollama
/// `/v1/embeddings`, Cohere compatibility API, etc. Always batches in one request.
///
/// Multivector: set `multi_*` options (or share the dense endpoint with a model that
/// returns nested `embedding: [[…],…]` arrays). Flat dense arrays on multi requests
/// are rejected.
#[cfg(feature = "rest")]
pub struct HttpEmbedder {
    endpoint: String,
    api_key: String,
    model: String,
    dimension: usize,
    multi_endpoint: Option<String>,
    multi_api_key: Option<String>,
    multi_model: Option<String>,
    multi_dimension: usize,
    image_endpoint: Option<String>,
    image_api_key: Option<String>,
    image_model: Option<String>,
    image_dimension: usize,
    rerank_endpoint: Option<String>,
    rerank_api_key: Option<String>,
    rerank_model: Option<String>,
    client: Client,
}

#[cfg(feature = "rest")]
impl HttpEmbedder {
    /// Build an embedder with only dense settings; see `try_with_options` for
    /// multi, image, and rerank configuration.
    pub fn new(
        endpoint: String,
        api_key: String,
        model: String,
        dimension: usize,
    ) -> Result<Self, QqlError> {
        Self::try_with_options(HttpEmbedderOptions {
            endpoint,
            api_key,
            model,
            dimension,
            ..Default::default()
        })
    }

    /// Validate options and build the embedder; endpoint, model, and positive
    /// dimension are required.
    pub fn try_with_options(opts: HttpEmbedderOptions) -> Result<Self, QqlError> {
        if opts.endpoint.trim().is_empty() {
            return Err(QqlError::execution(
                "QQL-EMBEDDING",
                "embedding endpoint is required",
                None,
            ));
        }
        if opts.model.trim().is_empty() {
            return Err(QqlError::execution(
                "QQL-EMBEDDING",
                "embedding model is required",
                None,
            ));
        }
        if opts.dimension == 0 {
            return Err(QqlError::execution(
                "QQL-EMBEDDING",
                "embedding dimension must be positive",
                None,
            ));
        }

        let client = Client::builder().build().map_err(|e| {
            QqlError::execution(
                "QQL-EMBEDDING",
                format!("failed to create HTTP client: {}", e),
                None,
            )
        })?;

        Ok(HttpEmbedder {
            endpoint: opts.endpoint,
            api_key: opts.api_key,
            model: opts.model,
            dimension: opts.dimension,
            multi_endpoint: opts.multi_endpoint.filter(|s| !s.trim().is_empty()),
            multi_api_key: opts.multi_api_key,
            multi_model: opts.multi_model.filter(|s| !s.trim().is_empty()),
            multi_dimension: opts.multi_dimension,
            image_endpoint: opts.image_endpoint.filter(|s| !s.trim().is_empty()),
            image_api_key: opts.image_api_key,
            image_model: opts.image_model.filter(|s| !s.trim().is_empty()),
            image_dimension: opts.image_dimension,
            rerank_endpoint: opts.rerank_endpoint.filter(|s| !s.trim().is_empty()),
            rerank_api_key: opts.rerank_api_key,
            rerank_model: opts.rerank_model.filter(|s| !s.trim().is_empty()),
            client,
        })
    }

    /// Attach multi/ColBERT settings after construction.
    pub fn with_multi(
        mut self,
        endpoint: Option<String>,
        api_key: Option<String>,
        model: Option<String>,
        dimension: usize,
    ) -> Self {
        self.multi_endpoint = endpoint.filter(|s| !s.trim().is_empty());
        self.multi_api_key = api_key;
        self.multi_model = model.filter(|s| !s.trim().is_empty());
        self.multi_dimension = dimension;
        self
    }

    /// Attach image/CLIP vision settings after construction.
    pub fn with_image(
        mut self,
        endpoint: Option<String>,
        api_key: Option<String>,
        model: Option<String>,
        dimension: usize,
    ) -> Self {
        self.image_endpoint = endpoint.filter(|s| !s.trim().is_empty());
        self.image_api_key = api_key;
        self.image_model = model.filter(|s| !s.trim().is_empty());
        self.image_dimension = dimension;
        self
    }

    /// Whether multi/ColBERT embedding is configured on this embedder.
    pub fn multi_enabled(&self) -> bool {
        self.multi_model.is_some() || self.multi_endpoint.is_some() || self.multi_dimension > 0
    }

    /// Whether image/CLIP vision embedding is configured on this embedder.
    pub fn image_enabled(&self) -> bool {
        self.image_model.is_some() || self.image_endpoint.is_some() || self.image_dimension > 0
    }

    /// Whether cross-encoder pair rerank is configured on this embedder.
    pub fn rerank_enabled(&self) -> bool {
        self.rerank_endpoint.is_some() || self.rerank_model.is_some()
    }

    /// Embed one sample input and return the observed vector dimension.
    pub async fn probe_dimension(&self, input: &str) -> Result<usize, QqlError> {
        let input = [input.to_string()];
        let body = EmbedRequest {
            model: &self.model,
            input: &input,
        };

        let resp = self
            .do_request(&self.endpoint, &self.api_key, &body)
            .await?;

        if resp.data.is_empty() {
            return Err(QqlError::execution(
                "QQL-EMBEDDING",
                "embedding response contained no vectors",
                None,
            ));
        }

        match &resp.data[0].embedding {
            EmbeddingPayload::Dense(v) => Ok(v.len()),
            EmbeddingPayload::Multi(rows) => rows.first().map(Vec::len).ok_or_else(|| {
                QqlError::execution(
                    "QQL-EMBEDDING",
                    "embedding response multivector was empty",
                    None,
                )
            }),
        }
    }

    async fn do_request(
        &self,
        endpoint: &str,
        api_key: &str,
        body: &EmbedRequest<'_>,
    ) -> Result<EmbedResponse, QqlError> {
        let mut req = self.client.post(endpoint).json(body);

        if !api_key.is_empty() {
            req = req.bearer_auth(api_key);
        }

        let resp = req.send().await.map_err(|e| {
            QqlError::execution(
                "QQL-EMBEDDING",
                format!("failed to call embedding endpoint: {}", e),
                None,
            )
        })?;

        let status = resp.status();
        if !status.is_success() {
            let text = resp.text().await.unwrap_or_default();
            return Err(QqlError::execution(
                "QQL-EMBEDDING",
                format!("embedding endpoint returned {}: {}", status, text),
                None,
            ));
        }

        let decoded: EmbedResponse = resp.json().await.map_err(|e| {
            QqlError::execution(
                "QQL-EMBEDDING",
                format!("failed to decode embedding response: {}", e),
                None,
            )
        })?;

        Ok(decoded)
    }

    fn resolve_dense_model<'a>(&'a self, model: &'a str) -> &'a str {
        if !model.is_empty() && model != "default" {
            model
        } else {
            self.model.as_str()
        }
    }

    fn resolve_multi_model<'a>(&'a self, model: &'a str) -> &'a str {
        if !model.is_empty() && model != "default" {
            model
        } else {
            self.multi_model.as_deref().unwrap_or(self.model.as_str())
        }
    }

    fn multi_url(&self) -> &str {
        self.multi_endpoint
            .as_deref()
            .unwrap_or(self.endpoint.as_str())
    }

    fn multi_key(&self) -> &str {
        self.multi_api_key
            .as_deref()
            .unwrap_or(self.api_key.as_str())
    }

    /// Embed a dense batch in one request, validating count, index coverage,
    /// and per-vector dimension against the configured `dimension`.
    pub async fn embed_batch_with_model(
        &self,
        inputs: &[String],
        model: &str,
    ) -> Result<Vec<Vec<f32>>, QqlError> {
        // One HTTP request for the full batch (OpenAI: up to 2048 inputs).
        if inputs.is_empty() {
            return Ok(Vec::new());
        }

        let model_name = self.resolve_dense_model(model);
        let body = EmbedRequest {
            model: model_name,
            input: inputs,
        };

        let decoded = self
            .do_request(&self.endpoint, &self.api_key, &body)
            .await?;

        if decoded.data.len() != inputs.len() {
            return Err(QqlError::execution(
                "QQL-EMBEDDING",
                format!(
                    "embedding response returned {} vector(s) for {} input(s)",
                    decoded.data.len(),
                    inputs.len()
                ),
                None,
            ));
        }

        let mut vectors: Vec<Option<Vec<f32>>> = vec![None; inputs.len()];
        for item in decoded.data {
            if item.index >= inputs.len() {
                return Err(QqlError::execution(
                    "QQL-EMBEDDING",
                    format!("embedding response index {} out of range", item.index),
                    None,
                ));
            }
            if vectors[item.index].is_some() {
                return Err(QqlError::execution(
                    "QQL-EMBEDDING",
                    format!("embedding response duplicated index {}", item.index),
                    None,
                ));
            }
            let dense = match item.embedding {
                EmbeddingPayload::Dense(v) => v,
                EmbeddingPayload::Multi(_) => {
                    return Err(QqlError::execution(
                        "QQL-EMBEDDING",
                        format!(
                            "embedding response index {} returned multivector; expected dense",
                            item.index
                        ),
                        None,
                    ));
                }
            };
            if dense.len() != self.dimension {
                return Err(QqlError::execution(
                    "QQL-EMBEDDING",
                    format!(
                        "embedding dimension mismatch for index {}: got {} want {}",
                        item.index,
                        dense.len(),
                        self.dimension
                    ),
                    None,
                ));
            }
            vectors[item.index] = Some(dense);
        }

        let mut result = Vec::with_capacity(vectors.len());
        for (i, v) in vectors.into_iter().enumerate() {
            if let Some(vec) = v {
                result.push(vec);
            } else {
                return Err(QqlError::execution(
                    "QQL-EMBEDDING",
                    format!("missing embedding vector at index {}", i),
                    None,
                ));
            }
        }

        Ok(result)
    }

    /// Embed a dense batch with the configured default model.
    pub async fn embed_batch(&self, inputs: &[String]) -> Result<Vec<Vec<f32>>, QqlError> {
        self.embed_batch_with_model(inputs, &self.model).await
    }

    async fn embed_multi_batch_with_model(
        &self,
        inputs: &[String],
        model: &str,
    ) -> Result<Vec<Vec<Vec<f32>>>, QqlError> {
        if inputs.is_empty() {
            return Ok(Vec::new());
        }
        if !self.multi_enabled() {
            return Err(qql_embed::embedder::multi_unsupported_error(model));
        }

        let model_name = self.resolve_multi_model(model);
        let body = EmbedRequest {
            model: model_name,
            input: inputs,
        };

        let decoded = self
            .do_request(self.multi_url(), self.multi_key(), &body)
            .await?;

        if decoded.data.len() != inputs.len() {
            return Err(QqlError::execution(
                "QQL-EMBEDDING-MULTI",
                format!(
                    "multi embedding response returned {} result(s) for {} input(s) (model={model_name})",
                    decoded.data.len(),
                    inputs.len()
                ),
                None,
            ));
        }

        let mut vectors: Vec<Option<Vec<Vec<f32>>>> = vec![None; inputs.len()];
        for item in decoded.data {
            if item.index >= inputs.len() {
                return Err(QqlError::execution(
                    "QQL-EMBEDDING-MULTI",
                    format!("multi embedding response index {} out of range", item.index),
                    None,
                ));
            }
            if vectors[item.index].is_some() {
                return Err(QqlError::execution(
                    "QQL-EMBEDDING-MULTI",
                    format!("multi embedding response duplicated index {}", item.index),
                    None,
                ));
            }
            let multi = match item.embedding {
                EmbeddingPayload::Multi(rows) => rows,
                EmbeddingPayload::Dense(flat) => {
                    return Err(QqlError::execution(
                        "QQL-EMBEDDING-MULTI",
                        format!(
                            "multi embedding endpoint returned a flat dense vector (len={}) for index {}; \
                             expected nested array [[f32,…],…] (token-level multivector). \
                             Point MODEL at a ColBERT multi service or set multi_embedding_endpoint.",
                            flat.len(),
                            item.index
                        ),
                        None,
                    ));
                }
            };
            if multi.is_empty() {
                return Err(QqlError::execution(
                    "QQL-EMBEDDING-MULTI",
                    format!("multi embedding returned empty bag at index {}", item.index),
                    None,
                ));
            }
            if self.multi_dimension > 0 {
                for (row_i, row) in multi.iter().enumerate() {
                    if row.len() != self.multi_dimension {
                        return Err(QqlError::execution(
                            "QQL-EMBEDDING-MULTI",
                            format!(
                                "multi embedding dimension mismatch at index {} row {}: got {} want {}",
                                item.index,
                                row_i,
                                row.len(),
                                self.multi_dimension
                            ),
                            None,
                        ));
                    }
                }
            }
            vectors[item.index] = Some(multi);
        }

        let mut result = Vec::with_capacity(vectors.len());
        for (i, v) in vectors.into_iter().enumerate() {
            if let Some(vec) = v {
                result.push(vec);
            } else {
                return Err(QqlError::execution(
                    "QQL-EMBEDDING-MULTI",
                    format!("missing multi embedding at index {}", i),
                    None,
                ));
            }
        }
        Ok(result)
    }

    fn resolve_image_model<'a>(&'a self, model: &'a str) -> &'a str {
        if !model.is_empty() && model != "default" {
            model
        } else {
            self.image_model.as_deref().unwrap_or(self.model.as_str())
        }
    }

    fn image_url(&self) -> &str {
        self.image_endpoint
            .as_deref()
            .unwrap_or(self.endpoint.as_str())
    }

    fn image_key(&self) -> &str {
        self.image_api_key
            .as_deref()
            .unwrap_or(self.api_key.as_str())
    }

    fn image_dim(&self) -> usize {
        if self.image_dimension > 0 {
            self.image_dimension
        } else {
            self.dimension
        }
    }

    /// Embed image paths/URLs via OpenAI-compatible dense endpoint.
    ///
    /// Contract: `POST { "model", "input": ["path-or-url", ...] }` → dense
    /// `embedding: [f32,…]` per input (same shape as text). Servers that accept
    /// filesystem paths or HTTP(S) image URLs can back CLIP vision this way.
    async fn embed_image_batch_with_model(
        &self,
        sources: &[String],
        model: &str,
    ) -> Result<Vec<Vec<f32>>, QqlError> {
        if sources.is_empty() {
            return Ok(Vec::new());
        }
        if !self.image_enabled() {
            return Err(qql_embed::embedder::image_unsupported_error(model));
        }

        let model_name = self.resolve_image_model(model);
        let body = EmbedRequest {
            model: model_name,
            input: sources,
        };
        let decoded = self
            .do_request(self.image_url(), self.image_key(), &body)
            .await?;

        if decoded.data.len() != sources.len() {
            return Err(QqlError::execution(
                "QQL-EMBEDDING-IMAGE",
                format!(
                    "image embedding response returned {} vector(s) for {} source(s)",
                    decoded.data.len(),
                    sources.len()
                ),
                None,
            ));
        }

        let want_dim = self.image_dim();
        let mut vectors: Vec<Option<Vec<f32>>> = vec![None; sources.len()];
        for item in decoded.data {
            if item.index >= sources.len() {
                return Err(QqlError::execution(
                    "QQL-EMBEDDING-IMAGE",
                    format!("image embedding response index {} out of range", item.index),
                    None,
                ));
            }
            let dense = match item.embedding {
                EmbeddingPayload::Dense(v) => v,
                EmbeddingPayload::Multi(_) => {
                    return Err(QqlError::execution(
                        "QQL-EMBEDDING-IMAGE",
                        format!(
                            "image embedding index {} returned multivector; expected dense (CLIP vision is single-vector)",
                            item.index
                        ),
                        None,
                    ));
                }
            };
            if dense.len() != want_dim {
                return Err(QqlError::execution(
                    "QQL-EMBEDDING-IMAGE",
                    format!(
                        "image embedding dimension mismatch for index {}: got {} want {}",
                        item.index,
                        dense.len(),
                        want_dim
                    ),
                    None,
                ));
            }
            vectors[item.index] = Some(dense);
        }

        let mut result = Vec::with_capacity(vectors.len());
        for (i, v) in vectors.into_iter().enumerate() {
            if let Some(vec) = v {
                result.push(vec);
            } else {
                return Err(QqlError::execution(
                    "QQL-EMBEDDING-IMAGE",
                    format!("missing image embedding at index {}", i),
                    None,
                ));
            }
        }
        Ok(result)
    }
}

#[cfg(feature = "rest")]
#[cfg_attr(target_arch = "wasm32", async_trait(?Send))]
#[cfg_attr(not(target_arch = "wasm32"), async_trait)]
impl Embedder for HttpEmbedder {
    fn dimension(&self) -> Option<usize> {
        Some(self.dimension)
    }

    fn multi_dimension(&self) -> Option<usize> {
        if self.multi_dimension > 0 {
            Some(self.multi_dimension)
        } else {
            None
        }
    }

    async fn embed_dense(&self, text: &str, model: &str) -> Result<Vec<f32>, QqlError> {
        let results = self
            .embed_batch_with_model(&[text.to_string()], model)
            .await?;
        Ok(results.into_iter().next().unwrap_or_default())
    }

    async fn embed_dense_batch(
        &self,
        texts: &[String],
        model: &str,
    ) -> Result<Vec<Vec<f32>>, QqlError> {
        self.embed_batch_with_model(texts, model).await
    }

    async fn embed_sparse_query(&self, text: &str, model: &str) -> Result<SparseVector, QqlError> {
        if !model.is_empty() && !model.eq_ignore_ascii_case("default") {
            return Err(qql_embed::sparse_model_unsupported_error(model));
        }
        Ok(qql_embed::sparse::embed_query(text))
    }

    async fn embed_sparse_document(
        &self,
        text: &str,
        model: &str,
    ) -> Result<SparseVector, QqlError> {
        if !model.is_empty() && !model.eq_ignore_ascii_case("default") {
            return Err(qql_embed::sparse_model_unsupported_error(model));
        }
        Ok(qql_embed::sparse::embed_document(text))
    }

    async fn embed_multi(&self, text: &str, model: &str) -> Result<Vec<Vec<f32>>, QqlError> {
        let results = self
            .embed_multi_batch_with_model(&[text.to_string()], model)
            .await?;
        results.into_iter().next().ok_or_else(|| {
            QqlError::execution(
                "QQL-EMBEDDING-MULTI",
                "multi embedding response was empty",
                None,
            )
        })
    }

    async fn embed_multi_batch(
        &self,
        texts: &[String],
        model: &str,
    ) -> Result<Vec<Vec<Vec<f32>>>, QqlError> {
        self.embed_multi_batch_with_model(texts, model).await
    }

    async fn embed_image(&self, source: &str, model: &str) -> Result<Vec<f32>, QqlError> {
        let results = self
            .embed_image_batch_with_model(&[source.to_string()], model)
            .await?;
        results.into_iter().next().ok_or_else(|| {
            QqlError::execution(
                "QQL-EMBEDDING-IMAGE",
                "image embedding response was empty",
                None,
            )
        })
    }

    async fn embed_image_batch(
        &self,
        sources: &[String],
        model: &str,
    ) -> Result<Vec<Vec<f32>>, QqlError> {
        self.embed_image_batch_with_model(sources, model).await
    }

    async fn rerank_pairs(
        &self,
        query: &str,
        documents: &[String],
        model: &str,
    ) -> Result<Vec<f32>, QqlError> {
        if documents.is_empty() {
            return Ok(Vec::new());
        }
        if !self.rerank_enabled() {
            return Err(qql_embed::cross_rerank_unsupported_error(model));
        }
        let endpoint = self
            .rerank_endpoint
            .as_deref()
            .ok_or_else(|| qql_embed::cross_rerank_unsupported_error(model))?;
        let model_name = if !model.is_empty() && model != "default" {
            model.to_string()
        } else {
            self.rerank_model
                .clone()
                .unwrap_or_else(|| "rerank".to_string())
        };
        let api_key = self
            .rerank_api_key
            .as_deref()
            .unwrap_or(self.api_key.as_str());

        // Cohere-compatible: { model, query, documents } → results[{index,relevance_score}]
        let body = serde_json::json!({
            "model": model_name,
            "query": query,
            "documents": documents,
        });
        let mut req = self.client.post(endpoint).json(&body);
        if !api_key.is_empty() {
            req = req.header("Authorization", format!("Bearer {api_key}"));
        }
        let resp = req.send().await.map_err(|e| {
            QqlError::execution(
                "QQL-RERANK-CROSS",
                format!("failed to call rerank endpoint: {e}"),
                None,
            )
        })?;
        let status = resp.status();
        if !status.is_success() {
            let text = resp.text().await.unwrap_or_default();
            return Err(QqlError::execution(
                "QQL-RERANK-CROSS",
                format!("rerank endpoint returned {status}: {text}"),
                None,
            ));
        }
        let value: serde_json::Value = resp.json().await.map_err(|e| {
            QqlError::execution(
                "QQL-RERANK-CROSS",
                format!("failed to decode rerank response: {e}"),
                None,
            )
        })?;
        // Accept results[] with index + relevance_score | score
        let results = value
            .get("results")
            .or_else(|| value.get("data"))
            .and_then(|v| v.as_array())
            .ok_or_else(|| {
                QqlError::execution(
                    "QQL-RERANK-CROSS",
                    "rerank response missing results array",
                    None,
                )
            })?;
        let mut scores = vec![0.0f32; documents.len()];
        let mut seen = vec![false; documents.len()];
        for item in results {
            let idx = item.get("index").and_then(|v| v.as_u64()).ok_or_else(|| {
                QqlError::execution("QQL-RERANK-CROSS", "rerank result missing index", None)
            })? as usize;
            if idx >= documents.len() {
                return Err(QqlError::execution(
                    "QQL-RERANK-CROSS",
                    format!("rerank result index {idx} out of range"),
                    None,
                ));
            }
            let score = item
                .get("relevance_score")
                .or_else(|| item.get("score"))
                .and_then(|v| v.as_f64())
                .unwrap_or(0.0) as f32;
            scores[idx] = score;
            seen[idx] = true;
        }
        if seen.iter().any(|s| !*s) {
            return Err(QqlError::execution(
                "QQL-RERANK-CROSS",
                "rerank response did not cover all documents",
                None,
            ));
        }
        Ok(scores)
    }
}