second-brain-api 0.13.5

HTTP API server for second-brain: REST endpoints for recall, remember, and ingest
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
pub mod bootstrap;
pub mod caching_store;
pub mod metrics;

use std::collections::HashSet;
use std::io::Write;
use std::path::Path;
use std::sync::Mutex;
use std::sync::atomic::{AtomicUsize, Ordering};
use std::thread;

use anyhow::Result;
use serde::{Deserialize, Serialize};
use uuid::Uuid;

use second_brain_core::embedding::Embedder;
use second_brain_core::index::Indexes;
use second_brain_core::query::{QueryEngine, QueryFilters, QueryRequest};
use second_brain_core::retrieval::{RerankScorer, Reranker, RetrievalTunables, Retriever};
use second_brain_core::sqlite_store::SqliteStore;
use second_brain_core::store::{Store, StoreAdmin};

/// Index-backed pieces the hybrid engines need beside the (possibly caching)
/// Store: candidates come from the real indexes over the snapshot SqliteStore,
/// ranking runs against the caching wrapper.
pub struct HybridContext<'a> {
    pub sqlite: &'a SqliteStore,
    pub indexes: &'a Indexes,
    pub reranker: Option<&'a Reranker>,
    pub tunables: RetrievalTunables,
}

#[derive(Debug, Clone, Deserialize, Serialize)]
pub struct EvalQuery {
    pub query_id: String,
    pub query: String,
    pub query_variant: String,
    pub seed_memory_id: Uuid,
    pub memory_type: String,
    pub relevant_memory_ids: Vec<Uuid>,
    #[serde(default)]
    pub note: Option<String>,
    #[serde(default)]
    pub tags: Vec<String>,
    /// Project the gold memory is filed under.
    #[serde(default)]
    pub source_project: Option<String>,
}

#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct QueryRecord {
    pub query_id: String,
    pub use_prefix: bool,
    pub ranked_ids: Vec<Uuid>,
    pub scores: Vec<f32>,
    pub first_relevant_rank: Option<usize>,
    pub gold_raw_rank: Option<usize>,
    pub gold_raw_similarity: Option<f32>,
    /// Engine-call wall time only. Under run_arm_parallel this includes
    /// rerank-mutex wait, so solo latency must come from a serial run.
    #[serde(default)]
    pub latency_ms: Option<u64>,
}

#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct ArmMetrics {
    pub recall_at_1: f32,
    pub recall_at_3: f32,
    pub recall_at_5: f32,
    pub mrr: f32,
    pub precision_at_5: f32,
}

#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct AggregateReport {
    pub bare: ArmMetrics,
    pub prefixed: ArmMetrics,
    pub delta_recall_at_3_ci: (f32, f32),
    pub delta_mrr_ci: (f32, f32),
    pub gated_out_rate_bare: f32,
    pub gated_out_rate_prefixed: f32,
}

#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct GatePoint {
    pub threshold: f32,
    pub recall_at_1: f32,
    pub recall_at_3: f32,
    pub recall_at_5: f32,
    pub precision_proxy: f32,
}

#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct GateSweepReport {
    pub frontier: Vec<GatePoint>,
    pub baseline_threshold: f32,
    pub chosen_threshold: f32,
    pub chosen_beats_baseline: bool,
}

#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct CorpusEntry {
    pub id: Uuid,
    pub content: String,
    pub memory_type: String,
    pub created_at: String,
    pub project_path: Option<String>,
}

const BASELINE_THRESHOLD: f32 = 0.59;

pub fn load_eval_set(path: &Path) -> Result<Vec<EvalQuery>> {
    let text = std::fs::read_to_string(path)?;
    let mut out = Vec::new();
    for line in text.lines() {
        let trimmed = line.trim();
        if trimmed.is_empty() {
            continue;
        }
        out.push(serde_json::from_str(trimmed)?);
    }
    Ok(out)
}

pub fn run_arm(
    store: &SqliteStore,
    embedder: &Embedder,
    queries: &[EvalQuery],
    use_prefix: bool,
    limit: usize,
) -> Result<Vec<QueryRecord>> {
    let engine = QueryEngine::new(store);
    let mut records = Vec::with_capacity(queries.len());

    for q in queries {
        let embedding = if use_prefix {
            embedder.embed_query(&q.query)?
        } else {
            embedder.embed(&q.query)?
        };

        let relevant: HashSet<Uuid> = q.relevant_memory_ids.iter().copied().collect();

        let request = QueryRequest {
            text: q.query.clone(),
            embedding: embedding.clone(),
            limit,
            filters: QueryFilters {
                entity_names: crate::link_query_entities(&q.query),
                ..QueryFilters::default()
            },
        };
        let started = std::time::Instant::now();
        let results = engine.recall(&request)?;
        let latency_ms = started.elapsed().as_millis() as u64;

        let ranked_ids: Vec<Uuid> = results.iter().map(|r| r.memory.id).collect();
        let scores: Vec<f32> = results.iter().map(|r| r.score).collect();
        let first_relevant_rank = ranked_ids
            .iter()
            .position(|id| relevant.contains(id))
            .map(|idx| idx + 1);

        let raw = store.vector_search(&embedding, limit * 3)?;
        let mut gold_raw_rank = None;
        let mut gold_raw_similarity = None;
        for (idx, (mem, sim)) in raw.iter().enumerate() {
            if relevant.contains(&mem.id) {
                gold_raw_rank = Some(idx + 1);
                gold_raw_similarity = Some(*sim);
                break;
            }
        }

        records.push(QueryRecord {
            query_id: q.query_id.clone(),
            use_prefix,
            ranked_ids,
            scores,
            first_relevant_rank,
            gold_raw_rank,
            gold_raw_similarity,
            latency_ms: Some(latency_ms),
        });
    }

    Ok(records)
}

pub struct EmbeddedQuery {
    pub query: EvalQuery,
    pub relevant: HashSet<Uuid>,
    pub bare_embedding: Vec<f32>,
    pub prefixed_embedding: Vec<f32>,
}

// The Embedder wraps the model in a Mutex, so embedding cannot run in parallel.
// We pre-compute every embedding serially here, in batches, to keep the model
// out of the parallel recall section entirely.
pub fn embed_all_queries(embedder: &Embedder, queries: &[EvalQuery]) -> Result<Vec<EmbeddedQuery>> {
    let bare_texts: Vec<&str> = queries.iter().map(|q| q.query.as_str()).collect();
    let bare = embedder.embed_batch(&bare_texts)?;

    let prefixed_owned: Vec<String> = queries
        .iter()
        .map(|q| second_brain_core::embedding::query_prompt(&q.query))
        .collect();
    let prefixed_texts: Vec<&str> = prefixed_owned.iter().map(|s| s.as_str()).collect();
    let prefixed = embedder.embed_batch(&prefixed_texts)?;

    let mut out = Vec::with_capacity(queries.len());
    for (i, q) in queries.iter().enumerate() {
        out.push(EmbeddedQuery {
            query: q.clone(),
            relevant: q.relevant_memory_ids.iter().copied().collect(),
            bare_embedding: bare[i].clone(),
            prefixed_embedding: prefixed[i].clone(),
        });
    }
    Ok(out)
}

fn record_for<S: Store + Sync>(
    embedded: &EmbeddedQuery,
    store: &S,
    use_prefix: bool,
    limit: usize,
    hybrid: Option<&HybridContext<'_>>,
) -> Result<QueryRecord> {
    let embedding = if use_prefix {
        &embedded.prefixed_embedding
    } else {
        &embedded.bare_embedding
    };

    let request = QueryRequest {
        text: embedded.query.query.clone(),
        embedding: embedding.clone(),
        limit,
        filters: QueryFilters {
            entity_names: crate::link_query_entities(&embedded.query.query),
            ..QueryFilters::default()
        },
    };
    let started = std::time::Instant::now();
    let results = match hybrid {
        None => QueryEngine::new(store).recall(&request)?,
        Some(ctx) => {
            let retriever = Retriever {
                tunables: ctx.tunables,
            };
            let pool = retriever.candidates(ctx.indexes, ctx.sqlite, &request)?;
            retriever.rank(
                store,
                &request,
                &pool,
                ctx.reranker.map(|r| r as &dyn RerankScorer),
            )?
        }
    };
    let latency_ms = started.elapsed().as_millis() as u64;

    let ranked_ids: Vec<Uuid> = results.iter().map(|r| r.memory.id).collect();
    let scores: Vec<f32> = results.iter().map(|r| r.score).collect();
    let first_relevant_rank = ranked_ids
        .iter()
        .position(|id| embedded.relevant.contains(id))
        .map(|idx| idx + 1);

    let raw = store.vector_search(embedding, limit * 3)?;
    let mut gold_raw_rank = None;
    let mut gold_raw_similarity = None;
    for (idx, (mem, sim)) in raw.iter().enumerate() {
        if embedded.relevant.contains(&mem.id) {
            gold_raw_rank = Some(idx + 1);
            gold_raw_similarity = Some(*sim);
            break;
        }
    }

    Ok(QueryRecord {
        query_id: embedded.query.query_id.clone(),
        use_prefix,
        ranked_ids,
        scores,
        first_relevant_rank,
        gold_raw_rank,
        gold_raw_similarity,
        latency_ms: Some(latency_ms),
    })
}

// SqliteStore serializes every store call behind its connection mutex, so
// sharing &store across scoped threads is safe; concurrent reads queue rather
// than race. The Embedder is absent here on purpose: embeddings are
// pre-computed in embed_all_queries.
pub fn run_arm_parallel<S: Store + Sync>(
    store: &S,
    embedded: &[EmbeddedQuery],
    use_prefix: bool,
    limit: usize,
    hybrid: Option<&HybridContext<'_>>,
) -> Result<Vec<QueryRecord>> {
    let total = embedded.len();
    if total == 0 {
        return Ok(Vec::new());
    }

    let workers = thread::available_parallelism()
        .map(|n| n.get().saturating_sub(1).max(1))
        .unwrap_or(1);
    let chunk_size = total.div_ceil(workers);

    let done = AtomicUsize::new(0);
    let collected: Mutex<Vec<(usize, QueryRecord)>> = Mutex::new(Vec::with_capacity(total));
    let error: Mutex<Option<anyhow::Error>> = Mutex::new(None);

    thread::scope(|scope| {
        for chunk in 0..workers {
            let start = chunk * chunk_size;
            if start >= total {
                break;
            }
            let end = (start + chunk_size).min(total);
            let done = &done;
            let collected = &collected;
            let error = &error;
            scope.spawn(move || {
                let mut local: Vec<(usize, QueryRecord)> = Vec::with_capacity(end - start);
                for (offset, eq) in embedded[start..end].iter().enumerate() {
                    if error.lock().unwrap().is_some() {
                        return;
                    }
                    match record_for(eq, store, use_prefix, limit, hybrid) {
                        Ok(rec) => local.push((start + offset, rec)),
                        Err(e) => {
                            *error.lock().unwrap() = Some(e);
                            return;
                        }
                    }
                    // eprintln progress because long runs were previously invisible.
                    let n = done.fetch_add(1, Ordering::Relaxed) + 1;
                    if n.is_multiple_of(25) || n == total {
                        eprintln!("  {n}/{total} queries");
                    }
                }
                collected.lock().unwrap().extend(local);
            });
        }
    });

    if let Some(e) = error.into_inner().unwrap() {
        return Err(e);
    }

    let mut indexed = collected.into_inner().unwrap();
    indexed.sort_by_key(|(i, _)| *i);
    Ok(indexed.into_iter().map(|(_, r)| r).collect())
}

pub fn aggregate(
    bare: &[QueryRecord],
    prefixed: &[QueryRecord],
    relevant_sets: &std::collections::HashMap<String, HashSet<Uuid>>,
) -> AggregateReport {
    const AGG_SEED: u64 = 0x4B1D_C0DE;
    let empty: HashSet<Uuid> = HashSet::new();

    let per_query = |rec: &QueryRecord| -> (f32, f32, f32, f32, f32) {
        let rel = relevant_sets.get(&rec.query_id).unwrap_or(&empty);
        (
            metrics::recall_at_k(&rec.ranked_ids, rel, 1),
            metrics::recall_at_k(&rec.ranked_ids, rel, 3),
            metrics::recall_at_k(&rec.ranked_ids, rel, 5),
            metrics::mrr(&rec.ranked_ids, rel),
            metrics::precision_at_k(&rec.ranked_ids, rel, 5),
        )
    };

    let arm = |records: &[QueryRecord]| -> ArmMetrics {
        if records.is_empty() {
            return ArmMetrics {
                recall_at_1: 0.0,
                recall_at_3: 0.0,
                recall_at_5: 0.0,
                mrr: 0.0,
                precision_at_5: 0.0,
            };
        }
        let n = records.len() as f32;
        let mut acc = (0.0, 0.0, 0.0, 0.0, 0.0);
        for r in records {
            let (r1, r3, r5, m, p5) = per_query(r);
            acc.0 += r1;
            acc.1 += r3;
            acc.2 += r5;
            acc.3 += m;
            acc.4 += p5;
        }
        ArmMetrics {
            recall_at_1: acc.0 / n,
            recall_at_3: acc.1 / n,
            recall_at_5: acc.2 / n,
            mrr: acc.3 / n,
            precision_at_5: acc.4 / n,
        }
    };

    let bare_idx: std::collections::HashMap<&str, &QueryRecord> =
        bare.iter().map(|r| (r.query_id.as_str(), r)).collect();

    let mut delta_r3 = Vec::new();
    let mut delta_mrr = Vec::new();
    for p_rec in prefixed {
        if let Some(b_rec) = bare_idx.get(p_rec.query_id.as_str()) {
            let (_, p_r3, _, p_mrr, _) = per_query(p_rec);
            let (_, b_r3, _, b_mrr, _) = per_query(b_rec);
            delta_r3.push(p_r3 - b_r3);
            delta_mrr.push(p_mrr - b_mrr);
        }
    }

    let gated_rate = |records: &[QueryRecord]| -> f32 {
        let flags: Vec<bool> = records
            .iter()
            .map(|r| match (r.gold_raw_rank, r.gold_raw_similarity) {
                (Some(_), Some(sim)) => sim < BASELINE_THRESHOLD,
                _ => false,
            })
            .collect();
        metrics::gated_out_rate(&flags)
    };

    AggregateReport {
        bare: arm(bare),
        prefixed: arm(prefixed),
        delta_recall_at_3_ci: bootstrap::paired_bootstrap_ci(&delta_r3, 10000, 0.95, AGG_SEED),
        delta_mrr_ci: bootstrap::paired_bootstrap_ci(&delta_mrr, 10000, 0.95, AGG_SEED),
        gated_out_rate_bare: gated_rate(bare),
        gated_out_rate_prefixed: gated_rate(prefixed),
    }
}

pub fn gate_sweep(prefixed: &[QueryRecord]) -> GateSweepReport {
    const GRID_SEED: u64 = 0x5EED_6A7E;

    let recalled_at = |rec: &QueryRecord, k: usize, t: f32| -> f32 {
        match (rec.gold_raw_rank, rec.gold_raw_similarity) {
            (Some(rank), Some(sim)) if rank <= k && sim >= t => 1.0,
            _ => 0.0,
        }
    };

    let mean = |vals: &[f32]| -> f32 {
        if vals.is_empty() {
            0.0
        } else {
            vals.iter().sum::<f32>() / vals.len() as f32
        }
    };

    let baseline_r3: Vec<f32> = prefixed
        .iter()
        .map(|r| recalled_at(r, 3, BASELINE_THRESHOLD))
        .collect();
    let baseline_r3_mean = mean(&baseline_r3);

    let mut frontier = Vec::with_capacity(41);
    let mut chosen_threshold = BASELINE_THRESHOLD;
    let mut chosen_beats_baseline = false;
    let mut best_recall_at_3 = baseline_r3_mean;

    for step in 0..=40u32 {
        let t = 0.40 + step as f32 * 0.01;

        let r1: Vec<f32> = prefixed.iter().map(|r| recalled_at(r, 1, t)).collect();
        let r3: Vec<f32> = prefixed.iter().map(|r| recalled_at(r, 3, t)).collect();
        let r5: Vec<f32> = prefixed.iter().map(|r| recalled_at(r, 5, t)).collect();

        let recall_at_3 = mean(&r3);
        let recall_at_5 = mean(&r5);

        frontier.push(GatePoint {
            threshold: t,
            recall_at_1: mean(&r1),
            recall_at_3,
            recall_at_5,
            precision_proxy: recall_at_5 / 5.0,
        });

        let deltas: Vec<f32> = r3
            .iter()
            .zip(baseline_r3.iter())
            .map(|(t_val, b_val)| t_val - b_val)
            .collect();
        let (lo, _hi) = bootstrap::paired_bootstrap_ci(&deltas, 2000, 0.95, GRID_SEED);

        if lo > 0.0 && recall_at_3 > best_recall_at_3 + 1e-6 {
            best_recall_at_3 = recall_at_3;
            chosen_threshold = t;
            chosen_beats_baseline = true;
        }
    }

    GateSweepReport {
        frontier,
        baseline_threshold: BASELINE_THRESHOLD,
        chosen_threshold,
        chosen_beats_baseline,
    }
}

/// One gold query scored for the exhausted-scope floor: the cosine of its
/// gold memory, which a recall scoped to the gold's own project must admit,
/// and the best cosine in an unrelated project, which a recall scoped there
/// should answer with nothing.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct ScopedPair {
    pub query_id: String,
    pub gold: f32,
    pub foreign_project: String,
    pub foreign_best: f32,
}

#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct FloorPoint {
    pub threshold: f32,
    /// Share of gold memories at or above the floor.
    pub gold_admitted: f32,
    /// Share of foreign-scope answers left empty by the floor.
    pub foreign_empty: f32,
}

#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct FloorSweepReport {
    pub pairs: usize,
    pub frontier: Vec<FloorPoint>,
    /// The highest floor that still admits `FLOOR_GOLD_TARGET` of the gold
    /// memories. Balancing the two rates instead (Youden's J) picked 0.71 on
    /// the 2026-09 corpus and dropped a third of gold answers from their own
    /// project, which is the one failure a caller who named the scope sees.
    pub chosen_threshold: f32,
}

/// A named scope is the caller asserting where the answer is, so the floor
/// may cost at most one gold answer in a hundred.
pub const FLOOR_GOLD_TARGET: f32 = 0.99;

pub fn exhausted_floor_sweep(pairs: &[ScopedPair]) -> FloorSweepReport {
    let n = pairs.len().max(1) as f32;
    let mut frontier = Vec::with_capacity(41);
    let mut chosen = 0.40_f32;
    for step in 0..=40u32 {
        let t = 0.40 + step as f32 * 0.01;
        let gold_admitted = pairs.iter().filter(|p| p.gold >= t).count() as f32 / n;
        let foreign_empty = pairs.iter().filter(|p| p.foreign_best < t).count() as f32 / n;
        if gold_admitted >= FLOOR_GOLD_TARGET {
            chosen = t;
        }
        frontier.push(FloorPoint {
            threshold: t,
            gold_admitted,
            foreign_empty,
        });
    }
    FloorSweepReport {
        pairs: pairs.len(),
        frontier,
        chosen_threshold: chosen,
    }
}

/// Scores every gold query whose gold memory is live in `store` against its
/// own vector and against the best match in a foreign project: one of
/// `projects` other than the gold's, picked by the query's position so the
/// pairing is deterministic and spread over every project.
pub fn scoped_pairs(
    store: &SqliteStore,
    queries: &[EmbeddedQuery],
    projects: &[String],
) -> Result<Vec<ScopedPair>> {
    let mut pairs = Vec::new();
    for (i, q) in queries.iter().enumerate() {
        let Some(own) = q.query.source_project.as_deref() else {
            continue;
        };
        let Some(gold) = store.get_memory_with_embedding(q.query.seed_memory_id)? else {
            continue;
        };
        if gold.embedding.is_empty() || gold.tier != second_brain_core::schema::Tier::Hot {
            continue;
        }
        let foreign: Vec<&String> = projects.iter().filter(|p| p.as_str() != own).collect();
        if foreign.is_empty() {
            continue;
        }
        let foreign_project = foreign[i % foreign.len()].clone();
        let emb = &q.prefixed_embedding;
        let foreign_best = store
            .filtered_vector_search(emb, None, None, Some(&foreign_project), 1)?
            .first()
            .map(|(_, score)| *score)
            .unwrap_or(0.0);
        pairs.push(ScopedPair {
            query_id: q.query.query_id.clone(),
            gold: second_brain_core::retrieval::cosine(emb, &gold.embedding),
            foreign_project,
            foreign_best,
        });
    }
    Ok(pairs)
}

pub fn extract_corpus(store: &SqliteStore, out: &Path) -> Result<usize> {
    let memories = store.all_memories_with_embeddings()?;
    let mut file = std::fs::File::create(out)?;
    let mut count = 0;
    // Only rows recall can return are eligible as gold: an archived or
    // superseded seed would score as a miss no matter how the ranker did.
    for m in memories.iter().filter(|m| {
        m.tier == second_brain_core::schema::Tier::Hot
            && m.superseded_by.is_none()
            && !m.embedding.is_empty()
    }) {
        let entry = CorpusEntry {
            id: m.id,
            content: m.content.clone(),
            memory_type: format!("{:?}", m.memory_type).to_lowercase(),
            created_at: m.created_at.to_rfc3339(),
            project_path: m.project_path.clone(),
        };
        writeln!(file, "{}", serde_json::to_string(&entry)?)?;
        count += 1;
    }
    Ok(count)
}

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

    fn pair(gold: f32, foreign_best: f32) -> ScopedPair {
        ScopedPair {
            query_id: "q".to_string(),
            gold,
            foreign_project: "p".to_string(),
            foreign_best,
        }
    }

    #[test]
    fn the_floor_sweep_picks_the_threshold_separating_gold_from_foreign() {
        let pairs = [pair(0.70, 0.50), pair(0.65, 0.52), pair(0.60, 0.48)];
        let report = exhausted_floor_sweep(&pairs);
        assert_eq!(report.pairs, 3);
        let chosen = report.chosen_threshold;
        assert!(
            (chosen - 0.60).abs() < 1e-4,
            "the highest floor that still admits every gold memory: {chosen}"
        );
        let at = |t: f32| {
            report
                .frontier
                .iter()
                .find(|p| (p.threshold - t).abs() < 1e-4)
                .unwrap()
                .clone()
        };
        assert!((at(0.66).gold_admitted - 1.0 / 3.0).abs() < 1e-4);
        assert!((at(0.50).foreign_empty - 1.0 / 3.0).abs() < 1e-4);
    }

    fn record(
        id_rank: Option<usize>,
        raw_rank: Option<usize>,
        raw_sim: Option<f32>,
    ) -> QueryRecord {
        QueryRecord {
            query_id: "q".to_string(),
            use_prefix: true,
            ranked_ids: Vec::new(),
            scores: Vec::new(),
            first_relevant_rank: id_rank,
            gold_raw_rank: raw_rank,
            gold_raw_similarity: raw_sim,
            latency_ms: None,
        }
    }

    #[test]
    fn load_eval_set_parses_one_object_per_line() {
        let dir = std::env::temp_dir();
        let path = dir.join(format!("eval_set_{}.jsonl", Uuid::new_v4()));
        let id_a = Uuid::new_v4();
        let id_b = Uuid::new_v4();
        let line1 = format!(
            r#"{{"query_id":"q1","query":"kuzu choice","query_variant":"literal","seed_memory_id":"{id_a}","memory_type":"decision","relevant_memory_ids":["{id_a}"]}}"#
        );
        let line2 = format!(
            r#"{{"query_id":"q2","query":"sync design","query_variant":"paraphrase","seed_memory_id":"{id_b}","memory_type":"architecture","relevant_memory_ids":["{id_b}","{id_a}"],"tags":["sync"]}}"#
        );
        let mut f = std::fs::File::create(&path).unwrap();
        writeln!(f, "{line1}").unwrap();
        writeln!(f, "{line2}").unwrap();
        drop(f);

        let queries = load_eval_set(&path).unwrap();
        std::fs::remove_file(&path).ok();

        assert_eq!(queries.len(), 2);
        assert_eq!(queries[0].query_id, "q1");
        assert_eq!(queries[0].seed_memory_id, id_a);
        assert_eq!(queries[1].relevant_memory_ids.len(), 2);
        assert_eq!(queries[1].tags, vec!["sync".to_string()]);
    }

    #[test]
    fn load_eval_set_tolerates_blank_lines() {
        let dir = std::env::temp_dir();
        let path = dir.join(format!("eval_blank_{}.jsonl", Uuid::new_v4()));
        let id = Uuid::new_v4();
        let line = format!(
            r#"{{"query_id":"q1","query":"x","query_variant":"v","seed_memory_id":"{id}","memory_type":"semantic","relevant_memory_ids":["{id}"]}}"#
        );
        std::fs::write(&path, format!("\n{line}\n\n")).unwrap();

        let queries = load_eval_set(&path).unwrap();
        std::fs::remove_file(&path).ok();

        assert_eq!(queries.len(), 1);
    }

    #[test]
    fn gate_sweep_emits_full_grid_and_monotone_recall() {
        let records = vec![
            record(Some(1), Some(1), Some(0.85)),
            record(Some(2), Some(2), Some(0.62)),
            record(Some(4), Some(4), Some(0.55)),
            record(None, None, None),
        ];

        let report = gate_sweep(&records);

        // grid 0.40..=0.80 step 0.01 inclusive is 41 points.
        assert_eq!(report.frontier.len(), 41);
        assert!((report.frontier.first().unwrap().threshold - 0.40).abs() < 1e-4);
        assert!((report.frontier.last().unwrap().threshold - 0.80).abs() < 1e-4);

        for w in report.frontier.windows(2) {
            assert!(
                w[0].recall_at_3 >= w[1].recall_at_3 - 1e-6,
                "recall must not increase as the gate tightens"
            );
        }
    }

    #[test]
    fn gate_sweep_recall_reflects_raw_rank_and_similarity() {
        let records = vec![
            record(Some(1), Some(1), Some(0.85)),
            record(Some(2), Some(2), Some(0.62)),
            record(Some(4), Some(4), Some(0.55)),
            record(None, None, None),
        ];

        let report = gate_sweep(&records);

        let at = |t: f32| {
            report
                .frontier
                .iter()
                .find(|p| (p.threshold - t).abs() < 1e-4)
                .unwrap()
        };

        // T=0.50: golds at sim 0.85, 0.62, 0.55 survive; ranks 1,2,4. recall@3 covers
        // the first two (rank<=3 and sim>=0.50) so 2/4 = 0.5; recall@5 covers all three = 0.75.
        let p050 = at(0.50);
        assert!(
            (p050.recall_at_1 - 0.25).abs() < 1e-6,
            "recall@1 was {}",
            p050.recall_at_1
        );
        assert!(
            (p050.recall_at_3 - 0.5).abs() < 1e-6,
            "recall@3 was {}",
            p050.recall_at_3
        );
        assert!(
            (p050.recall_at_5 - 0.75).abs() < 1e-6,
            "recall@5 was {}",
            p050.recall_at_5
        );

        // T=0.70: only the 0.85 gold survives, at rank 1.
        let p070 = at(0.70);
        assert!((p070.recall_at_1 - 0.25).abs() < 1e-6);
        assert!((p070.recall_at_3 - 0.25).abs() < 1e-6);
        assert!((p070.recall_at_5 - 0.25).abs() < 1e-6);
    }

    #[test]
    fn gate_sweep_keeps_baseline_when_nothing_beats_it() {
        let records = vec![record(Some(1), Some(1), Some(0.85))];
        let report = gate_sweep(&records);
        assert!((report.baseline_threshold - 0.59).abs() < 1e-6);
        assert!(!report.chosen_beats_baseline);
        assert!((report.chosen_threshold - 0.59).abs() < 1e-6);
    }
}