graphrt-rag 0.1.0

Retrieval building blocks for graphrt: embedders, vector stores, hybrid search, reranking, chunking and the knowledge-retrieval node.
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
//! [`KnowledgeBase`]: the service the `knowledge-retrieval` node uses.
//!
//! A knowledge base maps dataset ids to retrievers. A dataset is either backed
//! by the built-in pipeline (a [`VectorStore`] + [`Embedder`], optionally a
//! [`Reranker`]) or by any external [`Retriever`]. Register it in
//! `graphrt_core::Services` as a plain `KnowledgeBase` value:
//!
//! ```
//! # use std::sync::Arc;
//! # use graphrt_core::Services;
//! # use graphrt_rag::{KnowledgeBase, MemoryStore, HashEmbedder};
//! let kb = KnowledgeBase::new()
//!     .with_store_dataset("docs", Arc::new(MemoryStore::new()), Arc::new(HashEmbedder::new(64)), None);
//! let services = Services::new().with(kb);
//! ```

use std::sync::Arc;

use indexmap::IndexMap;
use serde::{Deserialize, Serialize};

use crate::error::RagError;
use crate::ingest::Ingestor;
use crate::pipeline::{StoreRetriever, finalize, rerank};
use crate::types::{Embedder, Hit, Reranker, RetrievalRequest, Retriever, VectorStore};

/// Public description of a dataset.
#[derive(Clone, Debug, Default, PartialEq, Eq, Serialize, Deserialize)]
pub struct DatasetInfo {
    pub id: String,
    #[serde(default)]
    pub name: String,
    #[serde(default)]
    pub description: String,
}

#[derive(Clone, Debug)]
struct DatasetEntry {
    info: DatasetInfo,
    retriever: Arc<dyn Retriever>,
    store: Option<(Arc<dyn VectorStore>, Arc<dyn Embedder>)>,
}

/// Dataset id -> retriever registry.
#[derive(Clone, Debug, Default)]
pub struct KnowledgeBase {
    datasets: IndexMap<String, DatasetEntry>,
    /// Reranker applied across datasets in multi-dataset retrieval.
    reranker: Option<Arc<dyn Reranker>>,
}

impl KnowledgeBase {
    pub fn new() -> Self {
        Self::default()
    }

    /// Add a dataset backed by the built-in store pipeline.
    pub fn with_store_dataset(
        mut self,
        id: impl Into<String>,
        store: Arc<dyn VectorStore>,
        embedder: Arc<dyn Embedder>,
        reranker: Option<Arc<dyn Reranker>>,
    ) -> Self {
        let id = id.into();
        self.add_store_dataset(DatasetInfo { id, ..DatasetInfo::default() }, store, embedder, reranker);
        self
    }

    /// Add (or replace) a store-backed dataset with descriptive info.
    pub fn add_store_dataset(
        &mut self,
        info: DatasetInfo,
        store: Arc<dyn VectorStore>,
        embedder: Arc<dyn Embedder>,
        reranker: Option<Arc<dyn Reranker>>,
    ) -> &mut Self {
        let retriever = StoreRetriever::new(store.clone(), embedder.clone()).with_reranker(reranker);
        self.datasets.insert(
            info.id.clone(),
            DatasetEntry { info, retriever: Arc::new(retriever), store: Some((store, embedder)) },
        );
        self
    }

    /// Add a dataset served by an external retriever.
    pub fn with_retriever(mut self, id: impl Into<String>, retriever: Arc<dyn Retriever>) -> Self {
        let id = id.into();
        self.add_retriever(DatasetInfo { id, ..DatasetInfo::default() }, retriever);
        self
    }

    /// Add (or replace) a dataset served by an external retriever.
    pub fn add_retriever(&mut self, info: DatasetInfo, retriever: Arc<dyn Retriever>) -> &mut Self {
        self.datasets.insert(info.id.clone(), DatasetEntry { info, retriever, store: None });
        self
    }

    /// Reranker applied to the merged results of multi-dataset retrieval.
    pub fn with_reranker(mut self, reranker: Option<Arc<dyn Reranker>>) -> Self {
        self.reranker = reranker;
        self
    }

    pub fn datasets(&self) -> impl Iterator<Item = &DatasetInfo> {
        self.datasets.values().map(|d| &d.info)
    }

    pub fn contains(&self, dataset_id: &str) -> bool {
        self.datasets.contains_key(dataset_id)
    }

    /// An ingestor for a store-backed dataset (`None` for external retrievers or unknown ids).
    pub fn ingestor(&self, dataset_id: &str) -> Option<Ingestor> {
        let entry = self.datasets.get(dataset_id)?;
        let (store, embedder) = entry.store.as_ref()?;
        Some(Ingestor::new(store.clone(), embedder.clone()))
    }

    fn entry(&self, id: &str) -> Result<&DatasetEntry, RagError> {
        self.datasets.get(id).ok_or_else(|| RagError::NotFound(format!("unknown dataset `{id}`")))
    }

    /// Query every dataset concurrently. Returns `(dataset id, hits)` in request order.
    async fn query_each(
        &self,
        request: &RetrievalRequest,
        rerank_each: bool,
    ) -> Result<Vec<(String, Vec<Hit>)>, RagError> {
        let mut jobs = Vec::with_capacity(request.datasets.len());
        for id in &request.datasets {
            let entry = self.entry(id)?;
            let sub = RetrievalRequest {
                datasets: vec![id.clone()],
                // Threshold and top-k are applied after merging.
                score_threshold: None,
                rerank: rerank_each,
                ..request.clone()
            };
            jobs.push(async move { entry.retriever.retrieve(&sub).await.map(|hits| (id.clone(), hits)) });
        }
        futures::future::try_join_all(jobs).await
    }

    /// Retrieve across all requested datasets and merge by score.
    ///
    /// When `request.rerank` is set and the knowledge base has a cross-dataset
    /// reranker, the merged list is reranked as a whole (making scores from
    /// different datasets comparable). Otherwise each dataset's own reranker, if
    /// any, is applied before merging.
    pub async fn retrieve(&self, request: &RetrievalRequest) -> Result<Vec<Hit>, RagError> {
        validate_request(request)?;
        let global_rerank = request.rerank && self.reranker.is_some() && request.datasets.len() > 1;
        let per_dataset = self.query_each(request, request.rerank && !global_rerank).await?;
        let mut merged: Vec<Hit> = per_dataset.into_iter().flat_map(|(_, hits)| hits).collect();
        if global_rerank {
            if let Some(reranker) = &self.reranker {
                merged = rerank(reranker.as_ref(), &request.query, merged).await?;
            }
        }
        Ok(finalize(merged, request.score_threshold, request.top_k))
    }

    /// Retrieve from the single best-matching dataset: every dataset is queried,
    /// and only the hits of the dataset whose top hit scores highest are kept.
    pub async fn retrieve_single(&self, request: &RetrievalRequest) -> Result<Vec<Hit>, RagError> {
        validate_request(request)?;
        let per_dataset = self.query_each(request, request.rerank).await?;
        let best = per_dataset
            .into_iter()
            .filter(|(_, hits)| !hits.is_empty())
            .max_by(|a, b| top_score(&a.1).total_cmp(&top_score(&b.1)));
        Ok(best.map(|(_, hits)| finalize(hits, request.score_threshold, request.top_k)).unwrap_or_default())
    }
}

/// Reject requests that would otherwise silently match nothing.
fn validate_request(request: &RetrievalRequest) -> Result<(), RagError> {
    if request.datasets.is_empty() {
        return Err(RagError::Input("no dataset ids given".into()));
    }
    if let Some(filter) = &request.filter {
        filter.validate().map_err(|e| RagError::Input(e.to_string()))?;
    }
    Ok(())
}

fn top_score(hits: &[Hit]) -> f32 {
    hits.iter().map(|h| h.score).fold(f32::MIN, f32::max)
}

// ---------------------------------------------------------------------------
// Configuration
// ---------------------------------------------------------------------------

/// Declarative knowledge-base configuration (`knowledge.yaml` / `.json`).
///
/// ```yaml
/// embedders:
///   default: { type: openai, base_url: "http://localhost:11434/v1", model: nomic-embed-text }
///   cloud:   { type: openai, model: text-embedding-3-small, api_key_env: OPENAI_API_KEY }
/// rerankers:
///   cohere: { type: http, base_url: "https://api.cohere.com/v2", model: rerank-v3.5, api_key_env: COHERE_API_KEY }
/// stores:
///   local:  { type: sqlite, path: ./kb.db }
///   json:   { type: memory, path: ./kb.json }
///   vector: { type: qdrant, url: "http://localhost:6333", collection: kb, api_key_env: QDRANT_API_KEY }
/// datasets:
///   - { id: handbook, name: Handbook, store: local, embedder: default, reranker: cohere }
///   - { id: crm, retriever: { type: http, url: "https://rag.internal/search", api_key_env: RAG_TOKEN } }
/// reranker: cohere   # optional, used across datasets
/// ```
///
/// Secrets are never written inline: every provider takes `api_key_env`, the
/// name of an environment variable resolved when the configuration is built.
#[derive(Clone, Debug, Default, PartialEq, Serialize, Deserialize)]
#[serde(deny_unknown_fields)]
pub struct KnowledgeConfig {
    #[serde(default)]
    pub embedders: IndexMap<String, serde_json::Value>,
    #[serde(default)]
    pub rerankers: IndexMap<String, serde_json::Value>,
    #[serde(default)]
    pub stores: IndexMap<String, serde_json::Value>,
    #[serde(default)]
    pub datasets: Vec<DatasetConfig>,
    /// Cross-dataset reranker (a key of `rerankers`).
    #[serde(default)]
    pub reranker: Option<String>,
}

/// One dataset entry.
#[derive(Clone, Debug, Default, PartialEq, Serialize, Deserialize)]
#[serde(deny_unknown_fields)]
pub struct DatasetConfig {
    pub id: String,
    #[serde(default)]
    pub name: String,
    #[serde(default)]
    pub description: String,
    /// Key of `stores` (store-backed datasets).
    #[serde(default)]
    pub store: Option<String>,
    /// Key of `embedders` (store-backed datasets).
    #[serde(default)]
    pub embedder: Option<String>,
    /// Key of `rerankers`.
    #[serde(default)]
    pub reranker: Option<String>,
    /// External retriever definition (instead of store + embedder).
    #[serde(default)]
    pub retriever: Option<serde_json::Value>,
}

/// Resolves environment variable names to values (`std::env::var` in production).
pub type EnvLookup<'a> = &'a dyn Fn(&str) -> Option<String>;

impl KnowledgeConfig {
    pub fn from_json(text: &str) -> Result<Self, RagError> {
        serde_json::from_str(text).map_err(|e| RagError::Config(format!("invalid knowledge config: {e}")))
    }

    #[cfg(feature = "yaml")]
    pub fn from_yaml(text: &str) -> Result<Self, RagError> {
        let options = serde_saphyr::options! { strict_booleans: true };
        serde_saphyr::from_str_with_options(text, options)
            .map_err(|e| RagError::Config(format!("invalid knowledge config: {e}")))
    }

    /// Build the knowledge base, resolving secrets with `std::env::var`.
    pub fn build(&self) -> Result<KnowledgeBase, RagError> {
        self.build_with_env(&|name| std::env::var(name).ok())
    }

    /// Build the knowledge base with a custom environment lookup.
    pub fn build_with_env(&self, env: EnvLookup<'_>) -> Result<KnowledgeBase, RagError> {
        let mut embedders: IndexMap<&str, Arc<dyn Embedder>> = IndexMap::new();
        for (name, raw) in &self.embedders {
            embedders.insert(name, build_embedder(name, raw, env)?);
        }
        let mut rerankers: IndexMap<&str, Arc<dyn Reranker>> = IndexMap::new();
        for (name, raw) in &self.rerankers {
            rerankers.insert(name, build_reranker(name, raw, env)?);
        }
        let mut stores: IndexMap<&str, Arc<dyn VectorStore>> = IndexMap::new();
        for (name, raw) in &self.stores {
            stores.insert(name, build_store(name, raw, env)?);
        }

        let lookup_reranker = |key: &Option<String>| -> Result<Option<Arc<dyn Reranker>>, RagError> {
            match key {
                None => Ok(None),
                Some(k) => rerankers
                    .get(k.as_str())
                    .cloned()
                    .map(Some)
                    .ok_or_else(|| RagError::Config(format!("unknown reranker `{k}`"))),
            }
        };

        let mut kb = KnowledgeBase::new().with_reranker(lookup_reranker(&self.reranker)?);
        for ds in &self.datasets {
            if ds.id.trim().is_empty() {
                return Err(RagError::Config("dataset id must not be empty".into()));
            }
            if kb.contains(&ds.id) {
                return Err(RagError::Config(format!("duplicate dataset `{}`", ds.id)));
            }
            let info = DatasetInfo { id: ds.id.clone(), name: ds.name.clone(), description: ds.description.clone() };
            if let Some(raw) = &ds.retriever {
                kb.add_retriever(info, build_retriever(&ds.id, raw, env)?);
                continue;
            }
            let store_key = ds.store.as_deref().ok_or_else(|| {
                RagError::Config(format!("dataset `{}` needs either `store` + `embedder` or `retriever`", ds.id))
            })?;
            let store = stores
                .get(store_key)
                .cloned()
                .ok_or_else(|| RagError::Config(format!("unknown store `{store_key}`")))?;
            let embedder_key = match ds.embedder.as_deref() {
                Some(k) => k,
                None if embedders.len() == 1 => embedders.keys().next().copied().unwrap_or_default(),
                None => return Err(RagError::Config(format!("dataset `{}` needs `embedder`", ds.id))),
            };
            let embedder = embedders
                .get(embedder_key)
                .cloned()
                .ok_or_else(|| RagError::Config(format!("unknown embedder `{embedder_key}`")))?;
            kb.add_store_dataset(info, store, embedder, lookup_reranker(&ds.reranker)?);
        }
        Ok(kb)
    }
}

/// Keys accepted by every HTTP-backed provider entry.
#[cfg_attr(not(feature = "http"), allow(dead_code))]
const HTTP_KEYS: &[&str] = &["type", "api_key", "api_key_env", "api_key_header", "headers", "timeout_ms"];

/// Reject keys a provider entry does not understand (typos would otherwise be
/// ignored silently, e.g. `file:` instead of `path:` turning a persistent store
/// into an in-memory one).
fn check_keys(name: &str, raw: &serde_json::Value, allowed: &[&str], with_http: bool) -> Result<(), RagError> {
    let obj = raw.as_object().ok_or_else(|| RagError::Config(format!("`{name}` must be an object")))?;
    for key in obj.keys() {
        let known = allowed.contains(&key.as_str()) || (with_http && HTTP_KEYS.contains(&key.as_str()));
        if !known && key != "type" {
            return Err(RagError::Config(format!("`{name}`: unknown key `{key}`")));
        }
    }
    Ok(())
}

fn provider_type(name: &str, raw: &serde_json::Value) -> Result<String, RagError> {
    raw.get("type")
        .and_then(|t| t.as_str())
        .map(str::to_owned)
        .ok_or_else(|| RagError::Config(format!("`{name}` needs a `type`")))
}

/// Strip `type` / `api_key_env`, reject inline secrets, and inject the resolved key.
#[cfg_attr(not(feature = "http"), allow(dead_code))]
fn with_secret(name: &str, raw: &serde_json::Value, env: EnvLookup<'_>) -> Result<serde_json::Value, RagError> {
    let mut obj = raw.as_object().cloned().ok_or_else(|| RagError::Config(format!("`{name}` must be an object")))?;
    obj.remove("type");
    if obj.contains_key("api_key") {
        return Err(RagError::Config(format!(
            "`{name}`: inline `api_key` is not allowed; use `api_key_env` with an environment variable name"
        )));
    }
    if let Some(var) = obj.remove("api_key_env") {
        let var = var.as_str().ok_or_else(|| RagError::Config(format!("`{name}`: `api_key_env` must be a string")))?;
        let value =
            env(var).ok_or_else(|| RagError::Config(format!("`{name}`: environment variable `{var}` is not set")))?;
        obj.insert("api_key".into(), serde_json::Value::String(value));
    }
    Ok(serde_json::Value::Object(obj))
}

#[cfg_attr(not(feature = "http"), allow(dead_code))]
fn parse<T: serde::de::DeserializeOwned>(name: &str, value: serde_json::Value) -> Result<T, RagError> {
    serde_json::from_value(value).map_err(|e| RagError::Config(format!("`{name}`: {e}")))
}

fn build_embedder(name: &str, raw: &serde_json::Value, env: EnvLookup<'_>) -> Result<Arc<dyn Embedder>, RagError> {
    let kind = provider_type(name, raw)?;
    match kind.as_str() {
        #[cfg(feature = "http")]
        "openai" | "openai-compatible" | "ollama" => {
            check_keys(
                name,
                raw,
                &["base_url", "model", "dimensions", "batch_size", "query_prefix", "document_prefix"],
                true,
            )?;
            let mut value = with_secret(name, raw, env)?;
            if kind == "ollama" {
                if let Some(obj) = value.as_object_mut() {
                    obj.entry("base_url").or_insert_with(|| "http://localhost:11434/v1".into());
                }
            }
            let cfg: crate::http::OpenAiEmbedderConfig = parse(name, value)?;
            Ok(Arc::new(crate::http::OpenAiEmbedder::new(cfg)?))
        }
        "hash" => {
            check_keys(name, raw, &["dimensions"], false)?;
            let dims = match raw.get("dimensions") {
                None => 256,
                Some(v) => v
                    .as_u64()
                    .and_then(|d| usize::try_from(d).ok())
                    .filter(|d| *d > 0)
                    .ok_or_else(|| RagError::Config(format!("`{name}`: `dimensions` must be a positive integer")))?,
            };
            let _ = env;
            Ok(Arc::new(crate::hash::HashEmbedder::new(dims)))
        }
        other => Err(RagError::Config(format!("embedder `{name}`: unsupported type `{other}` in this build"))),
    }
}

fn build_reranker(name: &str, raw: &serde_json::Value, env: EnvLookup<'_>) -> Result<Arc<dyn Reranker>, RagError> {
    let kind = provider_type(name, raw)?;
    match kind.as_str() {
        #[cfg(feature = "http")]
        "http" | "cohere" | "jina" | "voyage" => {
            check_keys(name, raw, &["base_url", "model"], true)?;
            let cfg: crate::http::HttpRerankerConfig = parse(name, with_secret(name, raw, env)?)?;
            Ok(Arc::new(crate::http::HttpReranker::new(cfg)?))
        }
        other => {
            let _ = env;
            Err(RagError::Config(format!("reranker `{name}`: unsupported type `{other}` in this build")))
        }
    }
}

fn build_store(name: &str, raw: &serde_json::Value, env: EnvLookup<'_>) -> Result<Arc<dyn VectorStore>, RagError> {
    let kind = provider_type(name, raw)?;
    match kind.as_str() {
        "memory" => {
            check_keys(name, raw, &["path"], false)?;
            let path = match raw.get("path") {
                None => None,
                Some(serde_json::Value::String(p)) => Some(p.as_str()),
                Some(_) => return Err(RagError::Config(format!("store `{name}`: `path` must be a string"))),
            };
            match path {
                #[cfg(feature = "fs")]
                Some(path) => Ok(Arc::new(crate::memory::MemoryStore::open(path)?)),
                #[cfg(not(feature = "fs"))]
                Some(_) => Err(RagError::Config(format!("store `{name}`: `path` requires the `fs` feature"))),
                None => Ok(Arc::new(crate::memory::MemoryStore::new())),
            }
        }
        #[cfg(feature = "sqlite")]
        "sqlite" => {
            check_keys(name, raw, &["path"], false)?;
            let path =
                raw.get("path").and_then(serde_json::Value::as_str).filter(|p| !p.trim().is_empty()).ok_or_else(
                    || RagError::Config(format!("store `{name}`: `path` (the database file) is required")),
                )?;
            Ok(Arc::new(crate::sqlite::SqliteStore::open(path)?))
        }
        #[cfg(feature = "qdrant")]
        "qdrant" => {
            check_keys(name, raw, &["url", "collection", "create_collection"], true)?;
            let mut value = with_secret(name, raw, env)?;
            if let Some(obj) = value.as_object_mut() {
                obj.entry("api_key_header").or_insert_with(|| "api-key".into());
            }
            let cfg: crate::qdrant::QdrantConfig = parse(name, value)?;
            Ok(Arc::new(crate::qdrant::QdrantStore::new(cfg)?))
        }
        other => {
            let _ = env;
            Err(RagError::Config(format!("store `{name}`: unsupported type `{other}` in this build")))
        }
    }
}

fn build_retriever(name: &str, raw: &serde_json::Value, env: EnvLookup<'_>) -> Result<Arc<dyn Retriever>, RagError> {
    let kind = provider_type(name, raw)?;
    match kind.as_str() {
        #[cfg(feature = "http")]
        "http" => {
            check_keys(name, raw, &["url"], true)?;
            let cfg: crate::http::HttpRetrieverConfig = parse(name, with_secret(name, raw, env)?)?;
            Ok(Arc::new(crate::http::HttpRetriever::new(cfg)?))
        }
        other => {
            let _ = env;
            Err(RagError::Config(format!("retriever for `{name}`: unsupported type `{other}` in this build")))
        }
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::hash::HashEmbedder;
    use crate::ingest::Document;
    use crate::memory::MemoryStore;

    fn kb_with_two_datasets() -> KnowledgeBase {
        let embedder: Arc<dyn Embedder> = Arc::new(HashEmbedder::new(128));
        KnowledgeBase::new()
            .with_store_dataset("pets", Arc::new(MemoryStore::new()), embedder.clone(), None)
            .with_store_dataset("space", Arc::new(MemoryStore::new()), embedder, None)
    }

    async fn ingest(kb: &KnowledgeBase, ds: &str, id: &str, text: &str) {
        kb.ingestor(ds).unwrap().ingest(&Document::new(ds, id, text)).await.unwrap();
    }

    #[tokio::test]
    async fn multi_dataset_merge_and_single_mode() {
        let kb = kb_with_two_datasets();
        ingest(&kb, "pets", "cats", "Cats are small furry pets that purr.").await;
        ingest(&kb, "pets", "dogs", "Dogs are loyal pets that bark.").await;
        ingest(&kb, "space", "mars", "Mars is the red planet in our solar system.").await;

        let req = RetrievalRequest::new("which pets purr", vec!["pets".into(), "space".into()]).top_k(2);
        let hits = kb.retrieve(&req).await.unwrap();
        assert_eq!(hits.len(), 2);
        assert_eq!(hits[0].document_id, "cats");

        let single = kb.retrieve_single(&req).await.unwrap();
        assert!(single.iter().all(|h| h.dataset_id == "pets"));

        let unknown = RetrievalRequest::new("x", vec!["nope".into()]);
        assert!(matches!(kb.retrieve(&unknown).await, Err(RagError::NotFound(_))));
    }

    #[test]
    fn config_rejects_inline_secrets_and_resolves_env() {
        let cfg = KnowledgeConfig::from_json(
            r#"{"embedders":{"e":{"type":"hash","dimensions":16}},
                "stores":{"m":{"type":"memory"}},
                "datasets":[{"id":"a","store":"m"}]}"#,
        )
        .unwrap();
        let kb = cfg.build_with_env(&|_| None).unwrap();
        assert!(kb.contains("a"));

        let bad = KnowledgeConfig::from_json(
            r#"{"rerankers":{"r":{"type":"cohere","base_url":"http://x","model":"m","api_key":"inline"}}}"#,
        )
        .unwrap();
        let err = bad.build_with_env(&|_| None).unwrap_err();
        assert!(err.to_string().contains("inline") || err.to_string().contains("unsupported"), "{err}");

        let missing = KnowledgeConfig::from_json(r#"{"datasets":[{"id":"a","store":"nope"}]}"#).unwrap();
        assert!(missing.build_with_env(&|_| None).is_err());
    }

    #[cfg(feature = "http")]
    #[test]
    fn config_env_secret_lookup() {
        let cfg = KnowledgeConfig::from_json(
            r#"{"embedders":{"e":{"type":"openai","base_url":"http://127.0.0.1:9","model":"m","api_key_env":"MY_KEY"}}}"#,
        )
        .unwrap();
        assert!(cfg.build_with_env(&|_| None).unwrap_err().to_string().contains("MY_KEY"));
        assert!(cfg.build_with_env(&|n| (n == "MY_KEY").then(|| "secret".to_owned())).is_ok());
    }

    #[cfg(feature = "yaml")]
    #[test]
    fn yaml_config() {
        let cfg = KnowledgeConfig::from_yaml(
            "embedders:\n  e: { type: hash }\nstores:\n  m: { type: memory }\ndatasets:\n  - { id: kb, store: m, embedder: e }\n",
        )
        .unwrap();
        assert_eq!(cfg.datasets.len(), 1);
        assert!(cfg.build_with_env(&|_| None).is_ok());
    }

    #[cfg(feature = "sqlite")]
    #[tokio::test]
    async fn sqlite_store_from_config_persists_between_builds() {
        let dir = std::env::temp_dir().join(format!("graphrt-rag-kb-sqlite-{}", std::process::id()));
        let _ = std::fs::remove_dir_all(&dir);
        std::fs::create_dir_all(&dir).unwrap();
        let path = dir.join("kb.db");
        let config = KnowledgeConfig::from_json(
            &serde_json::json!({
                "embedders": { "e": { "type": "hash", "dimensions": 64 } },
                "stores": { "local": { "type": "sqlite", "path": path } },
                "datasets": [{ "id": "faq", "store": "local", "embedder": "e" }]
            })
            .to_string(),
        )
        .unwrap();

        let first = config.build_with_env(&|_| None).unwrap();
        ingest(&first, "faq", "returns", "Items can be returned within 30 days.").await;
        drop(first);

        let second = config.build_with_env(&|_| None).unwrap();
        let hits = second.retrieve(&RetrievalRequest::new("returned within days", vec!["faq".into()])).await.unwrap();
        assert_eq!(hits.first().map(|h| h.document_id.as_str()), Some("returns"));
        let _ = std::fs::remove_dir_all(&dir);
    }

    #[cfg(feature = "sqlite")]
    #[test]
    fn sqlite_store_requires_a_path() {
        let config = KnowledgeConfig::from_json(r#"{"stores":{"local":{"type":"sqlite"}}}"#).unwrap();
        let error = config.build_with_env(&|_| None).unwrap_err();
        assert!(error.to_string().contains("`path` (the database file) is required"), "{error}");
    }
}