qql-edge 0.2.0

Embedded QQL executor backed by qdrant-edge and optional local FastEmbed inference
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
//! # QQL Edge — zero-network vector search
//!
//! Combines [fastembed-rs] for local ONNX embedding inference with
//! [qdrant-edge] for in-process HNSW vector search. No network hops,
//! no external services, no API keys — unless you choose an external provider.
//!
//! ## Embedder options
//!
//! | Function | Embedder | Network? |
//! |---|---|---|
//! | [`local_executor`] | fastembed (ONNX, local CPU) | ❌ none |
//! | [`local_executor_with_options`] | fastembed + model/cache selection | ❌ none |
//! | [`http_executor`] | OpenAI-compatible HTTP endpoint | ✅ provider only |
//! | [`custom_executor`] | Any `Arc<dyn Embedder>` | up to you |
//!
//! ## Quick start
//!
//! ```rust,no_run
//! use qql_edge::local_executor;
//! use qql::executor::OnError;
//!
//! # async fn example() -> Result<(), qql_core::error::QqlError> {
//! let mut executor = local_executor("/tmp/qql-edge-data", false)?;
//! let resp = executor.execute("CREATE COLLECTION docs HYBRID", OnError::Stop).await?;
//! # Ok(())
//! # }
//! ```
//!
//! [fastembed-rs]: https://crates.io/crates/fastembed
//! [qdrant-edge]: https://crates.io/crates/qdrant-edge

mod backend;
#[cfg(feature = "fastembed-local")]
mod embedder;

pub use backend::EdgeQdrant;
#[cfg(feature = "fastembed-local")]
pub use embedder::{
    list_embedding_models, resolve_embedding_model, resolve_image_model, resolve_multi_model,
    EmbeddingModelInfo, FastEmbedder, FastEmbedderOptions,
};

use qql::config::QqlConfig;
use qql::embedder::Embedder;
use qql::executor::Executor;
use std::path::PathBuf;
use std::sync::Arc;

/// Options for [`local_executor_with_options`].
#[derive(Debug, Clone, Default)]
pub struct LocalExecutorOptions {
    /// Store payloads on disk (default: `true` when constructed via bindings;
    /// this struct defaults to `false` for Rust ergonomics matching the
    /// historical `local_executor(path, false)` tests).
    pub on_disk_payload: bool,
    /// Local ONNX dense model name. See [`resolve_embedding_model`] for accepted forms.
    /// `None` → default `BGESmallENV15` (384-d).
    #[cfg(feature = "fastembed-local")]
    pub model: Option<String>,
    /// Offline sparse model (SPLADE or BGE-M3 via `SparseTextEmbedding`).
    /// e.g. `"splade"`, `"bge-m3"`. When set, `embed_sparse` uses real ONNX
    /// inference. `None` → local BM25 hashing for sparse requests.
    #[cfg(feature = "fastembed-local")]
    pub sparse_model: Option<String>,
    /// Offline multivector model (BGE-M3 ColBERT). e.g. `"bge-m3"`.
    /// When set, `embed_multi` / multivector RERANK work with no network.
    #[cfg(feature = "fastembed-local")]
    pub multi_model: Option<String>,
    /// Offline image / CLIP vision model. e.g. `"clip-vision"` / `ClipVitB32`.
    /// Pair with dense CLIP text (`model: Some("ClipVitB32".into())`) for multimodal.
    #[cfg(feature = "fastembed-local")]
    pub image_model: Option<String>,
    /// Offline cross-encoder (`bge-reranker-base`, `BGERerankerBase`, …).
    #[cfg(feature = "fastembed-local")]
    pub reranker_model: Option<String>,
    /// Override fastembed model cache directory.
    #[cfg(feature = "fastembed-local")]
    pub cache_dir: Option<PathBuf>,
    /// Show HuggingFace download progress bars (default: `false`).
    #[cfg(feature = "fastembed-local")]
    pub show_download_progress: bool,
}

/// Build a fully-local [`Executor`] backed by fastembed-rs and qdrant-edge.
///
/// Uses the default embedding model (`BGESmallENV15`, 384-d). Prefer
/// [`local_executor_with_options`] when you need a different model or cache dir.
///
/// No network calls are made at inference time — embedding runs on-device via ONNX.
/// Models are downloaded from HuggingFace on first use and cached locally.
#[cfg(feature = "fastembed-local")]
pub fn local_executor(
    data_dir: impl Into<PathBuf>,
    on_disk_payload: bool,
) -> Result<Executor, qql_core::error::QqlError> {
    local_executor_with_options(
        data_dir,
        LocalExecutorOptions {
            on_disk_payload,
            ..Default::default()
        },
    )
}

/// Build a fully-local [`Executor`] with explicit model / cache options.
#[cfg(feature = "fastembed-local")]
pub fn local_executor_with_options(
    data_dir: impl Into<PathBuf>,
    opts: LocalExecutorOptions,
) -> Result<Executor, qql_core::error::QqlError> {
    let client = Box::new(EdgeQdrant::new(data_dir, opts.on_disk_payload));
    let embedder = FastEmbedder::try_with_options(FastEmbedderOptions {
        model: opts.model,
        sparse_model: opts.sparse_model,
        multi_model: opts.multi_model,
        image_model: opts.image_model,
        reranker_model: opts.reranker_model,
        cache_dir: opts.cache_dir,
        show_download_progress: opts.show_download_progress,
    })?;

    // Pin collection vector size to the actual model dimension. Without this,
    // CREATE COLLECTION HYBRID always falls back to the hard-coded 384 and any
    // non-default model (768 / 1024-d) silently dimension-mismatches on upsert.
    let multi_dim = embedder.multi_dimension().unwrap_or(0);
    let image_dim = embedder.image_dimension().unwrap_or(0);
    let config = QqlConfig {
        inference_mode: "local".to_string(),
        embedding_dimension: embedder.dimension(),
        embedding_model: Some(embedder.model_name().to_string()),
        sparse_inference_model: embedder.sparse_model_code().map(str::to_string),
        multi_embedding_model: embedder.multi_model_code().map(str::to_string),
        multi_embedding_dimension: multi_dim,
        image_embedding_model: embedder.image_model_code().map(str::to_string),
        image_embedding_dimension: image_dim,
        rerank_model: embedder.reranker_model_code().map(str::to_string),
        ..Default::default()
    };

    let embedder = Some(Arc::new(embedder) as Arc<dyn Embedder>);
    Ok(Executor::with_embedder(client, Some(config), embedder))
}

/// Build an edge [`Executor`] that calls an external OpenAI-compatible embedding
/// endpoint instead of running fastembed locally.
///
/// Works with: OpenAI, Ollama (`/v1/embeddings`), Cohere, Together AI,
/// Mistral, and any other provider that follows the OpenAI embeddings spec.
///
/// - `endpoint` — full URL, e.g. `"https://api.openai.com/v1/embeddings"` or
///   `"http://localhost:11434/v1/embeddings"` for local Ollama.
/// - `api_key` — Bearer token. Pass `""` for unauthenticated local providers.
/// - `model` — model name sent in the request body, e.g. `"text-embedding-3-small"`.
/// - `dimension` — expected output dimension. Must match what the model returns.
#[cfg(feature = "http-embedding")]
pub fn http_executor(
    data_dir: impl Into<PathBuf>,
    on_disk_payload: bool,
    endpoint: impl Into<String>,
    api_key: impl Into<String>,
    model: impl Into<String>,
    dimension: usize,
) -> Result<Executor, qql_core::error::QqlError> {
    http_executor_with_options(
        data_dir,
        on_disk_payload,
        qql::embedder::HttpEmbedderOptions {
            endpoint: endpoint.into(),
            api_key: api_key.into(),
            model: model.into(),
            dimension,
            ..Default::default()
        },
    )
}

/// Edge executor with OpenAI-compatible dense + optional multi/ColBERT endpoints.
#[cfg(feature = "http-embedding")]
#[allow(clippy::too_many_arguments)]
pub fn http_executor_with_multi(
    data_dir: impl Into<PathBuf>,
    on_disk_payload: bool,
    endpoint: impl Into<String>,
    api_key: impl Into<String>,
    model: impl Into<String>,
    dimension: usize,
    multi_endpoint: Option<String>,
    multi_api_key: Option<String>,
    multi_model: Option<String>,
    multi_dimension: usize,
) -> Result<Executor, qql_core::error::QqlError> {
    http_executor_with_options(
        data_dir,
        on_disk_payload,
        qql::embedder::HttpEmbedderOptions {
            endpoint: endpoint.into(),
            api_key: api_key.into(),
            model: model.into(),
            dimension,
            multi_endpoint,
            multi_api_key,
            multi_model,
            multi_dimension,
            ..Default::default()
        },
    )
}

/// Edge executor from full [`qql::embedder::HttpEmbedderOptions`] (dense + multi + image/CLIP).
#[cfg(feature = "http-embedding")]
pub fn http_executor_with_options(
    data_dir: impl Into<PathBuf>,
    on_disk_payload: bool,
    opts: qql::embedder::HttpEmbedderOptions,
) -> Result<Executor, qql_core::error::QqlError> {
    let client = Box::new(EdgeQdrant::new(data_dir, on_disk_payload));
    let config = QqlConfig {
        inference_mode: "local".to_string(),
        embedding_dimension: opts.dimension,
        embedding_model: Some(opts.model.clone()),
        embedding_endpoint: None, // edge uses the Arc embedder, not config probing
        multi_embedding_endpoint: opts.multi_endpoint.clone(),
        multi_embedding_api_key: opts.multi_api_key.clone(),
        multi_embedding_model: opts.multi_model.clone(),
        multi_embedding_dimension: opts.multi_dimension,
        image_embedding_endpoint: opts.image_endpoint.clone(),
        image_embedding_api_key: opts.image_api_key.clone(),
        image_embedding_model: opts.image_model.clone(),
        image_embedding_dimension: opts.image_dimension,
        rerank_endpoint: opts.rerank_endpoint.clone(),
        rerank_api_key: opts.rerank_api_key.clone(),
        rerank_model: opts.rerank_model.clone(),
        ..Default::default()
    };
    let embedder = qql::embedder::HttpEmbedder::try_with_options(opts)?;
    Ok(Executor::with_embedder(
        client,
        Some(config),
        Some(Arc::new(embedder) as Arc<dyn Embedder>),
    ))
}

/// Build an edge [`Executor`] with a fully custom [`Embedder`].
///
/// Use this to plug in GPU-backed embedders, caching layers, ensemble
/// embedders, or any other custom implementation.
///
/// **Callers must set `embedding_dimension` on a [`QqlConfig`] themselves** if
/// they rely on `CREATE COLLECTION HYBRID` auto-sizing — pass the config via
/// [`Executor::with_embedder`] if the defaults (384) are wrong for your model.
pub fn custom_executor(
    data_dir: impl Into<PathBuf>,
    on_disk_payload: bool,
    embedder: Arc<dyn Embedder>,
) -> Result<Executor, qql_core::error::QqlError> {
    custom_executor_with_dimension(data_dir, on_disk_payload, embedder, None)
}

/// Build an edge [`Executor`] with a custom embedder and an optional explicit
/// dense dimension. The embedder-reported dimension takes precedence.
pub fn custom_executor_with_dimension(
    data_dir: impl Into<PathBuf>,
    on_disk_payload: bool,
    embedder: Arc<dyn Embedder>,
    dimension: Option<usize>,
) -> Result<Executor, qql_core::error::QqlError> {
    let client = Box::new(EdgeQdrant::new(data_dir, on_disk_payload));
    let config = QqlConfig {
        inference_mode: "local".to_string(),
        embedding_dimension: embedder.dimension().or(dimension).unwrap_or(0),
        ..Default::default()
    };
    Ok(Executor::with_embedder(
        client,
        Some(config),
        Some(embedder),
    ))
}

#[cfg(test)]
mod tests {
    use super::*;
    use async_trait::async_trait;
    use qql::embedder::Embedder;
    use qql::executor::OnError;
    use qql_core::error::QqlError;
    use qql_embed::sparse::SparseVector;

    struct TestEmbedder;

    #[async_trait]
    impl Embedder for TestEmbedder {
        fn dimension(&self) -> Option<usize> {
            Some(3)
        }

        async fn embed_dense(&self, _text: &str, _model: &str) -> Result<Vec<f32>, QqlError> {
            Ok(vec![1.0, 0.0, 0.0])
        }

        async fn embed_sparse(&self, _text: &str, _model: &str) -> Result<SparseVector, QqlError> {
            Ok(SparseVector {
                indices: vec![1],
                values: vec![1.0],
            })
        }
    }

    #[test]
    #[cfg(feature = "fastembed-local")]
    fn local_executor_options_default_has_no_sparse_model() {
        let opts = LocalExecutorOptions::default();
        assert!(opts.sparse_model.is_none());
    }

    #[test]
    #[cfg(feature = "fastembed-local")]
    fn local_executor_options_with_sparse_model() {
        let opts = LocalExecutorOptions {
            sparse_model: Some("splade".into()),
            ..Default::default()
        };
        assert_eq!(opts.sparse_model.as_deref(), Some("splade"));
    }

    #[test]
    fn custom_edge_executor_is_schema_aware_and_rejects_shards() {
        let runtime = tokio::runtime::Builder::new_current_thread()
            .enable_all()
            .build()
            .expect("test runtime");
        runtime.block_on(async {
            let data_dir =
                std::env::temp_dir().join(format!("qql-edge-test-{}", std::process::id()));
            let _ = std::fs::remove_dir_all(&data_dir);
            let executor =
                custom_executor(&data_dir, false, Arc::new(TestEmbedder)).expect("custom executor");

            let report = executor
                .execute("CREATE COLLECTION docs HYBRID", OnError::Stop)
                .await
                .expect("create collection");
            assert!(report.ok);
            let report = executor
                .execute(
                    "UPSERT INTO docs VALUES {id: 1, text: 'hello'}",
                    OnError::Stop,
                )
                .await
                .expect("schema-aware upsert");
            assert!(report.ok);
            let report = executor
                .execute(
                    "QUERY 'hello' FROM docs USING dense AS DENSE LIMIT 1",
                    OnError::Stop,
                )
                .await
                .expect("schema-aware query");
            assert!(report.ok);
            let report = executor
                .execute(
                    "CREATE COLLECTION dense_only (dense VECTOR(3, COSINE))",
                    OnError::Stop,
                )
                .await
                .expect("create dense-only collection");
            assert!(report.ok);
            let report = executor
                .execute(
                    "UPSERT INTO dense_only VALUES {id: 2, body: 'dense only'}",
                    OnError::Stop,
                )
                .await
                .expect("dense-only auto-embed");
            assert!(report.ok);
            let report = executor
                .execute(
                    "CREATE COLLECTION sparse_only (sparse SPARSE)",
                    OnError::Stop,
                )
                .await
                .expect("create sparse-only collection");
            assert!(report.ok);
            let report = executor
                .execute(
                    "UPSERT INTO sparse_only VALUES {id: 3, content: 'sparse only'}",
                    OnError::Stop,
                )
                .await
                .expect("sparse-only auto-embed");
            assert!(report.ok);
            let report = executor
                .execute(
                    "QUERY 'hello' FROM docs USING dense AS DENSE SHARD 'tenant-a' LIMIT 1",
                    OnError::Continue,
                )
                .await
                .expect("edge should report unsupported shard");
            assert!(!report.ok);
            assert!(report.results[0].message.contains("UNSUPPORTED-SHARD"));
            executor.close().await.expect("close edge executor");
            let _ = std::fs::remove_dir_all(data_dir);
        });
    }

    #[test]
    fn edge_reads_do_not_create_missing_collections() {
        // Regression (B-2): read operations against a missing collection must
        // error with QQL-EDGE-COLLECTION-NOT-FOUND and must not materialise a
        // ghost collection on disk — matching remote 404 semantics.
        let runtime = tokio::runtime::Builder::new_current_thread()
            .enable_all()
            .build()
            .expect("test runtime");
        runtime.block_on(async {
            let data_dir =
                std::env::temp_dir().join(format!("qql-edge-readtest-{}", std::process::id()));
            let _ = std::fs::remove_dir_all(&data_dir);
            let executor =
                custom_executor(&data_dir, false, Arc::new(TestEmbedder)).expect("custom executor");

            let cases = [
                "QUERY 'hello' FROM never_queried USING dense AS DENSE LIMIT 5",
                "SCROLL FROM never_scrolled LIMIT 5",
                "QUERY POINTS (1, 2) FROM never_points WITH PAYLOAD true",
                "COUNT FROM never_counted",
                "DELETE FROM never_deleted WHERE id = 1",
            ];
            for query in cases {
                let report = executor
                    .execute(query, OnError::Continue)
                    .await
                    .expect("read against missing collection must return a report");
                assert!(
                    !report.ok,
                    "expected failure for '{query}': {:?}",
                    report.results
                );
                assert!(
                    report.results[0]
                        .message
                        .contains("QQL-EDGE-COLLECTION-NOT-FOUND"),
                    "expected QQL-EDGE-COLLECTION-NOT-FOUND for '{query}', got {:?}",
                    report.results[0].message
                );
            }

            // No ghost collections were created on disk.
            let report = executor
                .execute("SHOW COLLECTIONS", OnError::Stop)
                .await
                .expect("list collections");
            assert!(report.ok);
            let cols = report.results[0]
                .data
                .as_ref()
                .and_then(|d| d["result"]["collections"].as_array())
                .expect("collections array");
            assert!(
                cols.is_empty(),
                "read operations must not create collections: {cols:?}"
            );

            executor.close().await.expect("close edge executor");
            let _ = std::fs::remove_dir_all(data_dir);
        });
    }

    #[test]
    fn edge_upsert_autocreate_and_get_points_shape() {
        // Regression (B-2 preserve + B-3): an embedding UPSERT to a missing
        // collection still auto-creates it with the inferred schema, and
        // QUERY POINTS on a populated collection returns non-empty hits.
        let runtime = tokio::runtime::Builder::new_current_thread()
            .enable_all()
            .build()
            .expect("test runtime");
        runtime.block_on(async {
            let data_dir =
                std::env::temp_dir().join(format!("qql-edge-pointstest-{}", std::process::id()));
            let _ = std::fs::remove_dir_all(&data_dir);
            let executor =
                custom_executor(&data_dir, false, Arc::new(TestEmbedder)).expect("custom executor");

            // Auto-create via embedding upsert (collection does not exist yet).
            let report = executor
                .execute(
                    "UPSERT INTO auto_docs VALUES {id: 1, text: 'hello'}, {id: 2, text: 'world'}",
                    OnError::Stop,
                )
                .await
                .expect("auto-creating upsert");
            assert!(report.ok, "auto-creating UPSERT failed: {:?}", report);

            let report = executor
                .execute("SHOW COLLECTIONS", OnError::Stop)
                .await
                .expect("list collections");
            let cols = report.results[0]
                .data
                .as_ref()
                .and_then(|d| d["result"]["collections"].as_array())
                .map(|c| c.len())
                .unwrap_or(0);
            assert_eq!(
                cols, 1,
                "UPSERT auto-create should materialise one collection"
            );

            // QUERY POINTS must surface the retrieved points (response shape
            // regression) and report the hit count.
            let report = executor
                .execute(
                    "QUERY POINTS (1, 2) FROM auto_docs WITH PAYLOAD true",
                    OnError::Stop,
                )
                .await
                .expect("points lookup");
            assert!(report.ok, "POINTS lookup failed: {:?}", report);
            assert_eq!(
                report.results[0].message, "Found 2 hits",
                "POINTS lookup should report 2 hits, got {:?}",
                report.results[0].message
            );
            let data = report.results[0].data.as_ref().expect("data");
            let hits = data.as_array().expect("hits array");
            assert_eq!(hits.len(), 2, "POINTS lookup should return 2 hits");

            executor.close().await.expect("close edge executor");
            let _ = std::fs::remove_dir_all(data_dir);
        });
    }
}