fetchium-core 1.0.0

Fetchium core library — search, extract, rank, validate, cache
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
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
//! Research pipeline orchestrator — 9-step PRD SS10 Mode B (PRD §10).

use crate::citation::evidence_graph::EvidenceGraphBuilder;
use crate::citation::formatter::CitationFormatter;
use crate::citation::types::SourceMeta;
use crate::config::FetchiumConfig;
use crate::error::FetchiumError;
use crate::extract::pipeline::extract;
use crate::http::client::HttpClient;
use crate::research::decompose::decompose_query;
use crate::research::{ResearchConfig, ResearchMeta, ResearchReport};
use crate::search::orchestrator::{OrchestratorConfig, SearchOrchestrator};
use crate::validate::authority::AuthorityScorer;
use crate::validate::calibration::ConfidenceCalibrator;
use crate::validate::content::{ContentInput, ContentValidator};
use crate::validate::cross_source::{CrossSourceVerifier, SourceContent as V4SourceContent};
use crate::validate::extraction::{ExtractionInput, ExtractionValidator};
use crate::validate::rar::{RarEngine, RarState};
use crate::validate::temporal::{SourceFreshness, TemporalValidator};
use std::time::Instant;

/// Extract the top `n` most query-relevant sentences from a block of text.
///
/// Used as a fallback synthesizer when AI is unavailable. Scores each sentence by
/// how many query keywords it contains, returns highest-scoring sentences joined
/// into a readable paragraph (≤ 200 chars each, ≥ 30 chars to exclude noise).
fn extract_key_sentences(text: &str, query_keywords: &[String], n: usize) -> String {
    if text.is_empty() {
        return String::new();
    }
    // Sentence split: ". " or "\n" boundaries, trim whitespace.
    let sentences: Vec<&str> = text
        .split('\n')
        .flat_map(|line| line.split(". "))
        .map(|s| s.trim())
        .filter(|s| s.len() >= 30 && s.len() <= 300)
        .collect();

    if sentences.is_empty() {
        // Nothing usable — return first 150 chars as a snippet.
        return text.chars().take(150).collect();
    }

    // Score by keyword overlap.
    let mut scored: Vec<(usize, &str)> = sentences
        .iter()
        .map(|s| {
            let sl = s.to_lowercase();
            let score = query_keywords
                .iter()
                .filter(|kw| sl.contains(kw.as_str()))
                .count();
            (score, *s)
        })
        .filter(|(score, _)| *score > 0)
        .collect();

    if scored.is_empty() {
        // No keyword hits at all — just return first sentence.
        return sentences.first().copied().unwrap_or("").to_string();
    }

    scored.sort_by(|a, b| b.0.cmp(&a.0));
    scored.truncate(n);
    scored
        .iter()
        .map(|(_, s)| *s)
        .collect::<Vec<_>>()
        .join(". ")
}

/// Research pipeline orchestrator.
///
/// Steps (PRD SS10 Mode B):
/// 1. Query decomposition
/// 2. Parallel multi-backend search (via SearchOrchestrator)
/// 3. Top sources fetched via CEP
/// 4. Content extracted via QATBE
/// 5. RAR reflection loop validates retrieval quality
/// 6. HyperFusion ranking
/// 7. Evidence mapping via EGP
/// 8. Synthesis with strict citation
/// 9. Validation layer (6-layer)
pub struct ResearchPipeline;

impl ResearchPipeline {
    /// Ensure source diversity by interleaving results from different backends.
    ///
    /// Without this, HyperFusion's BM25 signal causes long Wikipedia articles
    /// to dominate the top N (they contain more keyword matches). Premium backends
    /// (Tavily, Exa, Firecrawl, Serper) return targeted results that are more
    /// relevant but get ranked lower.
    ///
    /// Strategy: Round-robin from each backend, prioritizing premium backends,
    /// until we have enough results.
    fn diversify_sources(
        results: Vec<crate::types::ResultItem>,
        target: usize,
    ) -> Vec<crate::types::ResultItem> {
        use std::collections::HashMap;

        if results.len() <= target {
            return results;
        }

        // Group results by backend, preserving rank order within each group
        let mut by_backend: HashMap<String, Vec<crate::types::ResultItem>> = HashMap::new();
        for r in results {
            let key = format!("{}", r.backend);
            by_backend.entry(key).or_default().push(r);
        }

        // Priority order: premium first, then free backends
        let priority_order = [
            "tavily",
            "exa",
            "firecrawl",
            "serper", // premium: targeted results
            "searxng",
            "arxiv", // free: good quality
            "hackernews",
            "reddit",
            "stackoverflow", // free: community
            "wikipedia",
            "github", // free: generic
            "duckduckgo",
            "google",
            "bing",
            "brave", // free: web search
        ];

        let mut selected = Vec::with_capacity(target);
        let mut seen_urls: std::collections::HashSet<String> = std::collections::HashSet::new();

        // Phase 1: Take top result from each premium backend (if available)
        for backend_name in &priority_order[..4] {
            if selected.len() >= target {
                break;
            }
            if let Some(items) = by_backend.get_mut(*backend_name) {
                while !items.is_empty() && selected.len() < target {
                    let item = items.remove(0);
                    if seen_urls.insert(item.url.clone()) {
                        selected.push(item);
                        break; // One per backend in phase 1
                    }
                }
            }
        }

        // Phase 2: Round-robin from all backends until we reach target
        let mut round = 0;
        while selected.len() < target {
            let mut added_any = false;
            for backend_name in &priority_order {
                if selected.len() >= target {
                    break;
                }
                if let Some(items) = by_backend.get_mut(*backend_name) {
                    while !items.is_empty() {
                        let item = items.remove(0);
                        if seen_urls.insert(item.url.clone()) {
                            selected.push(item);
                            added_any = true;
                            break;
                        }
                    }
                }
            }
            round += 1;
            if !added_any || round > 20 {
                break;
            }
        }

        selected
    }

    /// Synthesize a research report using AI with numbered source context.
    ///
    /// Uses a oneshot channel + tokio::task::spawn_local pattern to keep
    /// the parent future Send-safe (required by axum handlers).
    /// Falls back to empty string on failure.
    async fn synthesize_with_ai(
        query: &str,
        sources: &[SourceMeta],
        citations: &[crate::citation::types::FormattedCitation],
        extracted_texts: &[String],
        fetchium_config: &FetchiumConfig,
        thinking: bool,
    ) -> String {
        use crate::ai::types::{AiConfig, ChatMessage};

        let mut ai_config = AiConfig::from_fetchium_config(fetchium_config);
        ai_config.thinking = thinking;
        if ai_config.providers.fallback_chain.is_empty() && ai_config.default_model.is_none() {
            return String::new();
        }

        // Build numbered source context with actual content excerpts (max 600 chars each).
        // This gives the AI real evidence to synthesize from, not just titles.
        let mut source_context = String::new();
        for (i, (meta, _citation)) in sources.iter().zip(citations.iter()).enumerate() {
            let excerpt = extracted_texts
                .get(i)
                .map(|t| {
                    let trimmed = t.trim();
                    let chars: String = trimmed.chars().take(500).collect();
                    if trimmed.chars().count() > 500 {
                        format!("{}...", chars)
                    } else {
                        chars
                    }
                })
                .unwrap_or_default();
            source_context.push_str(&format!(
                "[{}] {}\n{}\n{}\n\n",
                i + 1,
                meta.title,
                meta.url,
                excerpt,
            ));
        }

        let system_prompt = crate::ai::prompt::research_synthesis_prompt(query, sources.len());

        let messages = vec![
            ChatMessage {
                role: "system".into(),
                content: system_prompt,
            },
            ChatMessage {
                role: "user".into(),
                content: format!("SOURCES:\n{}\n\nSynthesize a report now.", source_context),
            },
        ];

        let providers = ai_config.providers.clone();

        // Use a oneshot channel to bridge across the Send boundary.
        // The non-Send part (dyn FnMut) is confined to a spawn_blocking thread.
        let (tx, rx) = tokio::sync::oneshot::channel::<String>();

        let ai_config_owned = ai_config;
        std::thread::spawn(move || {
            let rt = tokio::runtime::Builder::new_current_thread()
                .enable_all()
                .build()
                .unwrap();
            let result = rt.block_on(async move {
                let mut noop = |_: &str| {};
                match tokio::time::timeout(
                    std::time::Duration::from_secs(30),
                    crate::ai::provider_client::chat_with_fallback(
                        &messages,
                        None,
                        &ai_config_owned,
                        &providers,
                        false,
                        &mut noop,
                    ),
                )
                .await
                {
                    Ok(Ok(result)) => {
                        tracing::info!(
                            "AI synthesis completed: provider={}, model={}",
                            result.provider.slug(),
                            result.model_used
                        );
                        result.content
                    }
                    Ok(Err(e)) => {
                        tracing::warn!("AI synthesis failed: {e}");
                        String::new()
                    }
                    Err(_) => {
                        tracing::warn!("AI synthesis timed out (30s)");
                        String::new()
                    }
                }
            });
            let _ = tx.send(result);
        });

        rx.await.unwrap_or_default()
    }

    /// Execute the full research pipeline from a configuration.
    ///
    /// The orchestrator, extractor, and ranker are injected via the config and client.
    /// This implementation wires the complete pipeline end-to-end.
    pub async fn execute(
        config: &ResearchConfig,
        fetchium_config: &FetchiumConfig,
        http_client: &HttpClient,
    ) -> Result<ResearchReport, FetchiumError> {
        let start = Instant::now();

        // ── FAST PATH ──────────────────────────────────────────────────────
        // When AI synthesis is disabled, use a lightweight pipeline that skips
        // URL fetching entirely. Produces results from search snippets only,
        // achieving Exa-like speed (~2-3s) with no third-party dependencies.
        if !config.ai_synthesis {
            return Self::execute_fast(config, fetchium_config, http_client, start).await;
        }

        // Step 1: Decompose query into perspective-aware sub-queries
        let sub_queries = decompose_query(&config.query);

        // Step 2: Parallel Dispatch — search ALL sub-queries concurrently
        let search_budget = (config.max_sources as u32) * 3;
        let mut orch_config =
            OrchestratorConfig::from_fetchium_config(fetchium_config, search_budget);
        orch_config.max_total_results = search_budget;
        let orchestrator =
            std::sync::Arc::new(SearchOrchestrator::new(http_client.clone(), orch_config));

        // Fire all sub-queries in parallel for broader coverage
        let mut search_handles = Vec::new();
        for sq in &sub_queries {
            let orch = orchestrator.clone();
            let sq = sq.clone();
            let budget = search_budget;
            search_handles.push(tokio::spawn(async move {
                orch.search(&sq, Some(budget)).await.unwrap_or_default()
            }));
        }

        // Collect and merge results from all sub-queries
        let mut search_results = Vec::new();
        let mut seen_urls = std::collections::HashSet::new();
        for handle in search_handles {
            if let Ok(results) = handle.await {
                for r in results {
                    if seen_urls.insert(r.url.clone()) {
                        search_results.push(r);
                    }
                }
            }
        }

        let search_elapsed = start.elapsed();
        tracing::info!(
            "Search completed in {:.1}s ({} results from {} sub-queries)",
            search_elapsed.as_secs_f32(),
            search_results.len(),
            sub_queries.len()
        );

        // Source diversity selection
        search_results = Self::diversify_sources(search_results, config.max_sources);

        // Pre-filter: check title+snippet relevance BEFORE any fetching or synthesis
        let query_keywords_pre: Vec<String> = config
            .query
            .split_whitespace()
            .map(|w| {
                w.to_lowercase()
                    .trim_matches(|c: char| !c.is_alphanumeric())
                    .to_string()
            })
            .filter(|w| w.len() > 3)
            .collect();

        const PRE_GENERIC_WORDS: &[&str] = &[
            "software",
            "engineering",
            "development",
            "system",
            "technology",
            "computer",
            "application",
            "research",
            "study",
            "analysis",
            "paper",
            "using",
            "based",
            "approach",
            "model",
            "large",
            "deep",
            "data",
            "work",
            "show",
            "2024",
            "2025",
            "2026",
        ];

        let distinctive_pre: Vec<&String> = query_keywords_pre
            .iter()
            .filter(|w| !PRE_GENERIC_WORDS.contains(&w.as_str()))
            .collect();

        if !distinctive_pre.is_empty() {
            let pre_count = search_results.len();
            search_results.retain(|item| {
                let haystack = format!(
                    "{} {}",
                    item.title.to_lowercase(),
                    item.snippet
                        .chars()
                        .take(500)
                        .collect::<String>()
                        .to_lowercase()
                );
                distinctive_pre
                    .iter()
                    .any(|kw| haystack.contains(kw.as_str()))
            });
            tracing::debug!(
                "Pre-filter: {} → {} results",
                pre_count,
                search_results.len()
            );
        }

        // ── SPEED OPTIMIZATION: Start AI synthesis IMMEDIATELY using search snippets ──
        // Don't wait for URL fetching — search snippets (200-1500 chars) contain
        // enough content for a quality synthesis. This saves 5-8s of URL fetch time.
        // URL fetching still runs for validation but doesn't block the critical path.

        use crate::extract::ExtractedContent;
        use crate::types::BackendId;

        // Build snippet-based source data for immediate AI synthesis
        let formatter = CitationFormatter::new(config.citation_style);
        let mut snippet_metas = Vec::new();
        let mut snippet_citations = Vec::new();
        let mut snippet_texts = Vec::new();

        for (idx, item) in search_results.iter().enumerate() {
            let author = url::Url::parse(&item.url).ok().and_then(|u| {
                u.host_str()
                    .map(|h| h.trim_start_matches("www.").to_string())
            });
            let meta = SourceMeta {
                url: item.url.clone(),
                title: item.title.clone(),
                author,
                publisher: None,
                published_date: item.published_date.clone().or_else(|| Some("2026".into())),
                accessed_date: chrono::Utc::now().to_rfc3339(),
            };
            snippet_citations.push(formatter.format(&meta, idx + 1));
            snippet_metas.push(meta);
            snippet_texts.push(item.snippet.clone());
        }

        // Fire AI synthesis immediately (uses search snippets, no URL fetch needed)
        let ai_handle = if config.ai_synthesis && !snippet_metas.is_empty() {
            let query_owned = config.query.clone();
            let metas_owned = snippet_metas.clone();
            let citations_owned = snippet_citations.clone();
            let texts_owned = snippet_texts.clone();
            let fetchium_owned = fetchium_config.clone();
            let thinking = config.thinking;
            Some(tokio::spawn(async move {
                Self::synthesize_with_ai(
                    &query_owned,
                    &metas_owned,
                    &citations_owned,
                    &texts_owned,
                    &fetchium_owned,
                    thinking,
                )
                .await
            }))
        } else {
            None
        };

        // ── CONCURRENT: URL Fetch + Extract (for validation, runs while AI synthesizes) ──
        let mut extracted_sources = Vec::new();
        let mut fetch_tasks = Vec::new();

        for mut item in search_results {
            if item.url.contains("arxiv.org/abs/") {
                item.url = item.url.replace("arxiv.org/abs/", "arxiv.org/html/");
            }

            /// Rich snippet threshold — skip fetch for premium backends with enough content.
            const RICH_SNIPPET_THRESHOLD: usize = 300;
            let is_rich = matches!(
                item.backend,
                BackendId::Tavily | BackendId::Exa | BackendId::Firecrawl
            ) && item.snippet.len() >= RICH_SNIPPET_THRESHOLD;

            if is_rich {
                let title = item.title.clone();
                let text = item.snippet.clone();
                let extracted = ExtractedContent {
                    title,
                    text,
                    tokens: 0,
                    metadata: crate::extract::ContentMetadata {
                        author: None,
                        published_date: item.published_date.clone(),
                        language: Some("en".into()),
                        ..Default::default()
                    },
                    layer_used: crate::types::CepLayer::Layer1,
                };
                extracted_sources.push((item, extracted));
            } else {
                let client = http_client.clone();
                fetch_tasks.push(tokio::spawn(async move {
                    match tokio::time::timeout(
                        std::time::Duration::from_secs(6),
                        client.fetch_text(&item.url),
                    )
                    .await
                    {
                        Ok(Ok(html)) => (item, html),
                        _ => (item, String::new()),
                    }
                }));
            }
        }

        // Fetch remaining URLs (runs concurrently with AI synthesis)
        for task in fetch_tasks {
            if let Ok((item, html)) = task.await {
                if !html.is_empty() {
                    let extracted = extract(&html, &item.url);
                    extracted_sources.push((item, extracted));
                }
            }
        }

        // ── Build validation inputs from fetched content ──────────────────────
        let mut v1_inputs = Vec::new();
        let mut v2_inputs = Vec::new();
        let mut v3_inputs = Vec::new();
        let mut v4_inputs = Vec::new();
        let mut v5_inputs = Vec::new();

        // Intelligence layer: PIE observation (lightweight, non-blocking)
        let pie = crate::intelligence::pie::PersistentIntelligenceEngine::new().ok();
        if let Some(pie_engine) = &pie {
            let topic = crate::intelligence::pie::extract_topic(&config.query);
            let domains: Vec<String> = extracted_sources
                .iter()
                .filter_map(|(item, _)| {
                    url::Url::parse(&item.url)
                        .ok()
                        .and_then(|u| u.host_str().map(|h| h.to_string()))
                })
                .collect();
            let domain_strs: Vec<&str> = domains.iter().map(|s| s.as_str()).collect();
            let _ = pie_engine.observe_search(&config.query, &domain_strs, &topic);
        }

        // SGT: capture source data before consuming extracted_sources
        let sgt_hops: Vec<(String, String, String)> = if config.trace_sources {
            extracted_sources
                .iter()
                .map(|(item, ext)| {
                    let claim = ext
                        .text
                        .split(". ")
                        .next()
                        .unwrap_or(&ext.title)
                        .to_string();
                    (item.url.clone(), ext.title.clone(), claim)
                })
                .collect()
        } else {
            vec![]
        };

        // Intelligence layer: ACS validation
        let mut acs_flags = Vec::new();
        if config.trust_verify {
            let acs = crate::intelligence::acs::AdversarialContentShield::new();
            for (_item, ext) in &extracted_sources {
                let domain = url::Url::parse(&_item.url)
                    .map(|u| u.host_str().unwrap_or("unknown").to_string())
                    .unwrap_or_else(|_| "unknown".to_string());
                acs_flags.push(acs.analyze(&ext.text, &domain));
            }
        }

        // EGP data (before consuming extracted_sources)
        let egp_data: Vec<(String, String, String, f64)> = if config.evidence_graph {
            extracted_sources
                .iter()
                .map(|(item, ext)| {
                    (
                        item.url.clone(),
                        item.title.clone(),
                        ext.text.clone(),
                        item.score.unwrap_or(0.5),
                    )
                })
                .collect()
        } else {
            vec![]
        };

        let total_extracted = extracted_sources.len();

        // Use snippet-based metadata for the report (AI already has this data)
        let source_metas = snippet_metas;
        let formatted_citations = snippet_citations;
        let mut extracted_texts: Vec<String> = Vec::new();

        for (idx, (item, ext)) in extracted_sources.into_iter().enumerate() {
            extracted_texts.push(ext.text.clone());

            let has_ssl = item.url.starts_with("https");
            v1_inputs.push((item.url.clone(), has_ssl, 0));
            v2_inputs.push(ContentInput {
                url: item.url.clone(),
                text: ext.text.clone(),
            });
            v3_inputs.push(SourceFreshness {
                url: item.url.clone(),
                published_date: None,
                last_modified: None,
            });
            let claims: Vec<String> = ext
                .text
                .split(". ")
                .take(5)
                .map(|s| s.to_string())
                .collect();
            v4_inputs.push(V4SourceContent {
                url: item.url.clone(),
                index: idx,
                title: ext.title.clone(),
                claims,
                full_text: ext.text.clone(),
                confidence: item.score.unwrap_or(0.5),
            });
            v5_inputs.push(ExtractionInput {
                url: item.url.clone(),
                truncated: false,
                segment_count: 5,
                encoding_errors: 0,
            });
        }

        let sources_fetched = total_extracted;

        // Run Validation Layers (fast, local computation — <100ms total)
        let authority = AuthorityScorer::default();
        let l1 = authority.validate_sources(&v1_inputs);

        let content_val = ContentValidator::default();
        let l2 = content_val.validate(&v2_inputs, &config.query);

        let temporal = TemporalValidator::default();
        let l3 = temporal.validate(&v3_inputs, &config.query);

        let cross = CrossSourceVerifier::new();
        let l4 = cross.verify_with_query(&v4_inputs, &config.query);

        let l5 = ExtractionValidator::validate(&v5_inputs);

        // RAR reflection loop
        let rar_engine = RarEngine::default();
        let state = RarState {
            query: config.query.clone(),
            total_results: total_extracted,
            relevant_count: total_extracted,
            sufficiency_score: l2.score,
            support_ratio: l4.score,
            consistency_score: l4.score,
            unsupported_claims: vec![],
            contradictions: vec![],
            candidate_urls: vec![],
            low_relevance_terms: vec![],
        };
        let rar_iterations = vec![rar_engine.evaluate(&state, 0)];

        // Evidence mapping via EGP
        let egp = if config.evidence_graph {
            let mut builder = EvidenceGraphBuilder::new(&config.query);
            for (url, title, text, score) in &egp_data {
                builder.add_source(url, title, text, *score);
            }
            Some(builder.build())
        } else {
            None
        };

        // ── Await AI synthesis result ──────────────────────────────────────
        let mut synthesis = if let Some(handle) = ai_handle {
            handle.await.unwrap_or_default()
        } else {
            String::new()
        };

        // Fallback: extractive synthesis when AI is unavailable.
        if synthesis.is_empty() {
            if source_metas.is_empty() {
                synthesis = format!("No sources found for: **{}**\n\nTry broadening your query or checking your SearXNG connection.", config.query);
            } else {
                let mut fb = String::new();
                fb.push_str("> *Note: AI synthesis unavailable — showing extractive key-sentence summary.*\n\n");
                fb.push_str(&format!("## What Sources Say About: {}\n\n", config.query));

                let all_keywords: Vec<String> = config
                    .query
                    .split_whitespace()
                    .map(|w| {
                        w.to_lowercase()
                            .trim_matches(|c: char| !c.is_alphanumeric())
                            .to_string()
                    })
                    .filter(|w| w.len() > 3)
                    .collect();

                for (idx, meta) in source_metas.iter().enumerate() {
                    let text = extracted_texts.get(idx).map(|s| s.as_str()).unwrap_or("");
                    let key_sentences = extract_key_sentences(text, &all_keywords, 3);
                    let marker = &formatted_citations[idx].inline_marker;
                    fb.push_str(&format!("### [{}] {}\n", idx + 1, meta.title));
                    fb.push_str(&format!("*{}*\n\n", meta.url));
                    if key_sentences.is_empty() {
                        let preview: String = text.chars().take(120).collect();
                        if !preview.is_empty() {
                            fb.push_str(&format!("{} {}\n\n", preview, marker));
                        }
                    } else {
                        fb.push_str(&format!("{} {}\n\n", key_sentences, marker));
                    }
                }
                synthesis = fb;
            }
        }

        let reference_section = formatter.format_references(&source_metas);

        // SGT: append source genealogy report when trace_sources is requested
        if config.trace_sources && !sgt_hops.is_empty() {
            let chain = crate::intelligence::sgt::build_chain(sgt_hops);
            synthesis.push_str("\n\n");
            synthesis.push_str(&chain.to_markdown());
        }

        let calibrator = ConfidenceCalibrator::default();
        let validation_res = calibrator.build_result(
            config.validation_mode,
            vec![l1, l2, l3, l4, l5],
            vec![],
            vec![],
        );

        let mut overall_confidence = validation_res.confidence;

        // Intelligence layer: Confidence Calibration (CCE)
        if config.trust_verify {
            let db_path = crate::intelligence::intelligence_data_dir().join("calibration.db");
            if let Ok(cce) = crate::intelligence::cce::ConfidenceCalibrationEngine::new(&db_path) {
                let topic = crate::intelligence::pie::extract_topic(&config.query);
                if let Ok(calibrated) = cce.calibrate(&topic, overall_confidence) {
                    overall_confidence = calibrated.calibrated;
                }
            }
        }

        let pass_rate = if validation_res.passed { 1.0 } else { 0.5 };
        let duration_ms = start.elapsed().as_millis() as u64;

        Ok(ResearchReport {
            query: config.query.clone(),
            sub_queries,
            synthesis,
            sources: source_metas,
            citations: formatted_citations,
            reference_section,
            validation: validation_res,
            evidence_graph: egp,
            rar_iterations,
            meta: ResearchMeta {
                duration_ms,
                sources_fetched,
                sources_validated: sources_fetched,
                validation_pass_rate: pass_rate,
                overall_confidence,
                rar_loops_executed: 1,
            },
        })
    }

    /// Fast research pipeline — skips URL fetching and AI synthesis.
    ///
    /// Produces results from search result snippets only, achieving
    /// ~2-3s response time using only free/self-hosted backends.
    /// Used when `ai_synthesis = false` (via `--no-ai` flag).
    async fn execute_fast(
        config: &ResearchConfig,
        fetchium_config: &FetchiumConfig,
        http_client: &HttpClient,
        start: Instant,
    ) -> Result<ResearchReport, FetchiumError> {
        let sub_queries = decompose_query(&config.query);

        // Search with tight timeout for speed, larger budget for diversity
        let fast_budget = (config.max_sources as u32) * 3;
        let mut orch_config =
            OrchestratorConfig::from_fetchium_config(fetchium_config, fast_budget);
        orch_config.max_total_results = fast_budget;
        orch_config.backend_timeout = std::time::Duration::from_secs(5);
        let orchestrator =
            std::sync::Arc::new(SearchOrchestrator::new(http_client.clone(), orch_config));

        // Fire all sub-queries in parallel for broader coverage
        let mut search_handles = Vec::new();
        for sq in &sub_queries {
            let orch = orchestrator.clone();
            let sq = sq.clone();
            let budget = fast_budget;
            search_handles.push(tokio::spawn(async move {
                orch.search(&sq, Some(budget)).await.unwrap_or_default()
            }));
        }

        let mut raw_results = Vec::new();
        let mut seen_urls = std::collections::HashSet::new();
        for handle in search_handles {
            if let Ok(results) = handle.await {
                for r in results {
                    if seen_urls.insert(r.url.clone()) {
                        raw_results.push(r);
                    }
                }
            }
        }

        let search_results = Self::diversify_sources(raw_results, config.max_sources);

        // Pre-filter by relevance before building output
        let query_kws_pre: Vec<String> = config
            .query
            .split_whitespace()
            .map(|w| {
                w.to_lowercase()
                    .trim_matches(|c: char| !c.is_alphanumeric())
                    .to_string()
            })
            .filter(|w| w.len() > 3)
            .collect();

        const FAST_GENERIC: &[&str] = &[
            "software",
            "engineering",
            "development",
            "system",
            "technology",
            "computer",
            "application",
            "research",
            "study",
            "analysis",
            "paper",
            "using",
            "based",
            "approach",
            "model",
            "large",
            "deep",
            "data",
            "work",
            "show",
            "2024",
            "2025",
            "2026",
        ];

        let distinctive_fast: Vec<&String> = query_kws_pre
            .iter()
            .filter(|w| !FAST_GENERIC.contains(&w.as_str()))
            .collect();

        let search_results: Vec<_> = if !distinctive_fast.is_empty() {
            let filtered: Vec<_> = search_results
                .iter()
                .filter(|item| {
                    let haystack = format!(
                        "{} {}",
                        item.title.to_lowercase(),
                        item.snippet
                            .chars()
                            .take(500)
                            .collect::<String>()
                            .to_lowercase()
                    );
                    distinctive_fast
                        .iter()
                        .any(|kw| haystack.contains(kw.as_str()))
                })
                .cloned()
                .collect();
            // If filter is too aggressive (< 3 results), use unfiltered
            if filtered.len() >= 3 {
                filtered
            } else {
                search_results
            }
        } else {
            search_results
        };

        // Build sources directly from search snippets (no URL fetching)
        let formatter = CitationFormatter::new(config.citation_style);
        let mut source_metas = Vec::new();
        let mut formatted_citations = Vec::new();
        let mut extracted_texts = Vec::new();

        for (idx, item) in search_results.into_iter().enumerate() {
            let author = url::Url::parse(&item.url).ok().and_then(|u| {
                u.host_str()
                    .map(|h| h.trim_start_matches("www.").to_string())
            });
            let meta = SourceMeta {
                url: item.url.clone(),
                title: item.title.clone(),
                author,
                publisher: None,
                published_date: item.published_date.or_else(|| Some("2026".into())),
                accessed_date: chrono::Utc::now().to_rfc3339(),
            };
            source_metas.push(meta.clone());
            formatted_citations.push(formatter.format(&meta, idx + 1));
            extracted_texts.push(item.snippet.clone());
        }

        // Build extractive synthesis from snippets
        let mut synthesis = String::new();
        if source_metas.is_empty() {
            synthesis = format!(
                "No sources found for: **{}**\n\nTry broadening your query or checking your SearXNG connection.",
                config.query
            );
        } else {
            let query_kws: Vec<String> = config
                .query
                .split_whitespace()
                .map(|w| {
                    w.to_lowercase()
                        .trim_matches(|c: char| !c.is_alphanumeric())
                        .to_string()
                })
                .filter(|w| w.len() > 3)
                .collect();

            // Build a structured summary grouping related findings
            synthesis.push_str("## Overview\n\n");

            // Collect all key sentences for a combined overview
            let mut overview_sentences: Vec<(usize, String)> = Vec::new();
            for (idx, _meta) in source_metas.iter().enumerate() {
                let text = extracted_texts.get(idx).map(|s| s.as_str()).unwrap_or("");
                if !text.is_empty() {
                    let key = extract_key_sentences(text, &query_kws, 2);
                    if !key.is_empty() {
                        overview_sentences.push((idx, key));
                    }
                }
            }

            // Write overview paragraph with inline citations
            if !overview_sentences.is_empty() {
                for (idx, sentence) in &overview_sentences {
                    let marker = &formatted_citations[*idx].inline_marker;
                    synthesis.push_str(&format!("- {} {}\n", sentence, marker));
                }
                synthesis.push('\n');
            }

            // Source details section
            synthesis.push_str("## Sources\n\n");
            for (idx, meta) in source_metas.iter().enumerate() {
                let text = extracted_texts.get(idx).map(|s| s.as_str()).unwrap_or("");
                let marker = &formatted_citations[idx].inline_marker;
                let preview: String = text.chars().take(200).collect();
                synthesis.push_str(&format!(
                    "{}. **{}** — {} {}\n",
                    idx + 1,
                    meta.title,
                    if preview.is_empty() {
                        "No excerpt available"
                    } else {
                        &preview
                    },
                    marker,
                ));
            }
        }

        let reference_section = formatter.format_references(&source_metas);
        let sources_count = source_metas.len();
        let duration_ms = start.elapsed().as_millis() as u64;

        Ok(ResearchReport {
            query: config.query.clone(),
            sub_queries,
            synthesis,
            sources: source_metas,
            citations: formatted_citations,
            reference_section,
            validation: crate::validate::types::ValidationResult {
                layers_run: vec![],
                layer_results: vec![],
                passed: true,
                confidence: 0.5,
                contradictions: vec![],
                consensus: vec![],
                mode: config.validation_mode,
            },
            evidence_graph: None,
            rar_iterations: vec![],
            meta: ResearchMeta {
                duration_ms,
                sources_fetched: sources_count,
                sources_validated: sources_count,
                validation_pass_rate: 1.0,
                overall_confidence: 0.5,
                rar_loops_executed: 0,
            },
        })
    }
}

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

    #[tokio::test]
    #[ignore = "requires network-backed search providers"]
    async fn pipeline_executes_without_error() {
        let config = ResearchConfig {
            query: "what is Rust".into(),
            thinking: false,
            ..Default::default()
        };
        let fetchium_config = crate::config::FetchiumConfig::default();
        let http = crate::http::client::HttpClient::new(&fetchium_config).unwrap();
        let report = ResearchPipeline::execute(&config, &fetchium_config, &http)
            .await
            .unwrap();
        assert_eq!(report.query, "what is Rust");
        assert!(!report.sub_queries.is_empty());
    }

    #[tokio::test]
    #[ignore = "requires network-backed search providers"]
    async fn pipeline_builds_egp_when_requested() {
        let config = ResearchConfig {
            query: "Rust vs Go".into(),
            evidence_graph: true,
            thinking: false,
            ..Default::default()
        };
        let fetchium_config = crate::config::FetchiumConfig::default();
        let http = crate::http::client::HttpClient::new(&fetchium_config).unwrap();
        let report = ResearchPipeline::execute(&config, &fetchium_config, &http)
            .await
            .unwrap();
        assert!(report.evidence_graph.is_some());
    }
}