knot 1.9.7

Codebase Graph + Vector RAG Indexer for Java, TypeScript, JavaScript, Kotlin, Rust, Python, Groovy, C/C++, C#, Build Systems, and HTML/CSS codebases
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
//! Core search_hybrid_context logic shared between CLI and MCP
//!
//! Performs a hybrid search combining:
//! 1. Prefix name match via Neo4j (exact name prefix, case-insensitive)
//! 2. Semantic search via Qdrant vector similarity (understands code meaning)
//! 3. Kind-aware re-ranking (see [`rank`] — definitions outrank prose, tests
//!    and config/build entities for natural-language queries)
//! 4. Structural expansion via Neo4j graph relationships (understands architecture)
//!
//! Graph enrichment only *annotates* the entities already returned; it never
//! injects callers or helpers as substitute results.

use serde_json::json;
use std::collections::HashSet;
use std::path::Path;
use std::sync::{Arc, Mutex};

use crate::db::{
    graph::{DEFAULT_MAX_TARGETS, GraphDb, QueryExt},
    vector::{ExactNameProbe, UuidProbe, VectorDb, VectorSearchExt},
};
use crate::models::RepoScope;
use crate::pipeline::embed::Embedder;

pub mod rank;

/// Default result count when the caller does not ask for one.
/// Must equal the `default` advertised by the MCP schema
/// ([`crate::mcp_tools::search_hybrid_context`]).
pub const DEFAULT_MAX_RESULTS: usize = 5;

/// Hard ceiling for `max_results`. Must equal the `maximum` advertised by the
/// MCP schema — the drift guard test in `mcp_tools::search_hybrid_context`
/// pins the two together. There is no cursor or pagination: past this bound,
/// callers narrow the search with `kinds` / `path` / `repo_name` or refine
/// the query instead of raising the limit.
pub const MAX_RESULTS_CEILING: usize = 100;

/// A caller-requested result count resolved against the advertised bound.
///
/// Resolved values are always clamped into `1..=MAX_RESULTS_CEILING`:
/// values within the bound pass through unchanged, zero/negative requests
/// floor at 1 (a result count of 0 would mean "return nothing at all"),
/// and requests above the ceiling clamp to it — the caller is told via
/// [`ResolvedLimit::notice`], never silently.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub struct ResolvedLimit {
    /// The enforced value to use for the search.
    pub value: usize,
    /// What the caller originally asked for.
    pub requested: usize,
}

impl ResolvedLimit {
    /// Whether the request exceeded the advertised ceiling (or floored).
    pub fn was_clamped(&self) -> bool {
        self.requested != self.value
    }

    /// Caller-facing note when the request had to be adjusted; `None` when
    /// the request was served as asked.
    pub fn notice(&self) -> Option<String> {
        if !self.was_clamped() {
            return None;
        }
        if self.requested == 0 {
            return Some("> Note: `max_results` was floored from 0 to the minimum of 1.\n".into());
        }
        Some(format!(
            "> Note: `max_results` was clamped from {requested} to the advertised maximum of {MAX_RESULTS_CEILING}. \
This tool has no pagination — narrow the search with `kinds` / `path` / `repo_name`, or refine the query.\n",
            requested = self.requested
        ))
    }
}

/// Single source of truth for the `max_results` bound — used by the shared
/// search core, by the MCP tool layer and by the CLI so the enforced value
/// can never drift from the advertised schema.
pub fn resolve_max_results(requested: usize) -> ResolvedLimit {
    ResolvedLimit {
        value: requested.clamp(1, MAX_RESULTS_CEILING),
        requested,
    }
}

/// Bundled database and embedder dependencies for [`run_search_hybrid_context`].
///
/// Holding the three shared handles in one struct keeps the public function
/// within clippy's `too_many_arguments` threshold and lets callers (CLI bin,
/// MCP tool) hand over the same handles without copying.
#[derive(Clone)]
pub struct SearchContext<'a> {
    pub vector_db: &'a Arc<VectorDb>,
    pub graph_db: &'a Arc<GraphDb>,
    pub embedder: &'a Arc<Mutex<Embedder>>,
}

/// User-supplied optional filters, bundled so the shared search entry
/// point stays within clippy's arity threshold.
pub struct SearchFilters<'a> {
    /// Comma-separated kind filter (exact wire-format kinds or aliases).
    pub kinds: Option<&'a str>,
    /// Repo-relative directory prefix or glob (`list_files` matcher).
    pub path: Option<&'a str>,
}

/// Main search function called by both CLI and MCP.
///
/// `filters` bundles the optional restrictions; absent `kinds`/`path`
/// fields mean no filtering.
pub async fn run_search_hybrid_context(
    query: &str,
    max_results: usize,
    repo: &RepoScope,
    filters: SearchFilters<'_>,
    ctx: &SearchContext<'_>,
) -> anyhow::Result<serde_json::Value> {
    // Enforce the advertised bound before anything else so no caller (CLI or
    // MCP) can pull more than `MAX_RESULTS_CEILING` rows out of the pipeline.
    let max_results = resolve_max_results(max_results).value;
    let vector = ctx
        .embedder
        .lock()
        .unwrap()
        .embed_query(query)
        .map_err(|e| anyhow::anyhow!("Embedding failed: {}", e))?;

    let repo_names = repo.filter_names();
    let expanded_kinds = rank::parse_kinds(filters.kinds);
    let normalized_path = normalized_path(filters.path);

    // Path-restricted searches oversample: a large share of the cosine
    // window may fall outside the path, so the re-rank needs a wider pool
    // to fill `max_results` from matching files alone. The oversample is
    // capped so the widened pool cannot turn into an oversized Qdrant
    // round-trip at the top of the `max_results` range.
    const PATH_POOL_CEILING: usize = 600;
    let pool_window = if normalized_path.is_some() {
        rank::candidate_limit(max_results)
            .saturating_mul(3)
            .min(PATH_POOL_CEILING)
    } else {
        rank::candidate_limit(max_results)
    };

    let search_results = ctx
        .vector_db
        .search(&vector, pool_window, &repo_names, &expanded_kinds)
        .await?;

    // Boost: prepend entities whose name matches the query as a case-insensitive prefix
    let prefix_results = ctx
        .graph_db
        .find_entities_by_name_prefix(query, &repo_names, max_results)
        .await
        .unwrap_or_else(|_| json!([]));

    let mut seen_uuids: HashSet<String> = HashSet::new();
    let mut combined: Vec<serde_json::Value> = Vec::new();

    // Prefix hits keep the leading slots (name-match contract), filtered by
    // the kind and path filters when one is set.
    merge_prefix_hits(
        &prefix_results,
        &expanded_kinds,
        normalized_path.as_deref(),
        &mut seen_uuids,
        &mut combined,
    );

    // Vector hits: dedup, name/token probe recall and the caller bridge,
    // merged into one candidate pool for the kind-aware re-rank; the path
    // filter applies to every channel when one is set.
    let pool = CandidatePool {
        vector: &vector,
        query,
        search_results: &search_results,
        seen_uuids: &mut seen_uuids,
        repo_names: &repo_names,
        expanded_kinds: &expanded_kinds,
        candidate_limit: pool_window,
        ctx,
    };
    let mut vector_hits = pool.collect().await;
    if let Some(pattern) = &normalized_path {
        vector_hits.retain(|hit| {
            rank::path_allows(
                Some(pattern),
                hit.get("file_path").and_then(|v| v.as_str()).unwrap_or(""),
            )
        });
    }

    combined.extend(rank::rerank(vector_hits, query));

    if combined.is_empty() {
        return Ok(serde_json::Value::Null);
    }

    combined.truncate(max_results);

    let uuids: Vec<String> = combined
        .iter()
        .map(|entity| {
            entity
                .get("uuid")
                .and_then(|v| v.as_str())
                .unwrap_or("")
                .to_string()
        })
        .collect();

    let fetched = ctx
        .graph_db
        .get_entities_with_dependencies(&uuids, &repo_names)
        .await?;

    // `get_entities_with_dependencies` preserves uuid order, so the ranked
    // order survives the round-trip; drop any duplicate rows defensively.
    let context = serde_json::Value::Array(rank::dedup_by_identity(
        fetched.as_array().cloned().unwrap_or_default(),
    ));

    let entity_names: Vec<String> = context
        .as_array()
        .map(|entities| {
            entities
                .iter()
                .filter_map(|entity| {
                    entity
                        .get("name")
                        .and_then(|v| v.as_str())
                        .map(String::from)
                })
                .collect()
        })
        .unwrap_or_default();

    let enriched_context = enrich_with_relationships(&context, &entity_names, ctx.graph_db, repo)
        .await
        .unwrap_or(context);

    Ok(enriched_context)
}

/// Enrich search results with related entities (subclasses, implementers, usages).
///
/// Annotation-only: relationship data is attached to entities that are
/// already in the result list. Callers and helpers are shown as context of a
/// definition, never as substitute rows displacing it.
async fn enrich_with_relationships(
    context: &serde_json::Value,
    _entity_names: &[String],
    graph_db: &Arc<GraphDb>,
    repo: &RepoScope,
) -> anyhow::Result<serde_json::Value> {
    let mut enriched = context.clone();
    let repo_names = repo.filter_names();

    if let Some(entities) = enriched.as_array_mut() {
        for entity in entities.iter_mut() {
            if let Some(name) = entity.get("name").and_then(|v| v.as_str())
                && let Ok(references) = graph_db
                    // `kinds=all`: the call-site enrichment must keep seeing
                    // every reference edge regardless of the search's own
                    // kind filter semantics.
                    .find_references(name, &repo_names, DEFAULT_MAX_TARGETS, Some("all"))
                    .await
            {
                enrich_single_entity(entity, &references);
            }
        }
    }

    Ok(enriched)
}

/// Merge the raw vector hits with the recall channels into one candidate
/// pool, deduplicated against `seen_uuids` in place:
///
/// 1. the cosine search results themselves;
/// 2. the name/token probe — identifiers the query literally names, or the
///    identifier words it shares, entering the pool with their true cosine
///    however deep their pure similarity rank (`LookupMaps::build` sits at
///    cosine rank ~500 for "build lookup maps for reference resolution");
/// 3. the caller-recall bridge — the callers of the top semantic hits.
///
/// All recall channels are best effort: a failure drops the channel, never
/// the whole search.
pub(crate) struct CandidatePool<'a> {
    /// Query embedding.
    pub vector: &'a [f32],
    /// Raw query text (feeds the name/token probe).
    pub query: &'a str,
    /// Cosine search results, in pure similarity order.
    pub search_results: &'a [serde_json::Value],
    /// UUIDs already merged (prefix hits); advanced in place.
    pub seen_uuids: &'a mut HashSet<String>,
    /// Repository scope (empty = all).
    pub repo_names: &'a [String],
    /// Wire-format entity kinds to restrict hits to (empty = all).
    pub expanded_kinds: &'a [String],
    /// Over-fetch window size (also caps the caller bridge).
    pub candidate_limit: usize,
    /// Bounded shared handles (vector + graph + embedder).
    pub ctx: &'a SearchContext<'a>,
}

impl CandidatePool<'_> {
    pub async fn collect(self) -> Vec<serde_json::Value> {
        let mut vector_hits: Vec<serde_json::Value> = Vec::new();
        merge_hits(self.search_results, self.seen_uuids, &mut vector_hits);

        let probe_hits = name_probe_hits(
            self.vector,
            self.query,
            self.repo_names,
            self.expanded_kinds,
            self.ctx,
        )
        .await;
        merge_hits(&probe_hits, self.seen_uuids, &mut vector_hits);

        let (bridge, top_roots) = CallerBridge {
            vector: self.vector,
            search_results: self.search_results,
            seen_uuids: self.seen_uuids,
            repo_names: self.repo_names,
            expanded_kinds: self.expanded_kinds,
            ctx: self.ctx,
        }
        .run(self.candidate_limit)
        .await;
        merge_hits(&bridge, self.seen_uuids, &mut vector_hits);

        // Root-set coverage + FQN for the *whole* pool — cosine hits (the
        // entry point is usually among them), probe hits and bridge hits
        // alike. This is what makes the re-rank's caller-root signal reach
        // entities the direct search already returned; the bridge alone
        // never annotated those.
        annotate_root_coverage(&mut vector_hits, &top_roots, self.repo_names, self.ctx).await;
        vector_hits
    }
}

/// Append not-yet-seen hits to the pool, updating the seen-UUID set.
fn merge_hits(
    hits: &[serde_json::Value],
    seen_uuids: &mut HashSet<String>,
    pool: &mut Vec<serde_json::Value>,
) {
    for hit in hits {
        if let Some(uuid) = hit.get("uuid").and_then(|v| v.as_str())
            && seen_uuids.insert(uuid.to_string())
        {
            pool.push(hit.clone());
        }
    }
}

/// Normalize the caller's `path` input with the shared `explore_file`
/// resolver (`None`/empty → no restriction).
fn normalized_path(path: Option<&str>) -> Option<String> {
    path.filter(|p| !p.trim().is_empty()).map(|p| {
        crate::cli_tools::list_files::normalize_list_path(
            p,
            std::env::current_dir().ok().as_deref(),
            std::env::var("KNOT_REPO_PATH")
                .ok()
                .as_deref()
                .map(Path::new),
        )
    })
}

/// Merge name-prefix hits into the leading slots, honoring the kind and
/// path filters (name-match contract: prefix hits keep their leading
/// slots only if they pass the user's restrictions).
fn merge_prefix_hits(
    prefix_results: &serde_json::Value,
    expanded_kinds: &[String],
    normalized_path: Option<&str>,
    seen_uuids: &mut HashSet<String>,
    combined: &mut Vec<serde_json::Value>,
) {
    if let Some(arr) = prefix_results.as_array() {
        for entity in arr {
            let kind = entity.get("kind").and_then(|v| v.as_str()).unwrap_or("");
            let file_path = entity
                .get("file_path")
                .and_then(|v| v.as_str())
                .unwrap_or("");
            if rank::kinds_allow(expanded_kinds, kind)
                && rank::path_allows(normalized_path, file_path)
            {
                push_if_unique(entity, seen_uuids, combined);
            }
        }
    }
}

/// Run the name/token probe against Qdrant: entities whose name equals a
/// significant query token (lowercase + PascalCase variants, ≤ 4 tokens)
/// or whose identifier shares a query token, carrying their true cosine so
/// the re-rank can promote them. Best effort — fails to an empty list.
async fn name_probe_hits(
    vector: &[f32],
    query: &str,
    repo_names: &[String],
    expanded_kinds: &[String],
    ctx: &SearchContext<'_>,
) -> Vec<serde_json::Value> {
    let probe_tokens = rank::significant_query_tokens(query);
    if probe_tokens.is_empty() {
        return Vec::new();
    }
    let probe_names = rank::probe_name_variants(&probe_tokens);
    ctx.vector_db
        .search_exact_names(ExactNameProbe {
            vector,
            names: &probe_names,
            tokens: &probe_tokens,
            limit: 24,
            repo_names,
            kinds: expanded_kinds,
        })
        .await
        .unwrap_or_default()
}

/// Caller-recall bridge over the top semantic hits.
///
/// Roots are the highest-cosine vector hits ([`TOP_ROOTS`] capped); their
/// callers (in-repo, capped at [`CALLER_LINK_LIMIT`] Cypher rows) are scored
/// against the query vector in one Qdrant round-trip and returned as full
/// candidate rows with their true cosine. The bridge is **pure recall**: it
/// only adds unseen callers to the pool; the scoring signal (how many top
/// roots each candidate calls, directly or through one helper) is attached
/// to *every* pool row afterwards by [`annotate_root_coverage`], including
/// cosine hits the bridge never fetched (the entry point is usually already
/// in the cosine pool itself).
///
/// Skips prose and test-path roots — a Markdown section has no callers and
/// a test's callers are other tests. Failures collapse to an empty list;
/// the search continues without the bridge.
pub(crate) struct CallerBridge<'a> {
    /// Query embedding.
    pub vector: &'a [f32],
    /// Vector hits in pure cosine order — the bridge seeds from the top.
    pub search_results: &'a [serde_json::Value],
    /// UUIDs already in the candidate pool (skipped by the bridge).
    pub seen_uuids: &'a HashSet<String>,
    /// Repository scope (empty = all).
    pub repo_names: &'a [String],
    /// Wire-format entity kinds to restrict hits to (empty = all).
    pub expanded_kinds: &'a [String],
    /// Bounded shared handles (vector + graph databases).
    pub ctx: &'a SearchContext<'a>,
}

impl CallerBridge<'_> {
    /// Return `(bridge_hits, seeded_roots)`. `seeded_roots` is the UUID list
    /// the coverage annotation needs; `bridge_hits` is capped by the
    /// remaining pool capacity so the bridge can never displace what the
    /// direct search already gathered. Every hit is a genuine caller of at
    /// least one seeded root (the annotation pass re-checks nothing — the
    /// retain below already did).
    pub async fn run(&self, pool_cap: usize) -> (Vec<serde_json::Value>, Vec<String>) {
        let top_roots: Vec<String> = self
            .search_results
            .iter()
            .filter(|hit| {
                let kind = hit.get("kind").and_then(|v| v.as_str()).unwrap_or("");
                let file_path = hit.get("file_path").and_then(|v| v.as_str()).unwrap_or("");
                rank::kind_boost(kind) >= 0.0 && !rank::is_test_path(file_path)
            })
            .filter_map(|hit| hit.get("uuid").and_then(|v| v.as_str()))
            .take(TOP_ROOTS)
            .map(String::from)
            .collect();
        if top_roots.is_empty() {
            return (Vec::new(), top_roots);
        }

        let links = self
            .ctx
            .graph_db
            .find_caller_links(&top_roots, self.repo_names, CALLER_LINK_LIMIT)
            .await
            .unwrap_or_default();
        if links.is_empty() {
            return (Vec::new(), top_roots);
        }

        let caller_uuids: HashSet<String> =
            links.iter().map(|(caller, _)| caller.clone()).collect();
        let fetch_list: Vec<String> = caller_uuids
            .iter()
            .filter(|u| !self.seen_uuids.contains(*u))
            .cloned()
            .collect();
        if fetch_list.is_empty() {
            return (Vec::new(), top_roots);
        }

        let mut hits = self
            .ctx
            .vector_db
            .search_by_uuids(UuidProbe {
                vector: self.vector,
                uuids: &fetch_list,
                repo_names: self.repo_names,
                kinds: self.expanded_kinds,
                limit: pool_cap,
            })
            .await
            .unwrap_or_default();

        // Keep only rows that are genuine callers of a seeded root — Qdrant
        // keyword matching can return a point whose payload `uuid` differs
        // in case, and an annotated boost must always rest on real CALL
        // edges.
        hits.retain(|hit| {
            hit.get("uuid")
                .and_then(|v| v.as_str())
                .is_some_and(|u| caller_uuids.contains(u))
        });
        (hits, top_roots)
    }
}

/// Annotate every pool candidate with the graph evidence the re-rank needs:
/// its FQN (absent from the Qdrant payload, required by the generic-name
/// guard in [`rank`]) and how many of the top semantic roots it calls
/// directly / through exactly one helper ([`rank::root_coverage_boost`]
/// turns both counts into the root-set signal). One bounded round trip.
///
/// Best effort: a failed round trip leaves the pool unannotated and the
/// search proceeds on cosine + kind + lexical alone.
async fn annotate_root_coverage(
    pool: &mut [serde_json::Value],
    top_roots: &[String],
    repo_names: &[String],
    ctx: &SearchContext<'_>,
) {
    if pool.is_empty() || top_roots.is_empty() {
        return;
    }
    let pool_uuids: Vec<String> = pool
        .iter()
        .filter_map(|hit| hit.get("uuid").and_then(|v| v.as_str()))
        .map(String::from)
        .collect();
    if pool_uuids.is_empty() {
        return;
    }
    let coverage = ctx
        .graph_db
        .fetch_root_coverage(&pool_uuids, top_roots, repo_names, pool_uuids.len())
        .await
        .unwrap_or_default();
    if coverage.is_empty() {
        return;
    }
    let by_uuid: std::collections::HashMap<&str, &crate::db::graph::RootCoverage> = coverage
        .iter()
        .map(|row| (row.uuid.as_str(), row))
        .collect();

    // Who calls whom *inside the pool* shadows the coverage signal: a
    // candidate whose pool caller covers at least the same root set is an
    // internal step of that caller, so its provenance is redundant — the
    // boost is halved ([`rank::ORCHESTRATOR_STEP_ATTENUATION`]). Only a
    // production caller reassigns provenance: a test calling the same
    // helpers must not take the entry point's boost away (the test-path
    // guard of the coverage boost applies to the caller too).
    let pool_meta: std::collections::HashMap<&str, (&str, &str)> = pool
        .iter()
        .filter_map(|hit| {
            Some((
                hit.get("uuid")?.as_str()?,
                (hit.get("kind")?.as_str()?, hit.get("file_path")?.as_str()?),
            ))
        })
        .collect();
    let in_pool: HashSet<&str> = pool_uuids.iter().map(String::as_str).collect();
    let mut superseded: HashSet<String> = HashSet::new();
    if let Ok(links) = ctx
        .graph_db
        .find_caller_links(&pool_uuids, repo_names, pool_uuids.len() * 16)
        .await
    {
        for (caller, callee) in links.iter().filter(|(caller, callee)| {
            in_pool.contains(caller.as_str())
                && in_pool.contains(callee.as_str())
                && caller != callee
        }) {
            let (Some(caller_cov), Some(callee_cov)) =
                (by_uuid.get(caller.as_str()), by_uuid.get(callee.as_str()))
            else {
                continue;
            };
            let caller_production =
                pool_meta
                    .get(caller.as_str())
                    .is_some_and(|(kind, file_path)| {
                        rank::kind_boost(kind) >= 0.0 && !rank::is_test_path(file_path)
                    });
            if caller_production && caller_cov.direct_roots >= callee_cov.direct_roots {
                superseded.insert(callee.clone());
            }
        }
    }

    for hit in pool.iter_mut() {
        let uuid = hit.get("uuid").and_then(|v| v.as_str()).unwrap_or("");
        let Some(row) = by_uuid.get(uuid) else {
            continue;
        };
        let superseded_here = superseded.contains(uuid);
        let Some(obj) = hit.as_object_mut() else {
            continue;
        };
        // The pool already carries a `fqn` when the row came from a channel
        // that enriched it from the graph; only fill the Qdrant gap.
        obj.entry("fqn")
            .or_insert_with(|| serde_json::json!(row.fqn));
        obj.insert(
            "caller_roots".to_string(),
            serde_json::json!(row.direct_roots),
        );
        obj.insert(
            "caller_roots_transitive".to_string(),
            serde_json::json!(row.transitive_roots),
        );
        obj.insert(
            "caller_root_total".to_string(),
            serde_json::json!(top_roots.len()),
        );
        obj.insert(
            "caller_superseded".to_string(),
            serde_json::json!(superseded_here),
        );
    }
}

/// How many top vector hits seed the caller-recall bridge. Eight roots
/// cover the helpers a behavioral paraphrase ranks first even when its
/// shared caller's strongest helpers rank 4th–8th in pure cosine order
/// (three were not enough: the entry point then entered the pool with zero
/// coverage evidence). Seeding is capped by the pool window via `min` at
/// the call site (each seed must map to a real bridge request), and the
/// Neo4j fan-in stays bounded by [`CALLER_LINK_LIMIT`] on the Cypher side.
const TOP_ROOTS: usize = 8;

/// Neo4j row budget for the caller-recall bridge. Scaled with the widened
/// seed so a popular root cannot eat the whole pair list: the Cypher's
/// `ORDER BY caller_uuid, target_uuid LIMIT $limit` truncates by UUID
/// string, and a root followed by 30 callers would otherwise starve every
/// root sorted after it.
const CALLER_LINK_LIMIT: usize = TOP_ROOTS * 32;

fn push_if_unique(
    entity: &serde_json::Value,
    seen_uuids: &mut HashSet<String>,
    combined: &mut Vec<serde_json::Value>,
) {
    merge_hits(std::slice::from_ref(entity), seen_uuids, combined);
}

pub(crate) fn extract_subclass_names(extends_arr: &[serde_json::Value]) -> Vec<String> {
    extends_arr
        .iter()
        .filter_map(|e| e.get("name").and_then(|v| v.as_str()).map(String::from))
        .collect()
}

pub(crate) fn extract_implementer_names(implements_arr: &[serde_json::Value]) -> Vec<String> {
    implements_arr
        .iter()
        .filter_map(|e| e.get("name").and_then(|v| v.as_str()).map(String::from))
        .collect()
}

pub(crate) fn format_reference_samples(references_arr: &[serde_json::Value]) -> Vec<String> {
    references_arr
        .iter()
        .take(3)
        .map(|e| {
            let name = e.get("name").and_then(|v| v.as_str()).unwrap_or("?");
            let file = e.get("file_path").and_then(|v| v.as_str()).unwrap_or("?");
            format!("{} in {}", name, file)
        })
        .collect()
}

pub(crate) fn enrich_single_entity(entity: &mut serde_json::Value, references: &serde_json::Value) {
    if let Some(extends_arr) = references.get("extends").and_then(|v| v.as_array())
        && !extends_arr.is_empty()
    {
        let subclasses = extract_subclass_names(extends_arr);
        if let Some(obj) = entity.as_object_mut() {
            obj.insert("subclasses".to_string(), json!(subclasses));
        }
    }

    if let Some(implements_arr) = references.get("implements").and_then(|v| v.as_array())
        && !implements_arr.is_empty()
    {
        let implementers = extract_implementer_names(implements_arr);
        if let Some(obj) = entity.as_object_mut() {
            obj.insert("implementers".to_string(), json!(implementers));
        }
    }

    if let Some(references_arr) = references.get("references").and_then(|v| v.as_array())
        && !references_arr.is_empty()
    {
        let usage_count = references_arr.len();
        let samples = format_reference_samples(references_arr);
        if let Some(obj) = entity.as_object_mut() {
            obj.insert("type_usage_count".to_string(), json!(usage_count));
            obj.insert("type_usage_samples".to_string(), json!(samples));
        }
    }

    if let Some(calls_arr) = references.get("calls").and_then(|v| v.as_array())
        && !calls_arr.is_empty()
    {
        let caller_count = calls_arr.len();
        let samples = format_reference_samples(calls_arr);
        if let Some(obj) = entity.as_object_mut() {
            obj.insert("caller_count".to_string(), json!(caller_count));
            obj.insert("caller_samples".to_string(), json!(samples));
        }
    }
}

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

    // --- ResolvedLimit / resolve_max_results (advertised-bound contract) ---

    #[test]
    fn resolve_max_results_clamps_above_ceiling() {
        let resolved = resolve_max_results(10_000);
        assert_eq!(resolved.value, MAX_RESULTS_CEILING);
        assert_eq!(resolved.requested, 10_000);
        assert!(resolved.was_clamped());
    }

    #[test]
    fn resolve_max_results_preserves_values_within_bound() {
        for n in [1, DEFAULT_MAX_RESULTS, 20, MAX_RESULTS_CEILING] {
            let resolved = resolve_max_results(n);
            assert_eq!(resolved.value, n);
            assert!(!resolved.was_clamped());
            assert!(resolved.notice().is_none());
        }
    }

    #[test]
    fn resolve_max_results_floors_at_one() {
        // A result count of 0 would mean "return nothing at all"; floored.
        let resolved = resolve_max_results(0);
        assert_eq!(resolved.value, 1);
        assert!(resolved.was_clamped());
    }

    #[test]
    fn clamp_notice_states_bound_and_refine_rule() {
        let notice = resolve_max_results(500).notice().expect("clamped");
        assert!(notice.contains("500"));
        assert!(notice.contains(MAX_RESULTS_CEILING.to_string().as_str()));
        assert!(notice.contains("kinds"));
        assert!(notice.contains("path"));
        assert!(notice.contains("no pagination"));
    }

    #[test]
    fn extract_subclass_names_empty() {
        let refs = vec![];
        let names = extract_subclass_names(&refs);
        assert_eq!(names.len(), 0);
    }

    #[test]
    fn test_extract_subclass_names_with_data() {
        let refs = vec![
            json!({"name": "ChildClass1", "kind": "class"}),
            json!({"name": "ChildClass2", "kind": "class"}),
        ];
        let names = extract_subclass_names(&refs);
        assert_eq!(names.len(), 2);
        assert!(names.contains(&"ChildClass1".to_string()));
        assert!(names.contains(&"ChildClass2".to_string()));
    }

    #[test]
    fn test_extract_implementer_names() {
        let refs = vec![
            json!({"name": "ImplClass1", "kind": "class"}),
            json!({"name": "ImplClass2", "kind": "class"}),
        ];
        let names = extract_implementer_names(&refs);
        assert_eq!(names.len(), 2);
        assert!(names.contains(&"ImplClass1".to_string()));
    }

    #[test]
    fn test_format_reference_samples_limits_to_three() {
        let refs = vec![
            json!({"name": "usage1", "file_path": "file1.java"}),
            json!({"name": "usage2", "file_path": "file2.java"}),
            json!({"name": "usage3", "file_path": "file3.java"}),
            json!({"name": "usage4", "file_path": "file4.java"}),
            json!({"name": "usage5", "file_path": "file5.java"}),
        ];
        let samples = format_reference_samples(&refs);
        assert_eq!(samples.len(), 3);
        assert!(samples[0].contains("usage1"));
        assert!(samples[1].contains("usage2"));
        assert!(samples[2].contains("usage3"));
    }

    #[test]
    fn test_format_reference_samples() {
        let refs = vec![json!({"name": "caller1", "file_path": "caller.java"})];
        let samples = format_reference_samples(&refs);
        assert_eq!(samples.len(), 1);
        assert!(samples[0].contains("caller1"));
        assert!(samples[0].contains("caller.java"));
    }

    #[test]
    fn test_enrich_single_entity_with_subclasses() {
        let mut entity = json!({"name": "MyClass", "kind": "class"});
        let references = json!({
            "extends": [
                {"name": "Child1", "kind": "class"},
                {"name": "Child2", "kind": "class"}
            ],
            "implements": [],
            "references": [],
            "calls": []
        });

        enrich_single_entity(&mut entity, &references);

        assert_eq!(
            entity.get("subclasses"),
            Some(&json!(vec!["Child1", "Child2"]))
        );
    }

    #[test]
    fn test_enrich_single_entity_with_all_relationships() {
        let mut entity = json!({"name": "MyInterface"});
        let references = json!({
            "extends": [{"name": "Child1"}],
            "implements": [{"name": "Impl1"}, {"name": "Impl2"}],
            "references": [
                {"name": "ref1", "file_path": "ref1.java"},
                {"name": "ref2", "file_path": "ref2.java"}
            ],
            "calls": [{"name": "caller1", "file_path": "caller.java"}]
        });

        enrich_single_entity(&mut entity, &references);

        assert!(entity.get("subclasses").is_some());
        assert!(entity.get("implementers").is_some());
        assert!(entity.get("type_usage_count").is_some());
        assert!(entity.get("caller_count").is_some());
    }

    #[test]
    fn test_enrich_single_entity_ignores_empty_arrays() {
        let mut entity = json!({"name": "MyClass"});
        let references = json!({
            "extends": [],
            "implements": [],
            "references": [],
            "calls": []
        });

        enrich_single_entity(&mut entity, &references);

        assert!(entity.get("subclasses").is_none());
        assert!(entity.get("implementers").is_none());
        assert!(entity.get("type_usage_count").is_none());
        assert!(entity.get("caller_count").is_none());
    }
}