docling 2.67.1

DocumentConverter and format backends for docling.rs (a Rust port of docling).
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
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
//! docling JSON backend — reads docling's native `DoclingDocument` JSON
//! serialization and re-exports it. docling just pydantic-loads the model; here
//! we walk the `body` tree (children resolved through `$ref` into the
//! `texts`/`groups`/`tables`/`pictures` arrays, skipping `furniture`) and map
//! each item onto the crate's [`Node`] model so the shared Markdown serializer
//! reproduces docling-core's output. Each item's `prov` and the document's
//! `pages` ride along (#899): the JSON export of a docling-JSON input keeps
//! the page and box provenance docling gave the original document, and the
//! DocLang export carries the `<location>` tokens and `<page_break/>`s
//! docling-core's own serializer derives from them.

use serde_json::Value;

use crate::backend::markdown::escape_text;
use crate::backend::DeclarativeBackend;
use crate::error::ConversionError;
use crate::source::SourceDocument;
use docling_core::{CaptionParent, DoclingDocument, Node, PictureImage, Table};

pub struct DoclingJsonBackend;

impl DeclarativeBackend for DoclingJsonBackend {
    fn convert(&self, source: &SourceDocument) -> Result<DoclingDocument, ConversionError> {
        let text = source.text()?;
        let mut root: Value = match serde_json::from_str(&text) {
            Ok(v) => v,
            // Not JSON at all — but not a docling document either (its
            // schema name is nowhere in the text): the bytes as a json
            // code block, as MarkItDown reads any `.json` (#671). A broken
            // DoclingDocument keeps its parse error.
            Err(_) if !text.contains("DoclingDocument") => {
                return crate::backend::text::convert_json_text(source);
            }
            Err(e) => return Err(ConversionError::with_source("docling-json", e)),
        };
        // Generic JSON (#671) — no `schema_name`, no `body` to walk — is one
        // `json` code block, not an empty document.
        if !crate::backend::text::is_docling_document(&root) {
            return crate::backend::text::convert_json_text(source);
        }
        check_table_budget(&root)?;
        // Referenced images resolve against the JSON file's directory (#403).
        inline_referenced_images(
            &mut root,
            source.path.as_deref().and_then(std::path::Path::parent),
        );
        let root = root;
        let name = root["name"].as_str().unwrap_or(&source.name).to_string();
        let mut doc = DoclingDocument::new(name);
        // The `metadata` object (#679) a docling.rs export carried.
        if root["metadata"].is_object() {
            doc.metadata = docling_core::DocumentMetadata::from_json(&root["metadata"]);
        }
        let mut guard = WalkGuard {
            pages: page_sizes(&root),
            ..WalkGuard::default()
        };
        if let Some(children) = root["body"]["children"].as_array() {
            for c in children {
                walk(c, &root, 0, &mut doc, &mut guard);
            }
        }
        // A page no body item sits on (a blank page, a furniture-only one)
        // is still a page of the document: its marker keeps it in the
        // exported `pages` map. Trailing markers render nothing — Markdown's
        // page break goes only between two rendered blocks.
        for (&page_no, &(width, height)) in &guard.pages {
            if !guard.seen_pages.contains(&page_no) {
                doc.push(Node::PageInfo {
                    page_no,
                    width,
                    height,
                });
            }
        }
        Ok(doc)
    }
}

/// Every table of the file against the cell budget (#861; docling-core
/// 2.103.1's load-time validators): the declared `num_rows × num_cols` and
/// the `grid` the file carries, whichever is larger, per table and summed
/// over `tables` — all of them, reachable or not, as upstream counts them,
/// since the dense `rows` of each is what the reader materializes.
fn check_table_budget(root: &Value) -> Result<(), ConversionError> {
    let mut budget = docling_core::limits::TableCellBudget::new();
    for (i, table) in root["tables"].as_array().into_iter().flatten().enumerate() {
        let data = &table["data"];
        let declared = |key: &str| data[key].as_u64().unwrap_or(0).min(usize::MAX as u64) as usize;
        let grid = data["grid"].as_array();
        let rows = declared("num_rows").max(grid.map_or(0, Vec::len));
        let cols = declared("num_cols").max(
            grid.into_iter()
                .flatten()
                .filter_map(Value::as_array)
                .map(Vec::len)
                .max()
                .unwrap_or(0),
        );
        budget
            .table(&format!("docling-json #/tables/{i}"), rows, cols)
            .map_err(|e| ConversionError::with_source("docling-json", e))?;
    }
    Ok(())
}

/// Resolve a `{"$ref": "#/texts/3"}` reference into its array element.
fn resolve<'a>(reference: &Value, root: &'a Value) -> Option<&'a Value> {
    let path = reference["$ref"].as_str()?.strip_prefix("#/")?;
    let (kind, idx) = path.rsplit_once('/')?;
    root.get(kind)?.get(idx.parse::<usize>().ok()?)
}

fn ref_kind(reference: &Value) -> &str {
    reference["$ref"].as_str().unwrap_or("")
}

fn text_of(reference: &Value, root: &Value) -> String {
    resolve(reference, root)
        .map(formatted_text)
        .unwrap_or_default()
}

/// Escaped item text with docling-core's inline markers applied in order:
/// bold → italic → strikethrough → hyperlink (underline/script are no-ops in
/// Markdown).
fn formatted_text(item: &Value) -> String {
    let mut res = escape_text(item["text"].as_str().unwrap_or(""));
    let fmt = &item["formatting"];
    if fmt["bold"].as_bool() == Some(true) {
        res = format!("**{res}**");
    }
    if fmt["italic"].as_bool() == Some(true) {
        res = format!("*{res}*");
    }
    if fmt["strikethrough"].as_bool() == Some(true) {
        res = format!("~~{res}~~");
    }
    if let Some(url) = item["hyperlink"].as_str() {
        res = docling_core::hyperlink::markdown_link(&res, url);
    }
    res
}

/// Dispatch one body/child reference by the array it points into.
/// Guards the tree walk against a document that is not a tree: a `children`
/// entry that points back at an ancestor recursed until the stack overflowed
/// (an abort, not an error), a chain of 50 000 nested groups did the same, and
/// one item referenced from many parents multiplied the output with every
/// level. Every `$ref` is visited once — a docling document gives each item
/// exactly one parent — and nesting stops at [`MAX_DEPTH`].
#[derive(Default)]
struct WalkGuard {
    seen: std::collections::HashSet<String>,
    depth: usize,
    /// The document's `pages` (#899): page number → size in points.
    pages: std::collections::BTreeMap<usize, (f32, f32)>,
    /// The page the last provenanced item sat on (0 before the first).
    cur_page: usize,
    /// Pages a [`Node::PageInfo`] marker was emitted for.
    seen_pages: std::collections::BTreeSet<usize>,
}

/// The document's `pages` map: `{"1": {"size": {"width", "height"}, …}}`.
fn page_sizes(root: &Value) -> std::collections::BTreeMap<usize, (f32, f32)> {
    let mut pages = std::collections::BTreeMap::new();
    if let Some(map) = root["pages"].as_object() {
        for (key, page) in map {
            let no = page["page_no"]
                .as_u64()
                .map(|n| n as usize)
                .or_else(|| key.parse().ok());
            let (w, h) = (
                page["size"]["width"].as_f64(),
                page["size"]["height"].as_f64(),
            );
            if let (Some(no), Some(w), Some(h)) = (no, w, h) {
                if no > 0 {
                    pages.insert(no, (w as f32, h as f32));
                }
            }
        }
    }
    pages
}

/// An item's first `prov` entry (#899): the page it sits on, its box in the
/// page's own coordinate origin, and its character span. docling keeps a
/// list, one entry per page an item spans; the JSON export writes one box
/// per item, so the first — the only one in practice — is what survives.
struct Prov {
    page_no: usize,
    bbox: [f32; 4],
    bottom_left: bool,
    charspan: [usize; 2],
}

fn prov_of(item: &Value) -> Option<Prov> {
    let p = item["prov"].as_array()?.first()?;
    let page_no = p["page_no"].as_u64()? as usize;
    let b = &p["bbox"];
    let coord = |k: &str| b[k].as_f64().map(|v| v as f32);
    let bbox = [coord("l")?, coord("t")?, coord("r")?, coord("b")?];
    let bottom_left = b["coord_origin"].as_str() != Some("TOPLEFT");
    let span = p["charspan"].as_array();
    let at = |i: usize| span.and_then(|s| s.get(i)?.as_u64()).unwrap_or(0) as usize;
    Some(Prov {
        page_no,
        bbox,
        bottom_left,
        charspan: [at(0), at(1)],
    })
}

/// Before an item on page `page_no`: the [`Node::PageInfo`] marker that
/// tells the JSON export which page (and page size) the following items are
/// on, once per change of page, as the PDF and XLSX paths emit it. `true`
/// when the walk is on that page afterwards. A marker between two list items
/// would split the list (the serializers group a list from adjacent items),
/// so a list that runs onto the next page keeps that page's items in the
/// list, without a location — docling's own PDF documents never do, their
/// layout groups a list per page.
fn enter_page(
    page_no: usize,
    list_item: bool,
    doc: &mut DoclingDocument,
    g: &mut WalkGuard,
) -> bool {
    if page_no == 0 {
        return false;
    }
    if page_no == g.cur_page {
        return true;
    }
    let Some(&(width, height)) = g.pages.get(&page_no) else {
        return false;
    };
    if list_item && matches!(doc.nodes.last(), Some(Node::ListItem { .. })) {
        return false;
    }
    // docling-core's `_iterate_items` yields a page break between two items
    // on different pages (DocLang's `<page_break/>`; Markdown's placeholder,
    // which absorbs the marker pair into one). Not before the first item.
    if g.cur_page != 0 {
        doc.push(Node::PageBreak);
    }
    g.cur_page = page_no;
    g.seen_pages.insert(page_no);
    doc.push(Node::PageInfo {
        page_no,
        width,
        height,
    });
    true
}

/// Wrap the node at `index` (the first node the item pushed) in the item's
/// provenance: the exact box for the JSON export (`Node::Prov`, outside —
/// the JSON export prefers it) around, when the walk is on the item's page,
/// the location grid for DocLang (`Node::Located`, inside — its serializer
/// renders the node it wraps directly).
fn wrap_prov(doc: &mut DoclingDocument, index: usize, prov: &Prov, grid: Option<[u16; 4]>) {
    let inner = std::mem::replace(&mut doc.nodes[index], Node::PageBreak);
    let inner = match grid {
        Some(location) => Node::Located {
            location,
            inner: Box::new(inner),
        },
        None => inner,
    };
    doc.nodes[index] = Node::Prov {
        page_no: prov.page_no,
        bbox: prov.bbox,
        charspan: prov.charspan,
        bottom_left: prov.bottom_left,
        seq: None,
        inner: Box::new(inner),
    };
}

/// An item's provenance on the DocLang location grid (`[x0, y0, x1, y1]` on
/// 0–511, top-left origin): docling-core's
/// `_create_location_tokens_for_bbox` at resolution 512, the same
/// quantization as the PDF pipeline's. The `<location>` tokens of every
/// item, and the whole provenance of a list item, which the `location`
/// field carries: the serializers group a list from *adjacent* `ListItem`
/// nodes, so a `Prov` wrapper around one would split the list. The JSON
/// export denormalizes the grid against the page size, so a list item's box
/// comes back to within a grid cell (page size / 512 — under a point on A4)
/// and its charspan as `[0, len]`, which is what docling writes for a PDF
/// item anyway.
fn grid_location(prov: &Prov, g: &WalkGuard) -> Option<[u16; 4]> {
    let &(w, h) = g.pages.get(&prov.page_no)?;
    if w <= 0.0 || h <= 0.0 {
        return None;
    }
    let q =
        |v: f32, dim: f32| ((512.0 * v as f64 / dim as f64).round() as i64).clamp(0, 511) as u16;
    let [l, t, r, b] = prov.bbox;
    let (top, bottom) = if prov.bottom_left {
        (h - t, h - b)
    } else {
        (t, b)
    };
    Some([q(l, w), q(top, h), q(r, w), q(bottom, h)])
}

/// Deepest `children` nesting walked. docling's own JSON never goes past a
/// few dozen (sections in sections, lists in list items).
const MAX_DEPTH: usize = 256;

fn walk(reference: &Value, root: &Value, level: u8, doc: &mut DoclingDocument, g: &mut WalkGuard) {
    let key = reference["$ref"].as_str().unwrap_or("");
    if g.depth >= MAX_DEPTH || !g.seen.insert(key.to_string()) {
        return;
    }
    g.depth += 1;
    walk_inner(reference, root, level, doc, g);
    g.depth -= 1;
}

fn walk_inner(
    reference: &Value,
    root: &Value,
    level: u8,
    doc: &mut DoclingDocument,
    g: &mut WalkGuard,
) {
    let Some(item) = resolve(reference, root) else {
        return;
    };
    if item["content_layer"].as_str() == Some("furniture") {
        return;
    }
    let kind = ref_kind(reference);
    if kind.starts_with("#/texts/") {
        text_item(item, root, level, doc, g);
        // docling nests a section's content under its heading (and a list
        // item's sub-list under the item), so a text item's `children` are
        // body content too — without this walk a document collapses to its
        // first heading. Tables and pictures are not recursed: their children
        // are rich-cell / caption items already rendered with the parent.
        if let Some(children) = item["children"].as_array() {
            for c in children {
                walk(c, root, level, doc, g);
            }
        }
    } else if kind.starts_with("#/groups/") {
        group_item(item, root, level, doc, g);
    } else if kind.starts_with("#/tables/") {
        with_prov(item, doc, g, |doc, _| table_item(item, root, doc));
    } else if kind.starts_with("#/pictures/") {
        with_prov(item, doc, g, |doc, _| picture_item(item, root, doc));
    }
}

/// Run `push` for an item and attach the item's provenance (#899) to the
/// first node it pushed: the page marker ahead of it, the exact box around
/// it — or, for a list item, in its `location` field (`list_location`).
fn with_prov(
    item: &Value,
    doc: &mut DoclingDocument,
    g: &mut WalkGuard,
    push: impl FnOnce(&mut DoclingDocument, Option<[u16; 4]>),
) {
    let Some(prov) = prov_of(item) else {
        push(doc, None);
        return;
    };
    let list_item = item["label"].as_str() == Some("list_item");
    let on_page = enter_page(prov.page_no, list_item, doc, g);
    let start = doc.nodes.len();
    let grid = on_page.then(|| grid_location(&prov, g)).flatten();
    push(doc, grid);
    if doc.nodes.len() > start && !matches!(doc.nodes[start], Node::ListItem { .. }) {
        wrap_prov(doc, start, &prov, grid);
    }
}

fn text_item(item: &Value, root: &Value, level: u8, doc: &mut DoclingDocument, g: &mut WalkGuard) {
    with_prov(item, doc, g, |doc, location| {
        text_node(item, root, level, doc, location)
    });
}

fn text_node(
    item: &Value,
    root: &Value,
    level: u8,
    doc: &mut DoclingDocument,
    location: Option<[u16; 4]>,
) {
    let label = item["label"].as_str().unwrap_or("text");
    // docling does not serialize empty text items (an undecoded formula is the
    // one exception — it becomes a placeholder comment).
    if item["text"].as_str().unwrap_or("").is_empty() {
        if label == "formula" {
            doc.push(Node::Paragraph {
                text: "<!-- formula-not-decoded -->".into(),
            });
        }
        return;
    }
    let text = formatted_text(item);
    // Code and formulas are the two items docling serializes *unescaped*
    // (`escape_html = False`, `escape_underscores = False`): a SQL body keeps
    // its `VERIFY_GROUP_FOR_USER`, not `VERIFY\_GROUP\_FOR\_USER`.
    let raw = || item["text"].as_str().unwrap_or("").to_string();
    match label {
        "title" => doc.push(Node::Heading { level: 1, text }),
        "section_header" => {
            let lvl = item["level"].as_u64().unwrap_or(1) as u8;
            doc.push(Node::Heading {
                level: lvl + 1,
                text,
            });
        }
        "code" => {
            doc.push(Node::Code {
                language: item["code_language"]
                    .as_str()
                    .filter(|s| !s.is_empty() && *s != "unknown")
                    .map(String::from),
                text: raw(),
                orig: None,
                pretty: None,
            });
            // A `CodeItem` is docling's only *floating* text item, and the
            // text serializer appends a floating item's captions **after** its
            // own text — the opposite of a picture or a table, which lead with
            // theirs (`Listing 1: …` under the fence, not above it).
            if let Some(cap) = caption_of(item, root) {
                doc.push(Node::Paragraph { text: cap });
            }
        }
        "list_item" => doc.push(Node::ListItem {
            ordered: item["enumerated"].as_bool().unwrap_or(false),
            number: 1,
            first_in_list: true,
            text,
            level,
            marker: None,
            location,
            dclx: None,
            href: None,
            layer: None,
        }),
        // docling prefixes the task-list marker to the text of a checkbox item.
        "checkbox_selected" => doc.push(Node::CheckboxItem {
            checked: true,
            text,
        }),
        "checkbox_unselected" => doc.push(Node::CheckboxItem {
            checked: false,
            text,
        }),
        // A decoded formula renders as `$$…$$`, also unescaped.
        "formula" => doc.push(Node::Formula {
            latex: raw(),
            orig: item["orig"].as_str().unwrap_or("").to_string(),
            location: None,
        }),
        // A caption some table/picture/code claims renders with that element;
        // one nobody claims is an ordinary body item and renders where it sits
        // (docling's serializer only skips the refs a floating item consumed).
        "caption" => {
            if !caption_is_claimed(item, root) {
                doc.push(Node::Caption { text, href: None });
            }
        }
        _ => doc.push(Node::Paragraph { text }), // text, paragraph, footnote, …
    }
}

fn group_item(item: &Value, root: &Value, level: u8, doc: &mut DoclingDocument, g: &mut WalkGuard) {
    let label = item["label"].as_str().unwrap_or("unspecified");
    let empty = Vec::new();
    let children = item["children"].as_array().unwrap_or(&empty);
    match label {
        "list" | "ordered_list" => list_group(children, root, level, doc, g),
        "inline" => {
            // An inline group is one line: serialize each child and join with " ".
            let joined = children
                .iter()
                .map(|c| text_of(c, root))
                .collect::<Vec<_>>()
                .join(" ");
            if !joined.is_empty() {
                doc.push(Node::Paragraph { text: joined });
            }
        }
        // section / chapter / unspecified / sheet / comment_section → transparent
        _ => {
            for c in children {
                walk(c, root, level, doc, g);
            }
        }
    }
}

/// How docling-core's Markdown serializer renders a list item's marker.
///
/// * A marker it already considers valid Markdown — a bullet, or `12.` — is
///   printed **verbatim** and nothing is computed. This is what carries a
///   split reference list's real numbering (`18.` on the group that continues
///   over a page break) instead of restarting at 1.
/// * Any *other* non-empty marker (`a.`, `[7]`, `(3)`) forces a **bullet**:
///   the computed marker is a number only when the item carries no marker at
///   all (`… and (mode != AUTO or not item.marker)`). The original marker is
///   then kept after it when it holds a letter or digit, so docling renders
///   `- (1) Human Annotation`. In this node model it rides in the item's text,
///   which is where it lands on the rendered line either way; a marker with
///   no alphanumerics (a stray bullet glyph) is dropped, as upstream drops it.
/// * No marker at all leaves the group to decide: `{position}.` when its first
///   child is an enumerated item, `-` otherwise.
enum Marker {
    /// Print verbatim: `Some(n)` numbers the item, `None` bullets it.
    Verbatim(Option<u64>),
    /// Bullet, with this text (if any) kept in front of the item's own.
    Bullet(Option<String>),
    /// Nothing of its own — the group's kind and the item's position decide.
    FromGroup,
}

fn marker_of(item: &Value) -> Marker {
    let raw = item["marker"].as_str().unwrap_or("");
    if raw.is_empty() {
        return Marker::FromGroup;
    }
    if matches!(raw, "-" | "*" | "+") {
        return Marker::Verbatim(None);
    }
    if let Some(digits) = raw.strip_suffix('.') {
        if !digits.is_empty() && digits.bytes().all(|b| b.is_ascii_digit()) {
            return Marker::Verbatim(digits.parse().ok());
        }
    }
    Marker::Bullet(
        raw.chars()
            .any(|c| c.is_ascii_alphanumeric())
            .then(|| raw.to_string()),
    )
}

/// Emit a list group's items, recursing into nested lists at the next level.
fn list_group(
    children: &[Value],
    root: &Value,
    level: u8,
    doc: &mut DoclingDocument,
    g: &mut WalkGuard,
) {
    // docling-core's `first_item_is_enumerated`: the group renders computed
    // markers as numbers only when its *first child* is an enumerated item.
    let enumerated_group = children
        .first()
        .and_then(|c| resolve(c, root))
        .is_some_and(|it| {
            it["label"].as_str() == Some("list_item") && it["enumerated"].as_bool() == Some(true)
        });
    let mut first = true;
    for (pos, c) in children.iter().enumerate() {
        let kind = ref_kind(c);
        if kind.starts_with("#/groups/") {
            // A bare nested list (no enclosing item). It still occupies a
            // position in this group, which is why a later item can be
            // numbered past it.
            walk(c, root, level + 1, doc, g);
            continue;
        }
        let Some(item) = resolve(c, root) else {
            continue;
        };
        if item["label"].as_str() != Some("list_item") {
            continue;
        }
        let position = pos as u64 + 1;
        let body = formatted_text(item);
        let (ordered, number, text) = match marker_of(item) {
            Marker::Verbatim(Some(n)) => (true, n, body),
            Marker::Verbatim(None) => (false, position, body),
            Marker::Bullet(kept) => {
                let text = match kept {
                    Some(m) if !body.is_empty() => format!("{m} {body}"),
                    Some(m) => m,
                    None => body,
                };
                (false, position, text)
            }
            Marker::FromGroup => (enumerated_group, position, body),
        };
        let location = prov_of(item)
            .filter(|p| enter_page(p.page_no, true, doc, g))
            .and_then(|p| grid_location(&p, g));
        doc.push(Node::ListItem {
            ordered,
            number,
            first_in_list: first,
            text,
            level,
            marker: None,
            location,
            dclx: None,
            href: None,
            layer: None,
        });
        first = false;
        if let Some(sub) = item["children"].as_array() {
            for s in sub {
                walk(s, root, level + 1, doc, g);
            }
        }
    }
}

fn table_item(item: &Value, root: &Value, doc: &mut DoclingDocument) {
    let mut rows = Vec::new();
    // Per-cell flags carried through so DocTags export and the chunker's
    // triplet serialization see docling's header/span structure, not just the
    // flattened text grid.
    let mut structure = docling_core::TableStructure::default();
    if let Some(grid) = item["data"]["grid"].as_array() {
        for (r, row) in grid.iter().enumerate() {
            let Some(cells) = row.as_array() else {
                continue;
            };
            rows.push(
                cells
                    .iter()
                    .map(|cell| cell["text"].as_str().unwrap_or("").to_string())
                    .collect::<Vec<_>>(),
            );
            // A spanning cell is repeated at each grid position it covers; its
            // start offsets mark the anchor, everything past it a continuation.
            let past = |cell: &Value, key: &str, idx: usize| {
                cell[key].as_u64().is_some_and(|v| (v as usize) < idx)
            };
            structure
                .col_header
                .push(flags(cells, |c| c["column_header"].as_bool() == Some(true)));
            structure
                .row_header
                .push(flags(cells, |c| c["row_header"].as_bool() == Some(true)));
            structure.col_continuation.push(flags(cells.as_slice(), {
                let mut col = 0;
                move |c| {
                    let cont = past(c, "start_col_offset_idx", col);
                    col += 1;
                    cont
                }
            }));
            structure
                .row_continuation
                .push(flags(cells, |c| past(c, "start_row_offset_idx", r)));
        }
    }
    if !rows.is_empty() {
        let has_structure = structure.col_header.iter().flatten().any(|&b| b)
            || structure.row_header.iter().flatten().any(|&b| b)
            || structure.col_continuation.iter().flatten().any(|&b| b)
            || structure.row_continuation.iter().flatten().any(|&b| b);
        doc.push(Node::Table(Table {
            rows,
            location: None,
            structure: has_structure.then_some(structure),
            cell_blocks: None,
            cells: None,
            caption: caption_of(item, root),
            caption_parent: caption_parent_of(item, root),
            caption_location: None,
        }));
    }
}

/// Map each grid cell to a flag.
fn flags(cells: &[Value], f: impl FnMut(&Value) -> bool) -> Vec<bool> {
    cells.iter().map(f).collect()
}

fn picture_item(item: &Value, root: &Value, doc: &mut DoclingDocument) {
    doc.push(Node::Picture {
        caption: caption_of(item, root),
        caption_parent: caption_parent_of(item, root),
        caption_location: None,
        caption_href: None,
        image: picture_image(item),
        classification: picture_classification(item),
        description: picture_description(item),
    });
}

/// The picture's description (#975): docling's `meta.description`
/// (`DescriptionMetaField`: `text`, `created_by`), else the deprecated
/// `PictureDescriptionData` annotation (`kind: "description"`, `text`,
/// `provenance`) older documents carry alone. A description a picture-
/// description or picture-OCR pass computed is part of the document, and a
/// docling-JSON fed back in — to chunk it, to re-export it — keeps it; it
/// used to come back as `None`, so a document whose only text was its
/// pictures' descriptions chunked to nothing.
fn picture_description(item: &Value) -> Option<docling_core::PictureDescription> {
    let text_of = |v: &Value| {
        v["text"]
            .as_str()
            .filter(|t| !t.trim().is_empty())
            .map(str::to_string)
    };
    let meta = &item["meta"]["description"];
    if let Some(text) = text_of(meta) {
        return Some(docling_core::PictureDescription {
            text,
            provenance: meta["created_by"].as_str().unwrap_or_default().to_string(),
        });
    }
    item["annotations"]
        .as_array()?
        .iter()
        .find(|a| a["kind"].as_str() == Some("description"))
        .and_then(|a| {
            Some(docling_core::PictureDescription {
                text: text_of(a)?,
                provenance: a["provenance"].as_str().unwrap_or_default().to_string(),
            })
        })
}

/// The classifier's predictions (#975): `meta.classification.predictions`,
/// else the deprecated `classification` annotation's `predicted_classes`
/// (`class_name`, `confidence` each).
fn picture_classification(item: &Value) -> Option<Vec<docling_core::PictureClass>> {
    let preds = item["meta"]["classification"]["predictions"]
        .as_array()
        .or_else(|| {
            item["annotations"]
                .as_array()?
                .iter()
                .find(|a| a["kind"].as_str() == Some("classification"))?["predicted_classes"]
                .as_array()
        })?;
    let classes: Vec<_> = preds
        .iter()
        .filter_map(|p| {
            Some(docling_core::PictureClass {
                class_name: p["class_name"].as_str()?.to_string(),
                confidence: p["confidence"].as_f64().unwrap_or(0.0) as f32,
            })
        })
        .collect();
    (!classes.is_empty()).then_some(classes)
}

/// Where the item's first caption hangs in the source tree, so a re-export
/// keeps it there (#390): on the item (a PDF layout caption), on the item's
/// own parent — ahead of the item or behind it, by the parent's `children`
/// order — or, anywhere else, docling's `add_text` default, the body.
fn caption_parent_of(item: &Value, root: &Value) -> CaptionParent {
    let cap = &item["captions"][0];
    let Some(cap_ref) = cap["$ref"].as_str() else {
        return CaptionParent::Body;
    };
    let cap_parent = resolve(cap, root)
        .and_then(|c| c["parent"]["$ref"].as_str())
        .unwrap_or("#/body");
    let self_ref = item["self_ref"].as_str().unwrap_or("");
    let item_parent = &item["parent"];
    if cap_parent == self_ref {
        CaptionParent::Item
    } else if cap_parent == ref_kind(item_parent) && cap_parent != "#/body" {
        let pos = |r: &str| {
            resolve(item_parent, root)?
                .get("children")?
                .as_array()?
                .iter()
                .position(|c| c["$ref"] == r)
        };
        match (pos(cap_ref), pos(self_ref)) {
            (Some(c), Some(i)) if c > i => CaptionParent::ContainerAfter,
            _ => CaptionParent::Container,
        }
    } else {
        CaptionParent::Body
    }
}

/// The picture's embedded image — docling's `ImageRef` with a `data:` URI
/// (`{"mimetype", "dpi", "size", "uri"}`), which is how docling and our own
/// JSON export carry a picture's pixels. Without it every picture read back
/// from JSON was an empty placeholder, and `--images embedded`/`referenced`
/// silently had nothing to embed or write (#403). The dimensions come from the
/// bytes themselves (PNG IHDR / JPEG SOF), the declared `size` as a fallback.
fn picture_image(item: &Value) -> Option<PictureImage> {
    let image = item.get("image")?;
    let uri = image["uri"].as_str()?;
    let (mimetype, data) = parse_data_uri(uri)?;
    let mimetype = image["mimetype"]
        .as_str()
        .filter(|m| !m.is_empty())
        .unwrap_or(&mimetype)
        .to_string();
    let declared = |k: &str| image["size"][k].as_f64().unwrap_or(0.0).round().max(0.0) as u32;
    let (width, height) = crate::backend::rtf::image_size(&mimetype, &data)
        .unwrap_or((declared("width"), declared("height")));
    // The declared dpi round-trips (a PDF crop's 144 stays 144, #519).
    let dpi = image["dpi"]
        .as_u64()
        .and_then(|d| u32::try_from(d).ok())
        .filter(|&d| d > 0)
        .unwrap_or(PictureImage::DEFAULT_DPI);
    Some(PictureImage {
        mimetype,
        width,
        height,
        data,
        dpi,
        display_size: None,
    })
}

/// Split a `data:<mimetype>;base64,<payload>` URI into its media type and
/// decoded bytes. Anything else — a plain path, an `http(s)` or `file` URL —
/// is `None`: [`DoclingJsonBackend::convert`] inlines the local files it can
/// read beforehand, so only a genuinely unresolvable reference reaches here.
fn parse_data_uri(uri: &str) -> Option<(String, Vec<u8>)> {
    let rest = uri.strip_prefix("data:")?;
    let (meta, payload) = rest.split_once(',')?;
    let (mimetype, encoding) = meta.split_once(';').unwrap_or((meta, ""));
    if !encoding.eq_ignore_ascii_case("base64") {
        return None;
    }
    let data = docling_core::base64::decode(payload.trim())?;
    Some((mimetype.to_string(), data))
}

/// Turn each picture's *referenced* image (`uri` a filesystem path or `file:`
/// URL, as `--images referenced` writes them) into a `data:` URI, reading the
/// file relative to the JSON's own directory — docling's `ImageRef.pil_image`
/// opens such a URI from disk the same way. A file that cannot be read keeps
/// its `uri`, so the picture degrades to a placeholder rather than failing.
fn inline_referenced_images(root: &mut Value, base: Option<&std::path::Path>) {
    let Some(pictures) = root.get_mut("pictures").and_then(Value::as_array_mut) else {
        return;
    };
    for pic in pictures {
        let Some(image) = pic.get_mut("image").filter(|i| i.is_object()) else {
            continue;
        };
        let Some(uri) = image["uri"].as_str().map(str::to_string) else {
            continue;
        };
        if uri.starts_with("data:") || (uri.contains("://") && !uri.starts_with("file://")) {
            continue;
        }
        let path = std::path::PathBuf::from(uri.strip_prefix("file://").unwrap_or(&uri));
        let path = match base {
            Some(dir) if path.is_relative() => dir.join(path),
            _ => path,
        };
        let Ok(data) = std::fs::read(&path) else {
            continue;
        };
        let mimetype = image["mimetype"]
            .as_str()
            .filter(|m| !m.is_empty())
            .map(str::to_string)
            .unwrap_or_else(|| mime_of_path(&path));
        image["uri"] = Value::String(format!(
            "data:{mimetype};base64,{}",
            docling_core::base64::encode(&data)
        ));
    }
}

/// A media type from a file extension, for a referenced image whose `ImageRef`
/// carries none.
fn mime_of_path(path: &std::path::Path) -> String {
    match path
        .extension()
        .and_then(|e| e.to_str())
        .map(|e| e.to_ascii_lowercase())
        .as_deref()
    {
        Some("jpg") | Some("jpeg") => "image/jpeg",
        Some("gif") => "image/gif",
        Some("bmp") => "image/bmp",
        Some("webp") => "image/webp",
        Some("tif") | Some("tiff") => "image/tiff",
        _ => "image/png",
    }
    .to_string()
}

/// Whether a floating item lists this caption in its own `captions` — tables,
/// pictures and code items are the three that can. Scanning per caption keeps
/// the walk's signature unchanged; a document has few of either.
fn caption_is_claimed(item: &Value, root: &Value) -> bool {
    let Some(me) = item["self_ref"].as_str() else {
        return false;
    };
    ["tables", "pictures", "texts"].iter().any(|bucket| {
        root[bucket].as_array().is_some_and(|items| {
            items.iter().any(|it| {
                it["captions"]
                    .as_array()
                    .is_some_and(|caps| caps.iter().any(|c| c["$ref"].as_str() == Some(me)))
            })
        })
    })
}

/// An item's `captions` (refs into `texts`), joined as docling-core joins them
/// (`caption_delim`, a space). It belongs *on* the table or picture, not after
/// it: every serializer renders a caption before its element, and emitting the
/// caption items as trailing paragraphs put them on the wrong side (#384).
fn caption_of(item: &Value, root: &Value) -> Option<String> {
    let caps = item["captions"].as_array()?;
    let joined = caps
        .iter()
        .map(|c| text_of(c, root))
        .collect::<Vec<_>>()
        .join(" ");
    (!joined.trim().is_empty()).then_some(joined)
}

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

    fn md(json: &str) -> String {
        DoclingJsonBackend
            .convert(&SourceDocument::from_bytes(
                "t.json",
                InputFormat::JsonDocling,
                json.as_bytes().to_vec(),
            ))
            .unwrap()
            .export_to_markdown()
    }

    /// A `children` reference back to an ancestor, a chain of thousands of
    /// nested groups, and an item listed under many parents used to recurse
    /// without bound (or multiply the output). Each reference is walked once
    /// and nesting stops at `MAX_DEPTH`, so the documents convert and the
    /// cyclic one reads like its acyclic twin.
    #[test]
    fn cycles_and_deep_chains_convert() {
        let doc = |groups_children: &str| {
            format!(
                r##"{{"schema_name":"DoclingDocument","version":"1.7.0","name":"c",
                "body":{{"self_ref":"#/body","children":[{{"$ref":"#/groups/0"}}],"name":"_root_","label":"unspecified"}},
                "groups":[{{"self_ref":"#/groups/0","parent":{{"$ref":"#/body"}},"children":[{groups_children}],"name":"g","label":"section"}}],
                "texts":[{{"self_ref":"#/texts/0","parent":{{"$ref":"#/groups/0"}},"children":[],"label":"text","text":"hello","prov":[]}}]}}"##
            )
        };
        let plain = md(&doc(r##"{"$ref":"#/texts/0"}"##));
        assert!(plain.contains("hello"));
        let cyclic = md(&doc(
            r##"{"$ref":"#/texts/0"},{"$ref":"#/groups/0"},{"$ref":"#/body"}"##,
        ));
        assert_eq!(cyclic, plain);
        // 20 000 groups, each the only child of the previous one.
        let n = 20_000;
        let groups: Vec<String> = (0..n)
            .map(|i| {
                let child = if i + 1 < n { format!(r##"{{"$ref":"#/groups/{}"}}"##, i + 1) } else { r##"{"$ref":"#/texts/0"}"##.to_string() };
                let parent = if i == 0 { "#/body".to_string() } else { format!("#/groups/{}", i - 1) };
                format!(r##"{{"self_ref":"#/groups/{i}","parent":{{"$ref":"{parent}"}},"children":[{child}],"name":"g","label":"section"}}"##)
            })
            .collect();
        let deep = format!(
            r##"{{"schema_name":"DoclingDocument","version":"1.7.0","name":"d",
            "body":{{"self_ref":"#/body","children":[{{"$ref":"#/groups/0"}}],"name":"_root_","label":"unspecified"}},
            "groups":[{}],
            "texts":[{{"self_ref":"#/texts/0","parent":{{"$ref":"#/groups/19999"}},"children":[],"label":"text","text":"bottom","prov":[]}}]}}"##,
            groups.join(",")
        );
        // Converts (the text sits past the depth cap and is dropped).
        let _ = md(&deep);
    }

    /// #861 (docling-core 2.103.1's `validate_tree`): an item reachable
    /// through two references is walked once — 30 levels of a group whose
    /// two children are the same next group would be 2^30 walks without the
    /// guard — on every export; the leaf renders once.
    #[test]
    fn dag_shaped_tree_converts_once_and_in_linear_time() {
        let n = 30;
        let groups: Vec<String> = (0..n)
            .map(|i| {
                let child = if i + 1 < n {
                    format!(r##"{{"$ref":"#/groups/{}"}}"##, i + 1)
                } else {
                    r##"{"$ref":"#/texts/0"}"##.to_string()
                };
                let parent = if i == 0 { "#/body".to_string() } else { format!("#/groups/{}", i - 1) };
                format!(
                    r##"{{"self_ref":"#/groups/{i}","parent":{{"$ref":"{parent}"}},"children":[{child},{child}],"name":"g","label":"section"}}"##
                )
            })
            .collect();
        let dag = format!(
            r##"{{"schema_name":"DoclingDocument","version":"1.7.0","name":"d",
            "body":{{"self_ref":"#/body","children":[{{"$ref":"#/groups/0"}},{{"$ref":"#/groups/0"}}],"name":"_root_","label":"unspecified"}},
            "groups":[{}],
            "texts":[{{"self_ref":"#/texts/0","parent":{{"$ref":"#/groups/29"}},"children":[],"label":"text","text":"leaf","prov":[]}}]}}"##,
            groups.join(",")
        );
        let doc = DoclingJsonBackend
            .convert(&SourceDocument::from_bytes(
                "dag.json",
                InputFormat::JsonDocling,
                dag.into_bytes(),
            ))
            .unwrap();
        assert_eq!(doc.export_to_markdown().matches("leaf").count(), 1);
        // The JSON writes the text as `text` and `orig`: one item, two hits.
        assert_eq!(doc.export_to_json().matches("leaf").count(), 2);
        assert_eq!(doc.export_to_html().matches("leaf").count(), 1);
        assert_eq!(doc.export_to_latex().matches("leaf").count(), 1);
        assert_eq!(doc.export_to_doclang().matches("leaf").count(), 1);
        let pandoc = doc.export_to_pandoc_json();
        assert_eq!(pandoc.matches("leaf").count(), 1);
        let chunks = docling_core::chunker::HierarchicalChunker.chunk(&doc);
        assert_eq!(chunks.len(), 1, "{chunks:?}");
    }

    /// #861 (docling-core 2.103.1's `max_table_cells` /
    /// `max_total_table_cells`): a table declaring more cells than the
    /// budget — around one real cell — is refused before any grid is
    /// built, with the typed limit on the error's source chain; so is a
    /// document whose tables together pass the total.
    #[test]
    fn oversized_tables_are_rejected_with_a_typed_limit() {
        let doc = |tables: &str| {
            format!(
                r##"{{"schema_name":"DoclingDocument","version":"1.7.0","name":"t",
                "body":{{"self_ref":"#/body","children":[{{"$ref":"#/tables/0"}}],"name":"_root_","label":"unspecified"}},
                "tables":[{tables}]}}"##
            )
        };
        let table = |i: usize, rows: usize, cols: usize| {
            format!(
                r##"{{"self_ref":"#/tables/{i}","parent":{{"$ref":"#/body"}},"children":[],"label":"table","prov":[],
                "data":{{"num_rows":{rows},"num_cols":{cols},"table_cells":[{{"text":"x","start_row_offset_idx":0,"end_row_offset_idx":1,"start_col_offset_idx":0,"end_col_offset_idx":1,"row_span":1,"col_span":1}}],
                "grid":[[{{"text":"x","start_row_offset_idx":0,"end_row_offset_idx":1,"start_col_offset_idx":0,"end_col_offset_idx":1,"row_span":1,"col_span":1}}]]}}}}"##
            )
        };
        let convert = |json: String| {
            DoclingJsonBackend.convert(&SourceDocument::from_bytes(
                "t.json",
                InputFormat::JsonDocling,
                json.into_bytes(),
            ))
        };
        let err = convert(doc(&table(0, 1_000_000, 1_000_000))).unwrap_err();
        let limit = err.limit().expect("typed limit");
        assert_eq!(limit.variable, "DOCLING_RS_MAX_TABLE_CELLS");
        assert_eq!(limit.actual, 1_000_000 * 1_000_000);
        assert!(
            err.to_string().contains("DOCLING_RS_MAX_TABLE_CELLS"),
            "{err}"
        );
        // Rows alone past the cap, with no columns at all.
        assert!(convert(doc(&table(0, 1_000_001, 0)))
            .unwrap_err()
            .limit()
            .is_some());
        // Six tables of 999 999 cells pass each on their own and break the
        // document's total — reachable or not, as upstream counts them.
        let six: Vec<String> = (0..6).map(|i| table(i, 999, 1001)).collect();
        let err = convert(doc(&six.join(","))).unwrap_err();
        assert_eq!(
            err.limit().unwrap().variable,
            "DOCLING_RS_MAX_TOTAL_TABLE_CELLS"
        );
        // Under the caps the one real cell converts.
        let md = convert(doc(&table(0, 1000, 1000)))
            .unwrap()
            .export_to_markdown();
        assert!(md.contains("| x"), "{md}");
    }

    fn item(self_ref: &str, parent: &str, label: &str, extra: &str) -> String {
        format!(
            r#"{{"self_ref":"{self_ref}","parent":{{"$ref":"{parent}"}},"children":[],"label":"{label}","prov":[]{extra}}}"#
        )
    }

    /// #390: where a caption hangs is read back from the source tree, so a
    /// re-export keeps a PDF caption on its picture, an office chart's on the
    /// sheet ahead of the picture, and a declarative backend's on the body.
    #[test]
    fn caption_parents_round_trip() {
        let cap = |i: usize, parent: &str| {
            item(
                &format!("#/texts/{i}"),
                parent,
                "caption",
                r#","orig":"c","text":"c""#,
            )
        };
        let pic = |i: usize, parent: &str, cap: usize| {
            item(
                &format!("#/pictures/{i}"),
                parent,
                "picture",
                &format!(r##","captions":[{{"$ref":"#/texts/{cap}"}}],"annotations":[]"##),
            )
        };
        let json = format!(
            r##"{{"schema_name":"DoclingDocument","version":"1.10.0","name":"t",
            "body":{{"self_ref":"#/body","children":[{{"$ref":"#/pictures/0"}},{{"$ref":"#/groups/0"}},{{"$ref":"#/texts/1"}},{{"$ref":"#/pictures/2"}}],"name":"_root_","label":"unspecified"}},
            "furniture":{{"self_ref":"#/furniture","children":[],"name":"_root_","label":"unspecified"}},
            "groups":[{{"self_ref":"#/groups/0","parent":{{"$ref":"#/body"}},"children":[{{"$ref":"#/texts/2"}},{{"$ref":"#/pictures/1"}}],"name":"sheet","label":"section"}}],
            "texts":[{},{},{}],
            "pictures":[{},{},{}],
            "tables":[],"key_value_items":[],"form_items":[],"pages":{{}}}}"##,
            cap(0, "#/pictures/0"),
            cap(1, "#/body"),
            cap(2, "#/groups/0"),
            pic(0, "#/body", 0),
            pic(1, "#/groups/0", 2),
            pic(2, "#/body", 1),
        );
        let doc = DoclingJsonBackend
            .convert(&SourceDocument::from_bytes(
                "t.json",
                InputFormat::JsonDocling,
                json.into_bytes(),
            ))
            .unwrap();
        let parents: Vec<CaptionParent> = doc
            .nodes
            .iter()
            .filter_map(|n| match n {
                Node::Picture { caption_parent, .. } => Some(*caption_parent),
                _ => None,
            })
            .collect();
        assert_eq!(
            parents,
            [
                CaptionParent::Item,
                CaptionParent::Container,
                CaptionParent::Body
            ]
        );
        let v: Value = serde_json::from_str(&doc.export_to_json()).unwrap();
        let parent = |i: usize| {
            v["texts"][i]["parent"]["$ref"]
                .as_str()
                .unwrap()
                .to_string()
        };
        assert_eq!(parent(0), "#/pictures/0");
        // A `section` group is replayed transparently (only lists come back
        // as groups), so the container caption lands beside its picture on
        // the body — ahead of it, as a chart's title is.
        assert_eq!(parent(1), "#/body");
        assert_eq!(parent(2), "#/body");
        let body: Vec<&str> = v["body"]["children"]
            .as_array()
            .unwrap()
            .iter()
            .map(|r| r["$ref"].as_str().unwrap())
            .collect();
        assert_eq!(
            body,
            [
                "#/pictures/0",
                "#/texts/1",
                "#/pictures/1",
                "#/texts/2",
                "#/pictures/2"
            ]
        );
    }

    /// #975: a picture's description and classification come back from
    /// `meta` (docling 2.x) or from the deprecated annotations alone (an
    /// older document), so a described document fed back in chunks and
    /// re-exports with its descriptions; a picture carrying neither keeps
    /// `None`.
    #[test]
    fn a_pictures_description_and_classification_survive_the_round_trip() {
        let json_with = |pictures: &str| {
            format!(
                r##"{{"name":"t","body":{{"children":[{{"$ref":"#/pictures/0"}}]}},
                  "texts":[],"groups":[],"tables":[],"pictures":[{pictures}]}}"##
            )
        };
        let picture_of = |doc: &DoclingDocument| {
            doc.nodes
                .iter()
                .find_map(|n| match n {
                    Node::Picture {
                        description,
                        classification,
                        ..
                    } => Some((description.clone(), classification.clone())),
                    _ => None,
                })
                .expect("a picture")
        };
        let convert = |pictures: &str| {
            DoclingJsonBackend
                .convert(&SourceDocument::from_bytes(
                    "t.json",
                    InputFormat::JsonDocling,
                    json_with(pictures).into_bytes(),
                ))
                .unwrap()
        };
        // docling 2.x: `meta` carries both, the annotations repeat them.
        let doc = convert(
            r##"{"self_ref":"#/pictures/0","label":"picture","captions":[],"children":[],
                 "meta":{"description":{"created_by":"mock-vision","text":"A cat on a mat."},
                         "classification":{"predictions":[{"confidence":0.9,"created_by":"DocumentPictureClassifier","class_name":"natural_image"}]}},
                 "annotations":[{"kind":"description","text":"A cat on a mat.","provenance":"mock-vision"}]}"##,
        );
        let (desc, classes) = picture_of(&doc);
        let desc = desc.expect("the description");
        assert_eq!(
            (desc.text.as_str(), desc.provenance.as_str()),
            ("A cat on a mat.", "mock-vision")
        );
        let classes = classes.expect("the classification");
        assert_eq!(classes[0].class_name, "natural_image");
        assert!((classes[0].confidence - 0.9).abs() < 1e-6);
        // The description is text the chunker sees: a document whose only
        // text is a picture description chunks to that description.
        let chunks = docling_core::chunker::HierarchicalChunker.chunk(&doc);
        assert!(
            chunks.iter().any(|c| c.text.contains("A cat on a mat.")),
            "{chunks:?}"
        );
        // …and the re-export carries it the way the input did.
        let out: serde_json::Value = serde_json::from_str(&doc.export_to_json()).unwrap();
        assert_eq!(
            out["pictures"][0]["meta"]["description"]["text"],
            "A cat on a mat."
        );
        assert_eq!(
            out["pictures"][0]["meta"]["description"]["created_by"],
            "mock-vision"
        );

        // An older document: the deprecated annotations alone.
        let doc = convert(
            r##"{"self_ref":"#/pictures/0","label":"picture","captions":[],"children":[],
                 "annotations":[{"kind":"classification","provenance":"DocumentPictureClassifier",
                                 "predicted_classes":[{"class_name":"logo","confidence":0.75}]},
                                {"kind":"description","text":"Old style.","provenance":"old-model"}]}"##,
        );
        let (desc, classes) = picture_of(&doc);
        assert_eq!(
            desc.map(|d| (d.text, d.provenance)),
            Some(("Old style.".into(), "old-model".into()))
        );
        assert_eq!(
            classes.map(|c| c[0].class_name.clone()).as_deref(),
            Some("logo")
        );

        // Neither: nothing invented.
        let doc = convert(
            r##"{"self_ref":"#/pictures/0","label":"picture","captions":[],"children":[],"annotations":[]}"##,
        );
        assert_eq!(picture_of(&doc), (None, None));
    }

    /// #403: a picture's `image` (docling's `ImageRef`, a `data:` URI) comes
    /// back as the node's image — so `--images embedded`/`referenced` have
    /// pixels to work with — and a *referenced* image (a path relative to the
    /// JSON file, as `--images referenced` writes) is read from disk.
    #[test]
    fn a_pictures_image_survives_the_round_trip() {
        const PNG: &str = "iVBORw0KGgoAAAANSUhEUgAAADwAAAAoCAIAAAAt2Q6oAAAASUlEQVR4nO3OQQ3AIAAAMUAI/qUgCw97HFnSKug8e4+/Wa8DX0hXpCvSFemKdEW6Il2RrkhXpCvSFemKdEW6Il2RrkhXpCvSlQtBOQFUzKPJpQAAAABJRU5ErkJggg==";
        let json_with = |uri: &str| {
            format!(
                r##"{{"name":"t","body":{{"children":[{{"$ref":"#/pictures/0"}}]}},
                  "texts":[],"groups":[],"tables":[],
                  "pictures":[{{"self_ref":"#/pictures/0","label":"picture","captions":[],"children":[],
                    "image":{{"mimetype":"image/png","dpi":72,"size":{{"width":60.0,"height":40.0}},"uri":"{uri}"}}}}]}}"##
            )
        };
        let image_of = |doc: &DoclingDocument| {
            doc.nodes
                .iter()
                .find_map(|n| match n {
                    Node::Picture { image, .. } => image.clone(),
                    _ => None,
                })
                .expect("the picture keeps its image")
        };
        let doc = DoclingJsonBackend
            .convert(&SourceDocument::from_bytes(
                "t.json",
                InputFormat::JsonDocling,
                json_with(&format!("data:image/png;base64,{PNG}")).into_bytes(),
            ))
            .unwrap();
        let img = image_of(&doc);
        assert_eq!(
            (img.mimetype.as_str(), img.width, img.height),
            ("image/png", 60, 40)
        );
        assert_eq!(img.data_uri(), format!("data:image/png;base64,{PNG}"));
        assert!(doc
            .export_to_markdown_with_images(docling_core::ImageMode::Embedded, "artifacts")
            .0
            .contains(&format!("![Image](data:image/png;base64,{PNG})")));

        // Referenced: the JSON sits next to its `artifacts/` directory.
        let dir = std::env::temp_dir().join(format!(
            "docling-rs-json-pic-{}-{}",
            std::process::id(),
            std::time::SystemTime::now()
                .duration_since(std::time::UNIX_EPOCH)
                .unwrap()
                .as_nanos()
        ));
        std::fs::create_dir_all(dir.join("artifacts")).unwrap();
        std::fs::write(
            dir.join("artifacts/image_000000.png"),
            docling_core::base64::decode(PNG).unwrap(),
        )
        .unwrap();
        std::fs::write(dir.join("t.json"), json_with("artifacts/image_000000.png")).unwrap();
        let doc = DoclingJsonBackend
            .convert(&SourceDocument::from_file(dir.join("t.json")).unwrap())
            .unwrap();
        let img = image_of(&doc);
        assert_eq!((img.width, img.height), (60, 40));
        assert_eq!(img.data_uri(), format!("data:image/png;base64,{PNG}"));
        // An unreadable reference degrades to a placeholder, never an error.
        std::fs::write(dir.join("t.json"), json_with("artifacts/missing.png")).unwrap();
        let doc = DoclingJsonBackend
            .convert(&SourceDocument::from_file(dir.join("t.json")).unwrap())
            .unwrap();
        assert!(doc
            .nodes
            .iter()
            .any(|n| matches!(n, Node::Picture { image: None, .. })));
        let _ = std::fs::remove_dir_all(&dir);
    }

    /// #384: a caption leads its picture and its table, and trails its *code*
    /// block — a `CodeItem` is docling's only floating text item, and the text
    /// serializer appends a floating item's captions after its own text.
    #[test]
    fn captions_lead_a_picture_and_a_table_but_trail_code() {
        let json = r##"{
          "name":"n","body":{"children":[{"$ref":"#/pictures/0"},{"$ref":"#/tables/0"},{"$ref":"#/texts/2"}]},
          "texts":[
            {"self_ref":"#/texts/0","label":"caption","text":"Figure 1: a duck","children":[]},
            {"self_ref":"#/texts/1","label":"caption","text":"Table 1: the counts","children":[]},
            {"self_ref":"#/texts/2","label":"code","text":"let x = 1;","children":[],
             "captions":[{"$ref":"#/texts/3"}]},
            {"self_ref":"#/texts/3","label":"caption","text":"Listing 1: a binding","children":[]}
          ],
          "groups":[],
          "tables":[{"self_ref":"#/tables/0","label":"table","captions":[{"$ref":"#/texts/1"}],
                     "data":{"grid":[[{"text":"a"}]]},"children":[]}],
          "pictures":[{"self_ref":"#/pictures/0","label":"picture","captions":[{"$ref":"#/texts/0"}],"children":[]}]
        }"##;
        assert_eq!(
            md(json),
            "Figure 1: a duck\n\n<!-- image -->\n\nTable 1: the counts\n\n| a   |\n|-----|\n\n```\nlet x = 1;\n```\n\nListing 1: a binding\n"
        );
    }

    /// A caption no floating item claims is an ordinary body item and renders
    /// where it sits; one that is claimed renders only with its element.
    #[test]
    fn an_unclaimed_caption_still_renders() {
        let json = r##"{
          "name":"n","body":{"children":[{"$ref":"#/texts/0"},{"$ref":"#/pictures/0"},{"$ref":"#/texts/1"}]},
          "texts":[
            {"self_ref":"#/texts/0","label":"caption","text":"Figure 6: nobody claims me","children":[]},
            {"self_ref":"#/texts/1","label":"caption","text":"Figure 7: claimed","children":[]}
          ],
          "groups":[],"tables":[],
          "pictures":[{"self_ref":"#/pictures/0","label":"picture","captions":[{"$ref":"#/texts/1"}],"children":[]}]
        }"##;
        assert_eq!(
            md(json),
            "Figure 6: nobody claims me\n\nFigure 7: claimed\n\n<!-- image -->\n"
        );
    }

    /// docling-core prints a marker it already considers valid Markdown
    /// verbatim, so a reference list that continues over a page break keeps its
    /// real numbering instead of restarting at 1.
    #[test]
    fn a_numeric_marker_is_printed_verbatim() {
        let json = r##"{
          "name":"n","body":{"children":[{"$ref":"#/groups/0"}]},
          "texts":[
            {"self_ref":"#/texts/0","label":"list_item","text":"Xue, W.","marker":"18.","enumerated":true,"children":[]},
            {"self_ref":"#/texts/1","label":"list_item","text":"Ye, J.","marker":"19.","enumerated":true,"children":[]}
          ],
          "groups":[{"self_ref":"#/groups/0","label":"list","name":"list",
                     "children":[{"$ref":"#/texts/0"},{"$ref":"#/texts/1"}]}],
          "tables":[],"pictures":[]
        }"##;
        assert_eq!(md(json), "18. Xue, W.\n19. Ye, J.\n");
    }

    /// Any *other* non-empty marker forces a bullet and is kept in front of the
    /// text when it holds a letter or digit (docling: `- (1) Human Annotation`),
    /// because a number is computed only for an item with no marker at all.
    #[test]
    fn a_non_markdown_marker_forces_a_bullet_and_is_kept() {
        let json = r##"{
          "name":"n","body":{"children":[{"$ref":"#/groups/0"}]},
          "texts":[
            {"self_ref":"#/texts/0","label":"list_item","text":"Human Annotation","marker":"(1)","enumerated":true,"children":[]},
            {"self_ref":"#/texts/1","label":"list_item","text":"Red - PDF cells","marker":"a.","enumerated":true,"children":[]},
            {"self_ref":"#/texts/2","label":"list_item","text":"a stray glyph","marker":"\u0084","enumerated":false,"children":[]}
          ],
          "groups":[{"self_ref":"#/groups/0","label":"list","name":"list",
                     "children":[{"$ref":"#/texts/0"},{"$ref":"#/texts/1"},{"$ref":"#/texts/2"}]}],
          "tables":[],"pictures":[]
        }"##;
        assert_eq!(
            md(json),
            "- (1) Human Annotation\n- a. Red - PDF cells\n- a stray glyph\n"
        );
    }

    /// With no marker at all the group decides: its first child being an
    /// enumerated item numbers every item by its *position among the children*,
    /// nested groups included — so an item after a sublist is numbered past it.
    #[test]
    fn an_unmarked_item_is_numbered_by_its_position_in_the_group() {
        let json = r##"{
          "name":"n","body":{"children":[{"$ref":"#/groups/0"}]},
          "texts":[
            {"self_ref":"#/texts/0","label":"list_item","text":"one","marker":"","enumerated":true,"children":[]},
            {"self_ref":"#/texts/1","label":"list_item","text":"nested","marker":"","enumerated":true,"children":[]},
            {"self_ref":"#/texts/2","label":"list_item","text":"after","marker":"","enumerated":true,"children":[]}
          ],
          "groups":[
            {"self_ref":"#/groups/0","label":"list","name":"list",
             "children":[{"$ref":"#/texts/0"},{"$ref":"#/groups/1"},{"$ref":"#/texts/2"}]},
            {"self_ref":"#/groups/1","label":"list","name":"list","children":[{"$ref":"#/texts/1"}]}
          ],
          "tables":[],"pictures":[]
        }"##;
        // The nested group takes position 2, so "after" is the third child.
        assert_eq!(md(json), "1. one\n    1. nested\n3. after\n");
    }

    /// A group whose first child is not an enumerated item bullets every item,
    /// whatever each one's own flag says.
    #[test]
    fn a_group_starting_on_a_bullet_stays_bulleted() {
        let json = r##"{
          "name":"n","body":{"children":[{"$ref":"#/groups/0"}]},
          "texts":[
            {"self_ref":"#/texts/0","label":"list_item","text":"bullet first","marker":"","enumerated":false,"children":[]},
            {"self_ref":"#/texts/1","label":"list_item","text":"still a bullet","marker":"","enumerated":true,"children":[]}
          ],
          "groups":[{"self_ref":"#/groups/0","label":"list","name":"list",
                     "children":[{"$ref":"#/texts/0"},{"$ref":"#/texts/1"}]}],
          "tables":[],"pictures":[]
        }"##;
        assert_eq!(md(json), "- bullet first\n- still a bullet\n");
    }

    /// Checkbox items carry docling's task-list marker, and code and formulas
    /// are the two items it serializes unescaped.
    #[test]
    fn checkboxes_render_and_code_is_not_escaped() {
        let json = r##"{
          "name":"n","body":{"children":[{"$ref":"#/texts/0"},{"$ref":"#/texts/1"},{"$ref":"#/texts/2"},{"$ref":"#/texts/3"}]},
          "texts":[
            {"self_ref":"#/texts/0","label":"checkbox_selected","text":"done","children":[]},
            {"self_ref":"#/texts/1","label":"checkbox_unselected","text":"todo","children":[]},
            {"self_ref":"#/texts/2","label":"code","text":"VERIFY_GROUP_FOR_USER ( SESSION_USER )","children":[]},
            {"self_ref":"#/texts/3","label":"formula","text":"a_1 + b_2","orig":"a_1 + b_2","children":[]}
          ],
          "groups":[],"tables":[],"pictures":[]
        }"##;
        assert_eq!(
            md(json),
            "- [x] done\n\n- [ ] todo\n\n```\nVERIFY_GROUP_FOR_USER ( SESSION_USER )\n```\n\n$$a_1 + b_2$$\n"
        );
    }

    /// docling ≥ 2.5x nests body content under its `section_header`; the
    /// nested items must be walked, not dropped with the heading's subtree.
    #[test]
    fn walks_children_nested_under_headings() {
        let json = r##"{
          "name": "n", "body": {"children": [{"$ref":"#/texts/0"}]},
          "texts": [
            {"self_ref":"#/texts/0","label":"section_header","level":1,"text":"Intro",
             "children":[{"$ref":"#/texts/1"},{"$ref":"#/texts/2"}]},
            {"self_ref":"#/texts/1","label":"text","text":"First para","children":[]},
            {"self_ref":"#/texts/2","label":"section_header","level":2,"text":"Sub",
             "children":[{"$ref":"#/texts/3"}]},
            {"self_ref":"#/texts/3","label":"text","text":"Deep para","children":[]}
          ],
          "groups": [], "tables": [], "pictures": []
        }"##;
        let doc = DoclingJsonBackend
            .convert(&SourceDocument::from_bytes(
                "t.json",
                InputFormat::JsonDocling,
                json.as_bytes().to_vec(),
            ))
            .unwrap();
        assert_eq!(
            doc.export_to_markdown(),
            "## Intro\n\nFirst para\n\n### Sub\n\nDeep para\n"
        );
    }

    #[test]
    fn walks_body_tree_with_formatting_and_lists() {
        let json = r##"{
          "schema_name": "DoclingDocument", "name": "t",
          "body": {"children": [{"$ref":"#/texts/0"},{"$ref":"#/texts/1"},{"$ref":"#/groups/0"}]},
          "texts": [
            {"self_ref":"#/texts/0","label":"title","text":"Doc"},
            {"self_ref":"#/texts/1","label":"section_header","level":1,"text":"Sec","hyperlink":"http://x"},
            {"self_ref":"#/texts/2","label":"list_item","text":"one","enumerated":false},
            {"self_ref":"#/texts/3","label":"list_item","text":"two","enumerated":false,
             "formatting":{"bold":true,"italic":false,"strikethrough":false}}
          ],
          "groups": [{"self_ref":"#/groups/0","label":"list",
                      "children":[{"$ref":"#/texts/2"},{"$ref":"#/texts/3"}]}],
          "tables": [], "pictures": []
        }"##;
        let src =
            SourceDocument::from_bytes("t", InputFormat::JsonDocling, json.as_bytes().to_vec());
        let md = DoclingJsonBackend
            .convert(&src)
            .unwrap()
            .export_to_markdown();
        assert!(
            md.starts_with("# Doc\n\n## [Sec](http://x)\n\n- one\n- **two**"),
            "got:\n{md}"
        );
    }
}