crawlrs 0.2.0

crawlrs - Enterprise-grade web scraping and crawling platform built with Rust
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
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
// Copyright (c) 2025-2026 Kirky.X🌠
// SPDX-License-Identifier: Apache-2.0

//! LLM 提取 token 窗口分块(F8)
//!
//! 长文超出模型上下文窗口时:按窗口预算切块 → tokio JoinSet 并行提问 →
//! 单次 merge prompt 综合(可关,退化为字段级启发式拼接)→ TokenUsage 全链累加。
//!
//! 预算估算口径(零新依赖,不用 tokenizer 精确计数):
//! `budget_chars = (model_context_tokens − prompt_overhead_tokens − max_output_tokens)
//! × chars_per_token ÷ 2`(50% 余量覆盖估算偏差;CJK 场景建议把 chars_per_token 调低)。
//!
//! 切块边界三级回退:段落(`\n\n` / `\r\n\r\n`)→ 句子(`。!?.!?`)→ 按字符硬切;
//! 块间无重叠严格推进,拼接覆盖原文。
//!
//! 术语预检:schema 属性名拆词(驼峰/下划线/非字母数字切分,小写化),
//! 停用词过滤后全部不在原文中 → 确定性短路(省 token,规则 7)。

use serde_json::{json, Value};
use std::ops::Range;
use std::sync::Arc;

use crate::domain::services::llm::TokenUsage;

/// 分块提取的逐块上下文说明(让 LLM 对本块缺失字段输出 null)
const CHUNK_PART_NOTE: &str = "Note: this text is part {i} of {n} of a longer document. \
If the requested information is not present in this part, set the corresponding fields to null.";

/// merge prompt 指令前缀(复用 build_prompt 的 {{text}} 注入,schema 由模板注入)
const MERGE_INSTRUCTION: &str = "The following JSON array contains partial extraction results \
produced from consecutive parts of a longer document. Combine them into a single result that \
matches the requested schema: choose the best non-null value for each field, concatenate arrays \
for list fields, and output null for fields that are null in every part.\n\nPartial results:\n";

/// 内置停用词表:schema 通用属性名不参与术语预检(防误杀)
const TERM_STOPWORDS: &[&str] = &[
    "id",
    "name",
    "type",
    "value",
    "data",
    "time",
    "url",
    "title",
    "text",
    "key",
    "date",
    "count",
    "index",
    "description",
    "content",
    "items",
    "properties",
];

/// 分块配置(由 `[llm.chunking]` 七键 + enabled 构造)
#[derive(Debug, Clone, PartialEq)]
pub struct ChunkingConfig {
    /// 总开关:false 时长文单 prompt 直发(现状,向后兼容)
    pub enabled: bool,
    /// 模型上下文窗口(token)
    pub model_context_tokens: u32,
    /// 单次回答预留输出 token
    pub max_output_tokens: u32,
    /// prompt 模板/边界标记等注入开销预留 token
    pub prompt_overhead_tokens: u32,
    /// token→字符保守换算(÷2 余量在 [`budget_chars`] 内固定)
    pub chars_per_token: f64,
    /// JoinSet 并行上限
    pub max_parallel: usize,
    /// 分块结果单次 merge prompt 综合(false 时逐块结果直接按字段拼接)
    pub merge_enabled: bool,
    /// 术语预检短路开关
    pub term_precheck_enabled: bool,
}

impl Default for ChunkingConfig {
    fn default() -> Self {
        Self {
            enabled: false,
            model_context_tokens: 8192,
            max_output_tokens: 2048,
            prompt_overhead_tokens: 1024,
            chars_per_token: 4.0,
            max_parallel: 4,
            merge_enabled: true,
            term_precheck_enabled: true,
        }
    }
}

impl ChunkingConfig {
    /// 从 `[llm.chunking]` 配置构造
    pub fn from_settings(settings: &crate::config::settings::LlmChunkingSettings) -> Self {
        Self {
            enabled: settings.enabled,
            model_context_tokens: settings.model_context_tokens,
            max_output_tokens: settings.max_output_tokens,
            prompt_overhead_tokens: settings.prompt_overhead_tokens,
            chars_per_token: settings.chars_per_token,
            max_parallel: settings.max_parallel,
            merge_enabled: settings.merge_enabled,
            term_precheck_enabled: settings.term_precheck_enabled,
        }
    }
}

/// 分块预算(字符):(窗口 − 开销 − 输出) × chars_per_token ÷ 2,下限 1
pub fn budget_chars(cfg: &ChunkingConfig) -> usize {
    let avail = cfg
        .model_context_tokens
        .saturating_sub(cfg.prompt_overhead_tokens)
        .saturating_sub(cfg.max_output_tokens);
    let est = (avail as f64) * cfg.chars_per_token / 2.0;
    (est as usize).max(1)
}

/// 按三级边界把超预算文本切块(段落 → 句子 → 硬切)
///
/// 块间无重叠严格推进,拼接覆盖原文;每块字符数 ≤ 预算。
/// 文本未超预算时返回覆盖全文的单块。
pub fn plan_chunks(text: &str, cfg: &ChunkingConfig) -> Vec<Range<usize>> {
    let budget = budget_chars(cfg);
    if text.chars().count() <= budget {
        return vec![Range {
            start: 0,
            end: text.len(),
        }];
    }

    let mut chunks: Vec<Range<usize>> = Vec::new();
    let mut chunk_start: usize = 0;
    let mut chunk_chars: usize = 0;

    for seg in paragraph_segments(text) {
        let seg_chars = text[seg.clone()].chars().count();

        if seg_chars > budget {
            // 先冲刷当前累积,再对该段做句子/硬切细分(每子块独立成块)
            if chunk_chars > 0 {
                chunks.push(chunk_start..seg.start);
                chunk_chars = 0;
            }
            emit_oversized_segment(&mut chunks, text, &seg, budget);
            chunk_start = seg.end;
            continue;
        }

        if chunk_chars > 0 && chunk_chars + seg_chars > budget {
            chunks.push(chunk_start..seg.start);
            chunk_start = seg.start;
            chunk_chars = 0;
        }
        if chunk_chars == 0 {
            chunk_start = seg.start;
        }
        chunk_chars += seg_chars;
    }

    if chunk_start < text.len() {
        chunks.push(chunk_start..text.len());
    }
    chunks
}

/// 超预算段落的子块输出(自 [`plan_chunks`] 提取):段内 tiling 区间映射回全文偏移。
fn emit_oversized_segment(
    chunks: &mut Vec<Range<usize>>,
    text: &str,
    seg: &Range<usize>,
    budget: usize,
) {
    for sub in split_oversized_segment(&text[seg.clone()], budget) {
        chunks.push(seg.start + sub.start..seg.start + sub.end);
    }
}

/// 段落切分(tiling):`\n\n` / `\r\n\r\n` 为界,分隔符归前段
fn paragraph_segments(text: &str) -> Vec<Range<usize>> {
    let bytes = text.as_bytes();
    let mut segs: Vec<Range<usize>> = Vec::new();
    let mut start = 0usize;
    let mut i = 0usize;
    while i < bytes.len() {
        let sep_len = if bytes[i..].starts_with(b"\r\n\r\n") {
            4
        } else if bytes[i..].starts_with(b"\n\n") {
            2
        } else {
            0
        };
        if sep_len > 0 {
            let end = i + sep_len;
            segs.push(start..end);
            start = end;
            i = end;
        } else {
            i += 1;
        }
    }
    if start < text.len() {
        segs.push(start..text.len());
    }
    segs
}

/// 句子切分(tiling):句末标点(含 UTF-8 字节数)归前句
fn sentence_segments(seg: &str) -> Vec<Range<usize>> {
    let mut segs: Vec<Range<usize>> = Vec::new();
    let mut start = 0usize;
    for (bi, c) in seg.char_indices() {
        if matches!(c, '。' | '!' | '?' | '.' | '!' | '?') {
            let end = bi + c.len_utf8();
            segs.push(start..end);
            start = end;
        }
    }
    if start < seg.len() {
        segs.push(start..seg.len());
    }
    segs
}

/// 超预算段落细分:句子累积,超预算句子按字符硬切;返回段内 tiling 子区间
fn split_oversized_segment(seg: &str, budget: usize) -> Vec<Range<usize>> {
    let mut pieces: Vec<Range<usize>> = Vec::new();
    let mut acc_start = 0usize;
    let mut acc_chars = 0usize;

    for s in sentence_segments(seg) {
        let s_chars = seg[s.clone()].chars().count();

        if s_chars > budget {
            if acc_chars > 0 {
                pieces.push(acc_start..s.start);
                acc_chars = 0;
            }
            hard_split_sentence(seg, &s, budget, &mut pieces);
            acc_start = s.end;
            continue;
        }

        if acc_chars > 0 && acc_chars + s_chars > budget {
            pieces.push(acc_start..s.start);
            acc_start = s.start;
            acc_chars = 0;
        }
        if acc_chars == 0 {
            acc_start = s.start;
        }
        acc_chars += s_chars;
    }

    if acc_start < seg.len() && acc_chars > 0 {
        pieces.push(acc_start..seg.len());
    }
    pieces
}

/// 超预算句子的字符硬切:每 `budget` 个字符一块(UTF-8 字符边界安全),
/// 追加到 `pieces`
fn hard_split_sentence(seg: &str, s: &Range<usize>, budget: usize, pieces: &mut Vec<Range<usize>>) {
    let mut pos = s.start;
    let mut used = 0usize;
    for (bi, _) in seg[s.start..s.end].char_indices() {
        if used == budget {
            pieces.push(pos..s.start + bi);
            pos = s.start + bi;
            used = 0;
        }
        used += 1;
    }
    if pos < s.end {
        pieces.push(pos..s.end);
    }
}

/// 术语预检:schema 属性名拆词(停用词过滤后)全部不在文本中 → false
///
/// 无属性名 / 全停用词 → true(不误杀,跳过判定)。
///
/// 匹配按词元逐个做 ASCII 大小写不敏感滑窗查找(`split_terms` 产出的词元
/// 恒为 ASCII 小写),不构造整篇文本的小写副本。
pub fn precheck(schema: &Value, text: &str) -> bool {
    let mut terms = Vec::new();
    collect_schema_terms(schema, &mut terms);
    let meaningful: Vec<&String> = terms
        .iter()
        .filter(|t| !TERM_STOPWORDS.contains(&t.as_str()))
        .collect();
    if meaningful.is_empty() {
        return true;
    }
    meaningful
        .iter()
        .all(|t| contains_ascii_ci(text, t.as_str()))
}

/// ASCII 大小写不敏感子串查找(无分配;`needle` 须为 ASCII——`split_terms` 保证)
fn contains_ascii_ci(haystack: &str, needle: &str) -> bool {
    let n = needle.as_bytes();
    if n.is_empty() {
        return true;
    }
    let h = haystack.as_bytes();
    if n.len() > h.len() {
        return false;
    }
    h.windows(n.len()).any(|w| w.eq_ignore_ascii_case(n))
}

/// 递归收集 schema `properties` 的属性名
fn collect_schema_terms(schema: &Value, out: &mut Vec<String>) {
    match schema {
        Value::Object(map) => {
            if let Some(Value::Object(props)) = map.get("properties") {
                for (name, prop_schema) in props {
                    out.extend(split_terms(name));
                    collect_schema_terms(prop_schema, out);
                }
            }
            if let Some(items) = map.get("items") {
                collect_schema_terms(items, out);
            }
        }
        Value::Array(arr) => {
            for v in arr {
                collect_schema_terms(v, out);
            }
        }
        _ => {}
    }
}

/// 属性名拆词:按非字母数字 + 驼峰边界切分,小写化
fn split_terms(name: &str) -> Vec<String> {
    let mut terms = Vec::new();
    for seg in name.split(|c: char| !c.is_ascii_alphanumeric()) {
        if seg.is_empty() {
            continue;
        }
        split_camel_segment_into(seg, &mut terms);
    }
    terms
}

/// 段内驼峰边界切分并小写化(`FooBar` → `["foo", "bar"]`),追加到 `terms`
fn split_camel_segment_into(seg: &str, terms: &mut Vec<String>) {
    let mut cur = String::new();
    for c in seg.chars() {
        if !cur.is_empty() && c.is_ascii_uppercase() {
            terms.push(cur.to_lowercase());
            cur = String::new();
        }
        cur.push(c);
    }
    if !cur.is_empty() {
        terms.push(cur.to_lowercase());
    }
}

/// 字段级启发式 merge:对象逐字段取首个非 null(数组字段 concat),根数组直接 concat
pub fn merge_results(results: &[Value]) -> Value {
    if results.is_empty() {
        return json!({});
    }
    if results.iter().all(Value::is_array) {
        return concat_result_arrays(results);
    }
    let mut merged = serde_json::Map::new();
    for r in results {
        if let Value::Object(map) = r {
            merge_object_fields(&mut merged, map);
        }
    }
    Value::Object(merged)
}

/// 根数组结果按原顺序拼接为单个数组
fn concat_result_arrays(results: &[Value]) -> Value {
    let mut merged = Vec::new();
    for r in results {
        if let Value::Array(items) = r {
            merged.extend(items.iter().cloned());
        }
    }
    Value::Array(merged)
}

/// 将单个对象的字段并入聚合 map:数组字段拼接,其余取首个非 null
fn merge_object_fields(
    merged: &mut serde_json::Map<String, Value>,
    map: &serde_json::Map<String, Value>,
) {
    for (k, v) in map {
        match merged.get(k) {
            Some(Value::Array(existing)) if v.is_array() => {
                let mut concat = existing.clone();
                if let Value::Array(items) = v {
                    concat.extend(items.iter().cloned());
                }
                merged.insert(k.clone(), Value::Array(concat));
            }
            Some(Value::Null) | None => {
                merged.insert(k.clone(), v.clone());
            }
            Some(Value::Array(existing)) if v.is_null() => {
                // 已聚合数组 vs null:保留聚合结果
                merged.insert(k.clone(), Value::Array(existing.clone()));
            }
            _ => {
                // 已有非 null 值:保留首个非 null
            }
        }
    }
}

/// 分块提取编排:术语预检 → 预算判定 → 切块并行提取 → merge 综合
///
/// 回调由调用方注入(宿主 carrier 接 provider 路径;测试接 mock),使编排逻辑
/// 与具体 provider 解耦:
/// - `extract_one`:单次提取(切块后的文本 → 结构化结果)
/// - `merge_call`:merge prompt 调用(指令 + 块结果 JSON 数组 → 最终结果)
///
/// TokenUsage 累加 = Σ块 + merge 次调用。
/// 任一块提取失败 → 整体 Err(失败显性化,由上层 retry_policy 承接)。
pub async fn extract_chunked<E, FutE, M, FutM>(
    text: &str,
    schema: &Value,
    cfg: &ChunkingConfig,
    extract_one: E,
    merge_call: M,
) -> anyhow::Result<(Value, TokenUsage)>
where
    E: Fn(String) -> FutE + Clone + Send + 'static,
    FutE: std::future::Future<Output = anyhow::Result<(Value, TokenUsage)>> + Send,
    M: Fn(String) -> FutM + Clone + Send + 'static,
    FutM: std::future::Future<Output = anyhow::Result<(Value, TokenUsage)>> + Send,
{
    // 术语预检:确定性短路(省 token)
    if cfg.term_precheck_enabled && !precheck(schema, text) {
        log::warn!(
            "llm_chunking: term precheck failed — schema terms absent from text, short-circuit"
        );
        return Ok((json!({}), TokenUsage::default()));
    }

    // 未超预算:单发直走(现状语义)
    if text.chars().count() <= budget_chars(cfg) {
        return extract_one(text.to_string()).await;
    }

    let (chunk_texts, total) = plan_chunk_texts(text, cfg);
    log::info!(
        "llm_chunking: text_len={} chars, chunks={}, max_parallel={}, merge_enabled={}",
        text.chars().count(),
        total,
        cfg.max_parallel,
        cfg.merge_enabled
    );

    let per_chunk = extract_chunks(chunk_texts, extract_one, cfg.max_parallel).await?;
    let (chunk_values, mut total_usage) = collect_chunk_results(per_chunk);

    // merge:单次 merge prompt 综合;失败或关闭时确定性退化启发式拼接
    if cfg.merge_enabled && chunk_values.len() > 1 {
        if let Some(merged) =
            merge_chunk_results(&chunk_values, merge_call, &mut total_usage).await?
        {
            return Ok((merged, total_usage));
        }
    }

    Ok((merge_results(&chunk_values), total_usage))
}

/// 切块并在每块前加逐块上下文说明;返回(块文本, 块数)
fn plan_chunk_texts(text: &str, cfg: &ChunkingConfig) -> (Vec<String>, usize) {
    let ranges = plan_chunks(text, cfg);
    let total = ranges.len();
    let chunk_texts = ranges
        .iter()
        .enumerate()
        .map(|(i, r)| {
            format!(
                "{note}\n\n{chunk}",
                note = CHUNK_PART_NOTE
                    .replace("{i}", &(i + 1).to_string())
                    .replace("{n}", &total.to_string()),
                chunk = &text[r.clone()],
            )
        })
        .collect();
    (chunk_texts, total)
}

/// 并行块提取(JoinSet + 信号量限流,并发峰值 ≤ max_parallel),结果按块序回填
async fn extract_chunks<E, FutE>(
    chunk_texts: Vec<String>,
    extract_one: E,
    max_parallel: usize,
) -> anyhow::Result<Vec<(Value, TokenUsage)>>
where
    E: Fn(String) -> FutE + Clone + Send + 'static,
    FutE: std::future::Future<Output = anyhow::Result<(Value, TokenUsage)>> + Send,
{
    // JoinSet + 信号量限流:并发峰值 ≤ max_parallel
    let semaphore = Arc::new(tokio::sync::Semaphore::new(max_parallel.max(1)));
    let mut join_set: tokio::task::JoinSet<(usize, anyhow::Result<(Value, TokenUsage)>)> =
        tokio::task::JoinSet::new();
    let total = chunk_texts.len();
    for (idx, chunk_text) in chunk_texts.into_iter().enumerate() {
        let permit_sem = semaphore.clone();
        let extract_one = extract_one.clone();
        join_set.spawn(async move {
            let _permit = permit_sem.acquire_owned().await;
            (idx, extract_one(chunk_text).await)
        });
    }

    let mut per_chunk: Vec<Option<(Value, TokenUsage)>> = (0..total).map(|_| None).collect();
    while let Some(joined) = join_set.join_next().await {
        let (idx, result) = joined.map_err(|e| anyhow::anyhow!("chunk task panicked: {e}"))?;
        let (value, usage) = result?;
        per_chunk[idx] = Some((value, usage));
    }
    Ok(per_chunk.into_iter().flatten().collect())
}

/// 按块序收集结果值并累加 TokenUsage(Σ块)
fn collect_chunk_results(per_chunk: Vec<(Value, TokenUsage)>) -> (Vec<Value>, TokenUsage) {
    let mut total_usage = TokenUsage::default();
    let mut chunk_values: Vec<Value> = Vec::with_capacity(per_chunk.len());
    for (value, usage) in per_chunk {
        total_usage.prompt_tokens += usage.prompt_tokens;
        total_usage.completion_tokens += usage.completion_tokens;
        total_usage.total_tokens += usage.total_tokens;
        chunk_values.push(value);
    }
    (chunk_values, total_usage)
}

/// merge 调用:成功返回综合结果并累加其 TokenUsage;失败记 warn 并返回 None
/// (调用方退化为确定性启发式拼接)
async fn merge_chunk_results<M, FutM>(
    chunk_values: &[Value],
    merge_call: M,
    total_usage: &mut TokenUsage,
) -> anyhow::Result<Option<Value>>
where
    M: Fn(String) -> FutM,
    FutM: std::future::Future<Output = anyhow::Result<(Value, TokenUsage)>> + Send,
{
    let payload = format!(
        "{instruction}{results}",
        instruction = MERGE_INSTRUCTION,
        results = serde_json::to_string(chunk_values)?,
    );
    match merge_call(payload).await {
        Ok((merged, usage)) => {
            total_usage.prompt_tokens += usage.prompt_tokens;
            total_usage.completion_tokens += usage.completion_tokens;
            total_usage.total_tokens += usage.total_tokens;
            Ok(Some(merged))
        }
        Err(e) => {
            log::warn!("llm_chunking: merge call failed ({e}); falling back to heuristic merge");
            Ok(None)
        }
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use std::sync::atomic::{AtomicUsize, Ordering as AtomicOrdering};
    use std::sync::Mutex;

    // ===== 预算算术 =====

    /// 默认配置代入:(8192 − 1024 − 2048) × 4 ÷ 2 = 10240 字符
    #[test]
    fn budget_chars_default_config_arithmetic() {
        let cfg = ChunkingConfig::default();
        assert_eq!(budget_chars(&cfg), 10240);
    }

    /// 预算下限 1(退化配置不死循环/不 panic)
    #[test]
    fn budget_chars_floors_at_one() {
        let cfg = ChunkingConfig {
            model_context_tokens: 100,
            prompt_overhead_tokens: 200,
            max_output_tokens: 300,
            ..ChunkingConfig::default()
        };
        assert_eq!(budget_chars(&cfg), 1);
    }

    // ===== plan_chunks =====

    /// 未超预算文本返回覆盖全文的单块
    #[test]
    fn plan_chunks_single_block_when_under_budget() {
        let cfg = cfg_with_budget(100);
        let text = "short text".to_string();
        let chunks = plan_chunks(&text, &cfg);
        assert_eq!(chunks, vec![0..text.len()]);
    }

    /// 超预算文本:块数 = ceil(总字符/预算),每块 ≤ 预算,拼接覆盖原文(无缝无叠)
    #[test]
    fn plan_chunks_chunk_count_matches_ceil_and_tiles_text() {
        // 40 段 × (10 字符 + "\n\n"),末段无分隔符 → 478 字符;预算 100 → ceil = 5
        let mut text = String::new();
        for i in 0..40 {
            text.push_str(&format!("para-{:02}-x\n\n", i));
        }
        let text = text.trim_end_matches('\n').to_string();
        let cfg = cfg_with_budget(100);
        let chunks = plan_chunks(&text, &cfg);
        assert_eq!(
            chunks.len(),
            text.chars().count().div_ceil(100),
            "块数应为 ceil(len/budget)"
        );
        assert_tiles(text.as_str(), &chunks);
    }

    /// 段落边界优先:块切在 `\n\n` 处
    #[test]
    fn plan_chunks_prefers_paragraph_boundaries() {
        let cfg = cfg_with_budget(30);
        let text = "aa bb cc dd\n\nee ff gg hh\n\nii jj kk ll mm nn";
        let chunks = plan_chunks(text, &cfg);
        assert_tiles(text, &chunks);
        // 每块末尾应是段落分隔符或文本结尾
        for (i, r) in chunks.iter().enumerate() {
            let seg = &text[r.clone()];
            let is_last = i == chunks.len() - 1;
            assert!(
                is_last || seg.ends_with("\n\n"),
                "块应切在段落边界: {seg:?}"
            );
        }
    }

    /// CRLF 文本(Windows 行尾)段落边界同样生效
    #[test]
    fn plan_chunks_handles_crlf_paragraphs() {
        let cfg = cfg_with_budget(20);
        let text = "aa bb cc\r\n\r\ndd ee ff\r\n\r\ngg hh ii";
        let chunks = plan_chunks(text, &cfg);
        assert_tiles(text, &chunks);
    }

    /// 无段落时切句子边界
    #[test]
    fn plan_chunks_falls_back_to_sentence_boundaries() {
        let cfg = cfg_with_budget(25);
        let text = "First sentence here. Second sentence follows. Third one is longer!";
        let chunks = plan_chunks(text, &cfg);
        assert!(chunks.len() > 1, "超预算应多块");
        assert_tiles(text, &chunks);
        // 块应尽量切在句末标点后
        let last = chunks.last().unwrap();
        assert!(text[last.clone()].ends_with('!'));
    }

    /// 超长无标点句硬切:每块 ≤ 预算且拼接覆盖原文
    #[test]
    fn plan_chunks_hard_cuts_oversized_sentence() {
        let cfg = cfg_with_budget(10);
        let text = "abcdefghij".to_string();
        let text = format!("{text}{text}{text}"); // 30 字符无段落无句末标点
        let chunks = plan_chunks(&text, &cfg);
        assert_eq!(chunks.len(), 3);
        assert_tiles(&text, &chunks);
        for r in &chunks {
            assert!(text[r.clone()].chars().count() <= 10);
        }
    }

    /// UTF-8 字符边界安全:CJK 多字节字符硬切不 panic、不撕裂
    #[test]
    fn plan_chunks_hard_cut_respects_utf8_boundaries() {
        let cfg = cfg_with_budget(3);
        let text = "一二三四五六七八九十".to_string(); // 10 个 CJK 字符
        let chunks = plan_chunks(&text, &cfg);
        assert_tiles(&text, &chunks);
        for r in &chunks {
            assert!(text[r.clone()].chars().count() <= 3);
        }
    }

    /// 混合场景:段落内含超预算超长句
    #[test]
    fn plan_chunks_mixed_paragraph_with_oversized_sentence() {
        let cfg = cfg_with_budget(8);
        let text = format!("short.\n\n{}", "x".repeat(30));
        let chunks = plan_chunks(&text, &cfg);
        assert_tiles(&text, &chunks);
        for r in &chunks {
            assert!(text[r.clone()].chars().count() <= 8, "块 {:?} 超预算", r);
        }
    }

    /// ChunkingConfig::default() 是禁用态但保留默认预算参数(与 default.toml 对齐)
    #[test]
    fn default_config_is_disabled_with_spec_defaults() {
        let cfg = ChunkingConfig::default();
        assert!(!cfg.enabled);
        assert_eq!(cfg.model_context_tokens, 8192);
        assert_eq!(cfg.max_output_tokens, 2048);
        assert_eq!(cfg.prompt_overhead_tokens, 1024);
        assert!((cfg.chars_per_token - 4.0).abs() < f64::EPSILON);
        assert_eq!(cfg.max_parallel, 4);
        assert!(cfg.merge_enabled);
        assert!(cfg.term_precheck_enabled);
    }

    // ===== 术语预检 =====

    /// schema 字段 product_price 在含 "product price" 文本中通过
    #[test]
    fn precheck_passes_when_terms_present_in_text() {
        let schema = json!({
            "type": "object",
            "properties": { "product_price": { "type": "number" } }
        });
        assert!(precheck(&schema, "The product price is 42 dollars."));
    }

    /// 驼峰命名等价拆词
    #[test]
    fn precheck_splits_camel_case_terms() {
        let schema = json!({
            "type": "object",
            "properties": { "productPrice": { "type": "number" } }
        });
        assert!(precheck(&schema, "the product price"));
    }

    /// 字段 qwzx 不在文本中 → false
    #[test]
    fn precheck_fails_when_terms_absent_from_text() {
        let schema = json!({
            "type": "object",
            "properties": { "qwzx": { "type": "string" } }
        });
        assert!(!precheck(&schema, "nothing relevant here at all."));
    }

    /// 嵌套 schema 属性名递归收集
    #[test]
    fn precheck_collects_nested_property_names() {
        let schema = json!({
            "type": "object",
            "properties": {
                "author": {
                    "type": "object",
                    "properties": { "biography": { "type": "string" } }
                }
            }
        });
        assert!(precheck(&schema, "author biography mentioned"));
        assert!(!precheck(&schema, "neither word appears here"));
    }

    /// 停用词:schema 仅含 {"id","name"} 时不误杀(全停用词跳过判定)
    #[test]
    fn precheck_all_stopwords_skips_verdict() {
        let schema = json!({
            "type": "object",
            "properties": {
                "id": { "type": "string" },
                "name": { "type": "string" }
            }
        });
        assert!(precheck(&schema, "completely unrelated text"));
    }

    /// 无属性名的 schema(如 {"type":"array"})不误杀
    #[test]
    fn precheck_schema_without_properties_passes() {
        let schema = json!({ "type": "array", "description": "list of items" });
        assert!(precheck(&schema, "any text"));
    }

    // ===== 启发式 merge =====

    /// 对象:逐字段取首个非 null
    #[test]
    fn merge_results_takes_first_non_null_per_field() {
        let results = vec![
            json!({"title": "from chunk 1", "x": null}),
            json!({"title": "from chunk 2"}),
            json!({"title": null, "y": 2}),
        ];
        let merged = merge_results(&results);
        assert_eq!(merged["title"], "from chunk 1");
        assert_eq!(merged["x"], Value::Null);
        assert_eq!(merged["y"], 2);
    }

    /// 数组字段 concat
    #[test]
    fn merge_results_concatenates_array_fields() {
        let results = vec![json!({"items": [1, 2]}), json!({"items": [3]})];
        let merged = merge_results(&results);
        assert_eq!(merged["items"], json!([1, 2, 3]));
    }

    /// 根数组 schema:直接 concat
    #[test]
    fn merge_results_concatenates_root_arrays() {
        let results = vec![json!([{"a": 1}]), json!([{"b": 2}])];
        let merged = merge_results(&results);
        assert_eq!(merged, json!([{"a": 1}, {"b": 2}]));
    }

    /// 空结果退化空对象
    #[test]
    fn merge_results_empty_input_yields_empty_object() {
        assert_eq!(merge_results(&[]), json!({}));
    }

    // ===== extract_chunked 编排 =====

    /// 未超预算:单发直走,回调收到原文(不切块、不加块注记)
    #[tokio::test]
    async fn extract_chunked_single_shot_when_under_budget() {
        let cfg = cfg_with_budget(1000);
        let calls = Arc::new(Mutex::new(Vec::<String>::new()));
        let calls_c = calls.clone();
        let (value, usage) = extract_chunked(
            "plain short text about products",
            &json!({"type": "object", "properties": {"product": {"type": "string"}}}),
            &cfg,
            move |text: String| {
                let calls = calls_c.clone();
                async move {
                    calls.lock().unwrap().push(text);
                    Ok((json!({"product": "p"}), usage_of(10, 5)))
                }
            },
            move |_payload: String| async { anyhow::bail!("merge should not be called") },
        )
        .await
        .unwrap();
        assert_eq!(value, json!({"product": "p"}));
        assert_eq!(usage.total_tokens, 15);
        let calls = calls.lock().unwrap();
        assert_eq!(calls.len(), 1);
        assert_eq!(calls[0], "plain short text about products");
    }

    /// 术语预检失败:确定性短路,不调用 provider,返回空对象零用量
    #[tokio::test]
    async fn extract_chunked_precheck_short_circuits() {
        let cfg = cfg_with_budget(1000);
        let calls = Arc::new(AtomicUsize::new(0));
        let calls_c = calls.clone();
        let (value, usage) = extract_chunked(
            "nothing relevant here",
            &json!({"type": "object", "properties": {"qwzx": {"type": "string"}}}),
            &cfg,
            move |_text: String| {
                let calls = calls_c.clone();
                async move {
                    calls.fetch_add(1, AtomicOrdering::SeqCst);
                    Ok((json!({}), usage_of(1, 1)))
                }
            },
            move |_payload: String| async { anyhow::bail!("merge should not be called") },
        )
        .await
        .unwrap();
        assert_eq!(value, json!({}));
        assert_eq!(usage.total_tokens, 0);
        assert_eq!(calls.load(AtomicOrdering::SeqCst), 0, "provider 不应被调用");
    }

    /// term_precheck_enabled=false 时不短路
    #[tokio::test]
    async fn extract_chunked_precheck_disabled_proceeds() {
        let cfg = ChunkingConfig {
            term_precheck_enabled: false,
            ..cfg_with_budget(1000)
        };
        let calls = Arc::new(AtomicUsize::new(0));
        let calls_c = calls.clone();
        let _ = extract_chunked(
            "nothing relevant here",
            &json!({"type": "object", "properties": {"qwzx": {"type": "string"}}}),
            &cfg,
            move |_text: String| {
                let calls = calls_c.clone();
                async move {
                    calls.fetch_add(1, AtomicOrdering::SeqCst);
                    Ok((json!({"qwzx": "found"}), usage_of(1, 1)))
                }
            },
            move |_payload: String| async { anyhow::bail!("merge should not be called") },
        )
        .await
        .unwrap();
        assert_eq!(calls.load(AtomicOrdering::SeqCst), 1);
    }

    /// 分块并行:每块都带 part 注记,并发峰值 ≤ max_parallel(且确实并行)
    #[tokio::test]
    async fn extract_chunked_runs_parallel_within_max_parallel() {
        let cfg = ChunkingConfig {
            max_parallel: 2,
            ..cfg_with_budget(20)
        };
        // 4 段 × (17 字符 + "\n\n"),预算 20 → 每段一段一块,共 4 块
        let text = "aaaa bbbb cccc dd\neeee ffff gggg hh\niiii jjjj kkkk ll\nmmmm nnnn oooo pp"
            .to_string();
        // schema 用停用词字段,绕过术语预检(预检行为由专属测试覆盖)
        let schema = json!({"type": "object", "properties": {"text": {"type": "string"}}});

        let calls = Arc::new(AtomicUsize::new(0));
        let inflight = Arc::new(AtomicUsize::new(0));
        let peak = Arc::new(AtomicUsize::new(0));
        let texts = Arc::new(Mutex::new(Vec::<String>::new()));

        let (calls_c, inflight_c, peak_c, texts_c) =
            (calls.clone(), inflight.clone(), peak.clone(), texts.clone());
        let (value, usage) = extract_chunked(
            &text,
            &schema,
            &cfg,
            move |chunk: String| {
                let (calls, inflight, peak, texts) = (
                    calls_c.clone(),
                    inflight_c.clone(),
                    peak_c.clone(),
                    texts_c.clone(),
                );
                async move {
                    calls.fetch_add(1, AtomicOrdering::SeqCst);
                    let now = inflight.fetch_add(1, AtomicOrdering::SeqCst) + 1;
                    peak.fetch_max(now, AtomicOrdering::SeqCst);
                    tokio::time::sleep(std::time::Duration::from_millis(20)).await;
                    inflight.fetch_sub(1, AtomicOrdering::SeqCst);
                    texts.lock().unwrap().push(chunk);
                    Ok((json!({"text": "p"}), usage_of(7, 3)))
                }
            },
            move |_payload: String| async { anyhow::bail!("merge should not be called") },
        )
        .await
        .unwrap();

        assert_eq!(calls.load(AtomicOrdering::SeqCst), 4, "4 块各提取一次");
        assert!(
            peak.load(AtomicOrdering::SeqCst) <= 2,
            "并发峰值必须 ≤ max_parallel"
        );
        assert!(
            peak.load(AtomicOrdering::SeqCst) >= 2,
            "JoinSet 应真实并行(峰值 > 1)"
        );
        // 逐块注记存在
        let texts = texts.lock().unwrap();
        for (i, t) in texts.iter().enumerate() {
            assert!(t.contains("part"), "块 {i} 应带 part 注记");
        }
        // merge 回调失败 → 启发式兜底;4 块结果一致 → 取首个非 null
        assert_eq!(value, json!({"text": "p"}));
        // TokenUsage 累加 = Σ块(merge 失败退化,无 merge 用量)
        assert_eq!(usage.prompt_tokens, 28);
        assert_eq!(usage.completion_tokens, 12);
        assert_eq!(usage.total_tokens, 40);
    }

    /// merge:三块各出 {"title": ...} → merge prompt 收到 JSON 数组,返回合并结果;
    /// TokenUsage 累加 = Σ块 + merge 次调用
    #[tokio::test]
    async fn extract_chunked_merge_prompt_combines_chunk_results() {
        let cfg = ChunkingConfig {
            merge_enabled: true,
            ..cfg_with_budget(20)
        };
        // 3 段均 ≤ 预算 20(含分隔符)→ 3 块
        let text = "alpha beta gamma\n\nzeta eta theta\n\niota kappa lamb".to_string();
        let merge_payloads = Arc::new(Mutex::new(Vec::<String>::new()));
        let merge_payloads_c = merge_payloads.clone();

        let (value, usage) = extract_chunked(
            &text,
            &json!({"type": "object", "properties": {"title": {"type": "string"}}}),
            &cfg,
            move |chunk: String| {
                let title = if chunk.contains("zeta") {
                    "title-zeta"
                } else {
                    "title-other"
                };
                async move { Ok((json!({ "title": title }), usage_of(10, 2))) }
            },
            move |payload: String| {
                let payloads = merge_payloads_c.clone();
                async move {
                    payloads.lock().unwrap().push(payload);
                    Ok((json!({"title": "merged-best"}), usage_of(30, 5)))
                }
            },
        )
        .await
        .unwrap();

        assert_eq!(value, json!({"title": "merged-best"}));
        let payloads = merge_payloads.lock().unwrap();
        assert_eq!(payloads.len(), 1, "merge 只调一次");
        let payload = &payloads[0];
        assert!(payload.contains("Combine them into a single result"));
        assert!(payload.contains("\"title\""), "应含块结果 JSON 数组字段");
        assert!(
            payload.contains("title-zeta") && payload.contains("title-other"),
            "应含各块结果: {payload}"
        );
        // Σ块 = 3 × (10+2=12) = 36 + merge (30+5=35) = 71
        assert_eq!(usage.prompt_tokens, 30 + 30);
        assert_eq!(usage.completion_tokens, 6 + 5);
        assert_eq!(usage.total_tokens, 36 + 35);
    }

    /// merge_enabled=false:不调 merge prompt,启发式拼接(非 null 首个)
    #[tokio::test]
    async fn extract_chunked_merge_disabled_uses_heuristic_merge() {
        let cfg = ChunkingConfig {
            merge_enabled: false,
            ..cfg_with_budget(20)
        };
        let text = "alpha beta gamma\n\nzeta eta theta\n\niota kappa lamb".to_string();
        let (value, usage) = extract_chunked(
            &text,
            &json!({"type": "object", "properties": {"title": {"type": "string"}}}),
            &cfg,
            move |chunk: String| {
                let title = if chunk.contains("zeta") {
                    "winner-zeta"
                } else {
                    ""
                };
                async move {
                    Ok((
                        if title.is_empty() {
                            json!({"title": null})
                        } else {
                            json!({"title": title})
                        },
                        usage_of(10, 2),
                    ))
                }
            },
            move |_payload: String| async { anyhow::bail!("merge must not be called") },
        )
        .await
        .unwrap();
        assert_eq!(value, json!({"title": "winner-zeta"}), "取首个非 null");
        assert_eq!(usage.total_tokens, 36, "仅 Σ块用量");
    }

    /// merge prompt 失败 → 显性 warn + 启发式兜底(不整体失败)
    #[tokio::test]
    async fn extract_chunked_merge_failure_falls_back_to_heuristic() {
        let cfg = cfg_with_budget(20);
        let text = "alpha beta gamma\n\nzeta eta theta\n\niota kappa lamb".to_string();
        let (value, usage) = extract_chunked(
            &text,
            &json!({"type": "object", "properties": {"title": {"type": "string"}}}),
            &cfg,
            move |_chunk: String| async { Ok((json!({"title": "chunk-title"}), usage_of(10, 2))) },
            move |_payload: String| async { anyhow::bail!("merge provider down") },
        )
        .await
        .unwrap();
        assert_eq!(
            value,
            json!({"title": "chunk-title"}),
            "启发式兜底取首个非 null"
        );
        assert_eq!(usage.total_tokens, 36);
    }

    /// 任一块提取失败 → 整体 Err(失败显性化,不静默丢块)
    #[tokio::test]
    async fn extract_chunked_chunk_failure_propagates_error() {
        let cfg = cfg_with_budget(20);
        let text =
            "alpha beta gamma delta\n\nepsilon zeta eta theta\n\niota kappa lambda mu".to_string();
        let result = extract_chunked(
            &text,
            &json!({"type": "object", "properties": {"title": {"type": "string"}}}),
            &cfg,
            move |chunk: String| async move {
                if chunk.contains("epsilon") {
                    anyhow::bail!("provider 429")
                }
                Ok((json!({"title": "ok"}), usage_of(1, 1)))
            },
            move |_payload: String| async { anyhow::bail!("merge must not be called") },
        )
        .await;
        let err = result.expect_err("块失败必须整体 Err");
        assert!(err.to_string().contains("429"), "错误信息保留: {err}");
    }

    /// TokenUsage 在无块产出时为零值
    #[tokio::test]
    async fn extract_chunked_empty_usage_default() {
        let cfg = cfg_with_budget(1000);
        let (_, usage) = extract_chunked(
            "tiny text",
            &json!({"type": "object", "properties": {"product": {"type": "string"}}}),
            &cfg,
            move |_text: String| async { Ok((json!({}), TokenUsage::default())) },
            move |_payload: String| async { anyhow::bail!("merge must not be called") },
        )
        .await
        .unwrap();
        assert_eq!(usage, TokenUsage::default());
    }

    // ===== helpers =====

    fn usage_of(prompt: u32, completion: u32) -> TokenUsage {
        TokenUsage {
            prompt_tokens: prompt,
            completion_tokens: completion,
            total_tokens: prompt + completion,
        }
    }

    /// 构造预算 = n 字符的配置:(8192 − 1024 − out) × 4 ÷ 2 = n → out = 8192 − 1024 − n/2
    fn cfg_with_budget(n: usize) -> ChunkingConfig {
        ChunkingConfig {
            enabled: true,
            max_output_tokens: (8192 - 1024 - (n / 2)) as u32,
            ..ChunkingConfig::default()
        }
    }

    /// 断言块区间 tiling:无缝无叠、按序拼接覆盖原文
    fn assert_tiles(text: &str, chunks: &[Range<usize>]) {
        assert!(!chunks.is_empty(), "至少一块");
        let mut cursor = 0usize;
        for r in chunks {
            assert_eq!(r.start, cursor, "块无缝衔接");
            assert!(r.end >= r.start);
            cursor = r.end;
        }
        assert_eq!(cursor, text.len(), "块拼接覆盖全文");
        let joined: String = chunks.iter().map(|r| &text[r.clone()]).collect();
        assert_eq!(joined, text);
    }
}