docling_core/tree.rs
1//! docling's **item tree**, for a backend that knows the exact shape upstream
2//! gives a document and wants the JSON export to reproduce it.
3//!
4//! [`DoclingDocument::nodes`](crate::DoclingDocument::nodes) is a flat,
5//! reading-order stream tuned for Markdown / DocLang / LaTeX; the JSON export
6//! rebuilds docling's parent/child structure from it with generic rules (runs
7//! of list items become list groups, a heading is a flat sibling of the text
8//! that follows it). Upstream's backends do not all agree on that structure:
9//! the HTML backend nests everything after a heading *under* the heading,
10//! splits a paragraph of mixed formatting into an `inline` group of one text
11//! item per formatting run, parents a rich table cell's content to a group
12//! under the table, keeps site chrome on the `furniture` layer… and numbers
13//! every item in the order it *creates* them. A backend that ports those
14//! rules call-for-call (HTML's `html_tree.rs`, DOCX's `docx_tree.rs`) records
15//! the result here — an arena of items in
16//! creation order, each with its parent and children — and the JSON export
17//! ([`DoclingDocument::export_to_json`](crate::DoclingDocument::export_to_json))
18//! serializes this tree instead of deriving one from the nodes. Every other
19//! serializer keeps reading the flat nodes, so their output is unaffected.
20
21use crate::{ContentLayer, FieldItem, PictureImage, Script, Table};
22
23/// docling-core's `Formatting`: the inline styles an item carries in JSON.
24#[derive(Debug, Clone, Copy, PartialEq, Eq, Default)]
25pub struct Formatting {
26 pub bold: bool,
27 pub italic: bool,
28 pub underline: bool,
29 pub strikethrough: bool,
30 pub script: Script,
31}
32
33/// A `list_item`'s docling fields.
34#[derive(Debug, Clone, PartialEq, Eq, Default)]
35pub struct ListMeta {
36 pub enumerated: bool,
37 /// docling's `marker` — the HTML backend writes `""` unless an ordered
38 /// list carries an explicit `start`, then `"{n}."`.
39 pub marker: String,
40}
41
42/// What an item in the tree is. Mirrors the docling-core item classes the
43/// JSON `texts` / `groups` / `tables` / `pictures` / `field_regions` buckets
44/// hold.
45#[derive(Debug, Clone, PartialEq)]
46pub enum TreeKind {
47 /// A `TextItem` / `TitleItem` / `SectionHeaderItem` / `ListItem`, told
48 /// apart by `label` (`text`, `title`, `section_header`, `list_item`,
49 /// `caption`, `checkbox_selected`, `checkbox_unselected`, …).
50 Text {
51 label: String,
52 text: String,
53 /// docling's `orig` when it differs from `text` (the heading text
54 /// before unicode cleanup, say); `None` = same as `text`.
55 orig: Option<String>,
56 formatting: Option<Formatting>,
57 hyperlink: Option<String>,
58 /// `section_header` only: docling's heading level.
59 level: Option<u8>,
60 /// `list_item` only.
61 list: Option<ListMeta>,
62 },
63 /// A `CodeItem`.
64 Code {
65 text: String,
66 orig: Option<String>,
67 /// The language hint (a highlighter class token such as `python`),
68 /// mapped onto docling's `CodeLanguageLabel` at export; `None` →
69 /// `unknown`.
70 language: Option<String>,
71 formatting: Option<Formatting>,
72 hyperlink: Option<String>,
73 },
74 /// A `GroupItem`: `label` is docling's `GroupLabel` value (`inline`,
75 /// `list`, `section`, `unspecified`, …), `name` its name (`group`, `list`,
76 /// `ordered list`, `header-2`, `rich_cell_group_1_0_3`, …).
77 Group { label: String, name: String },
78 /// A `TableItem`. `rich_cells` marks the cells docling serialized as a
79 /// `RichTableCell`: `(row, col)` grid anchor → the group item (a child of
80 /// the table) that holds the cell's content. `captions` are caption text
81 /// items in the tree.
82 Table {
83 table: Table,
84 rich_cells: Vec<(usize, usize, usize)>,
85 captions: Vec<usize>,
86 },
87 /// A `PictureItem`, its caption text items and optional payload.
88 /// `classification` is a `PictureClassificationLabel` value written as
89 /// the picture's `meta.classification` (an HTML `<stamp>` / `<signature>`).
90 Picture {
91 captions: Vec<usize>,
92 image: Option<PictureImage>,
93 classification: Option<String>,
94 /// The prediction's `confidence`, when the backend writes one: the
95 /// DocLang deserializer stamps `1.0`; the office and HTML backends
96 /// leave it out (`None`).
97 confidence: Option<f64>,
98 /// A native chart's data grid (docling's `meta.tabular_chart.chart_data`,
99 /// the series reconstructed as a `TableData`), for a DOCX chart drawing.
100 chart: Option<Table>,
101 /// The `ImageRef.dpi` docling writes for `image` when the backend
102 /// read one from the file (python-pptx's `Image.dpi`: PIL's `dpi`
103 /// info, rounded, 72 when absent or out of 1–2048); `None` → 72,
104 /// which is what upstream's other office backends pass.
105 dpi: Option<u32>,
106 },
107 /// A form key-value region (`field_regions` / `field_items`).
108 FieldRegion { items: Vec<FieldItem> },
109 /// A `KeyValueItem` (`key_value_items`): docling's `GraphData` of key and
110 /// value cells and their links, written verbatim.
111 KeyValueGraph {
112 cells: Vec<crate::GraphCell>,
113 links: Vec<crate::GraphLink>,
114 },
115}
116
117/// docling's `ProvenanceItem` for a tree item, written verbatim: the
118/// backend's own geometry in the page's units — a PPTX shape's EMU box,
119/// whose `pages` entry is the slide size in EMU — rather than the 0–511
120/// DocLang grid the flat [`Node::Located`](crate::Node::Located) carries
121/// (which cannot round-trip those integers).
122#[derive(Debug, Clone, PartialEq)]
123pub struct TreeProv {
124 /// 1-based page (slide) number.
125 pub page_no: usize,
126 /// `[l, t, r, b]`, exactly as docling computed them.
127 pub bbox: [f64; 4],
128 /// docling's `coord_origin` tag. The office backends tag their
129 /// top-left-based boxes `TOPLEFT` (the PPTX backend since docling#4294 —
130 /// it used to tag them `BOTTOMLEFT`, which read the tuple as `(l, b, r,
131 /// t)` and swapped the vertical edges); a backend that really works in a
132 /// bottom-left space sets this.
133 pub bottom_left: bool,
134 /// `[0, len(text)]` in characters for a text item, `[0, 0]` for a table
135 /// or picture.
136 pub charspan: [usize; 2],
137}
138
139/// docling's `TrackSource` — where in a time-based track (a WebVTT cue) a
140/// text item came from. Written as the item's `source: [{"kind": "track", …}]`.
141#[derive(Debug, Clone, PartialEq)]
142pub struct TreeTrack {
143 /// The cue's start offset in seconds (docling's `WebVTTTimestamp.seconds`:
144 /// `h*3600 + m*60 + s + millis/1000.0`, so the float is bit-identical).
145 pub start_time: f64,
146 /// The cue's end offset in seconds.
147 pub end_time: f64,
148 /// The cue identifier line, when the cue has one.
149 pub identifier: Option<String>,
150 /// The `<v …>` voice annotation the text sits in, when any.
151 pub voice: Option<String>,
152}
153
154/// One item of an [`ItemTree`].
155#[derive(Debug, Clone, PartialEq)]
156pub struct TreeItem {
157 /// The parent item's index; `None` = the document body.
158 pub parent: Option<usize>,
159 /// Child item indices, in docling's `children` order.
160 pub children: Vec<usize>,
161 /// The content layer; `None` = `body`.
162 pub layer: Option<ContentLayer>,
163 pub kind: TreeKind,
164 /// The item's `prov` entry, when the backend has page geometry for it
165 /// (`None` → `prov: []`, what the HTML and DOCX backends write).
166 pub prov: Option<TreeProv>,
167 /// docling's `DocItem.comments`: the `comment_section` groups (or note
168 /// text items) annotating this item, as item indices — written after
169 /// `prov` when non-empty.
170 pub comments: Vec<usize>,
171 /// docling's `DocItem.source`: the track segment a text item was taken
172 /// from (WebVTT cues) — written after `prov` when set.
173 pub source: Option<TreeTrack>,
174 /// Removed by [`ItemTree::delete`] (docling's `delete_items`): the slot
175 /// stays so every other index keeps its meaning, but the item is not
176 /// numbered or written.
177 pub deleted: bool,
178 /// Footnotes / endnotes referenced from inside this text item (#538): the
179 /// note call's position (chars into the item's text) and the note's text.
180 /// docling keeps notes as unlinked furniture `footnote` items, so the
181 /// JSON never shows this; the Pandoc AST writes each as a `Note` there.
182 pub notes: Vec<TreeNote>,
183 /// This furniture `footnote` item is the body of a note some text item
184 /// calls (it travels in that item's [`Self::notes`]); the Pandoc AST then
185 /// leaves it out as a standalone block.
186 pub note_body: bool,
187}
188
189/// A note call inside a text item ([`TreeItem::notes`]).
190#[derive(Debug, Clone, PartialEq, Eq)]
191pub struct TreeNote {
192 /// The call's position, in chars into the item's `text`.
193 pub offset: usize,
194 /// The note's plain text.
195 pub text: String,
196}
197
198impl ItemTree {
199 /// Where each note call of one paragraph lands (#538): `full_text` is the
200 /// paragraph's text, `offsets` the calls' positions in it (chars), and
201 /// the paragraph's items are those created from `first_new` on. Each text
202 /// item is a trimmed slice of the paragraph (a formatting run, a link, or
203 /// the whole heading / list item), matched in order; a call inside an
204 /// item lands there, one between items ends the earlier (or starts the
205 /// first). When no item matches, the call ends the paragraph's last text
206 /// item; with no text item at all it is `None`.
207 pub fn place_note_calls(
208 &self,
209 first_new: usize,
210 full_text: &str,
211 offsets: &[usize],
212 ) -> Vec<Option<(usize, usize)>> {
213 let texts: Vec<(usize, &str)> = (first_new..self.items.len())
214 .filter(|&i| !self.items[i].deleted)
215 .filter_map(|i| match &self.items[i].kind {
216 TreeKind::Text { text, .. } | TreeKind::Code { text, .. } => {
217 Some((i, text.as_str()))
218 }
219 _ => None,
220 })
221 .collect();
222 // Each found item's (id, start, end) in chars of `full_text`.
223 let mut spans: Vec<(usize, usize, usize)> = Vec::new();
224 let mut cursor = 0usize; // bytes
225 for &(item, text) in &texts {
226 if text.is_empty() {
227 continue;
228 }
229 if let Some(pos) = full_text[cursor..].find(text) {
230 let start_b = cursor + pos;
231 let start = full_text[..start_b].chars().count();
232 spans.push((item, start, start + text.chars().count()));
233 cursor = start_b + text.len();
234 }
235 }
236 offsets
237 .iter()
238 .map(|&offset| {
239 spans
240 .iter()
241 .find(|&&(_, start, end)| offset >= start && offset <= end)
242 .map(|&(item, start, _)| (item, offset - start))
243 .or_else(
244 || match spans.iter().rev().find(|&&(_, _, end)| end <= offset) {
245 Some(&(item, start, end)) => Some((item, end - start)),
246 None => spans.first().map(|&(item, _, _)| (item, 0)),
247 },
248 )
249 .or_else(|| {
250 texts
251 .last()
252 .map(|&(item, text)| (item, text.chars().count()))
253 })
254 })
255 .collect()
256 }
257}
258
259/// docling's item tree in creation order (see the [module docs](self)).
260#[derive(Debug, Clone, PartialEq, Default)]
261pub struct ItemTree {
262 /// Every item, indexed by creation order — which is how docling numbers
263 /// `#/texts/N`, `#/groups/N`, … within each bucket.
264 pub items: Vec<TreeItem>,
265 /// The body's `children`, as item indices.
266 pub body: Vec<usize>,
267}
268
269impl ItemTree {
270 /// Append an item under `parent` (`None` = body) on `layer`, registering
271 /// it as its parent's last child — docling's `add_*` calls do exactly that.
272 pub fn add(
273 &mut self,
274 parent: Option<usize>,
275 layer: Option<ContentLayer>,
276 kind: TreeKind,
277 ) -> usize {
278 let id = self.items.len();
279 self.items.push(TreeItem {
280 parent,
281 children: Vec::new(),
282 layer,
283 kind,
284 prov: None,
285 comments: Vec::new(),
286 source: None,
287 deleted: false,
288 notes: Vec::new(),
289 note_body: false,
290 });
291 match parent {
292 Some(p) => self.items[p].children.push(id),
293 None => self.body.push(id),
294 }
295 id
296 }
297
298 /// [`add`](Self::add) with the item's provenance — docling's
299 /// `add_text(…, prov=prov)`.
300 pub fn add_with_prov(
301 &mut self,
302 parent: Option<usize>,
303 layer: Option<ContentLayer>,
304 kind: TreeKind,
305 prov: TreeProv,
306 ) -> usize {
307 let id = self.add(parent, layer, kind);
308 self.items[id].prov = Some(prov);
309 id
310 }
311
312 /// Append every item of `other` after this tree's, renumbering its
313 /// indices (parents, children, comments, table/picture caption and
314 /// rich-cell refs) and adding its body children to this body — so a
315 /// backend can build independent fragments in parallel (one per PPTX
316 /// slide) and still hand the export one tree in creation order, exactly
317 /// as if it had been built sequentially.
318 pub fn append(&mut self, other: ItemTree) {
319 let off = self.items.len();
320 let shift = |i: usize| i + off;
321 for mut item in other.items {
322 item.parent = item.parent.map(shift);
323 for c in item.children.iter_mut().chain(item.comments.iter_mut()) {
324 *c = shift(*c);
325 }
326 match &mut item.kind {
327 TreeKind::Table {
328 rich_cells,
329 captions,
330 ..
331 } => {
332 for (_, _, g) in rich_cells.iter_mut() {
333 *g = shift(*g);
334 }
335 for c in captions.iter_mut() {
336 *c = shift(*c);
337 }
338 }
339 TreeKind::Picture { captions, .. } => {
340 for c in captions.iter_mut() {
341 *c = shift(*c);
342 }
343 }
344 _ => {}
345 }
346 self.items.push(item);
347 }
348 self.body.extend(other.body.into_iter().map(shift));
349 }
350
351 /// Move `id` under `new_parent`, dropping it from its current parent's
352 /// children and appending it to the new one's — docling's
353 /// `group_cell_elements` re-parenting of a rich cell's items.
354 pub fn reparent(&mut self, id: usize, new_parent: Option<usize>) {
355 let old = self.items[id].parent;
356 let siblings = match old {
357 Some(p) => &mut self.items[p].children,
358 None => &mut self.body,
359 };
360 siblings.retain(|&c| c != id);
361 self.items[id].parent = new_parent;
362 match new_parent {
363 Some(p) => self.items[p].children.push(id),
364 None => self.body.push(id),
365 }
366 }
367
368 /// Re-number the items in traversal order — a pre-order walk of the body
369 /// through every layer, which is how docling's
370 /// `DoclingDocument.concatenate` (and `_normalize_references`) re-creates
371 /// a document's items: a group created after the content it was later
372 /// wrapped around comes before that content afterwards. Items the walk
373 /// does not reach (deleted ones) are dropped. Returns each old id's new
374 /// id.
375 pub fn renumber_in_traversal_order(&mut self) -> Vec<Option<usize>> {
376 let mut order = Vec::with_capacity(self.items.len());
377 let mut stack: Vec<usize> = self.body.iter().rev().copied().collect();
378 while let Some(id) = stack.pop() {
379 if self.items[id].deleted {
380 continue;
381 }
382 order.push(id);
383 stack.extend(self.items[id].children.iter().rev());
384 }
385 let mut new_of: Vec<Option<usize>> = vec![None; self.items.len()];
386 for (new, &old) in order.iter().enumerate() {
387 new_of[old] = Some(new);
388 }
389 let remap = |ids: &mut Vec<usize>| {
390 *ids = ids.iter().filter_map(|&i| new_of[i]).collect();
391 };
392 let mut old_items: Vec<Option<TreeItem>> = std::mem::take(&mut self.items)
393 .into_iter()
394 .map(Some)
395 .collect();
396 for &old in &order {
397 let mut item = old_items[old].take().expect("each item visited once");
398 item.parent = item.parent.and_then(|p| new_of[p]);
399 remap(&mut item.children);
400 remap(&mut item.comments);
401 match &mut item.kind {
402 TreeKind::Table {
403 rich_cells,
404 captions,
405 ..
406 } => {
407 rich_cells.retain_mut(|(_, _, g)| match new_of[*g] {
408 Some(n) => {
409 *g = n;
410 true
411 }
412 None => false,
413 });
414 remap(captions);
415 }
416 TreeKind::Picture { captions, .. } => remap(captions),
417 _ => {}
418 }
419 self.items.push(item);
420 }
421 remap(&mut self.body);
422 new_of
423 }
424
425 /// Remove `id` from the tree — docling's `delete_items`, which the DOCX
426 /// backend uses to drop the empty text item a blank spacer paragraph left
427 /// between two items of a resumed list. The item leaves its parent's
428 /// children and is neither numbered nor written; its slot stays so the
429 /// indices held elsewhere stay valid.
430 pub fn delete(&mut self, id: usize) {
431 match self.items[id].parent {
432 Some(p) => self.items[p].children.retain(|&c| c != id),
433 None => self.body.retain(|&c| c != id),
434 }
435 self.items[id].deleted = true;
436 }
437
438 /// docling-core's `DoclingDocument.validate_misplaced_list_items` (a
439 /// model validator, so it runs whenever docling-core serializes or loads
440 /// a document): every `list_item` whose parent is not a `list` group is
441 /// re-homed into a new one. A pre-order walk of the body (groups
442 /// included) collects them; consecutive misplaced items directly on the
443 /// body share one group, any other misplaced item gets its own. Working
444 /// from the last run back, each run gets a `ListGroup` (name `group`)
445 /// inserted where its first item stood, the items are deleted and
446 /// re-added under the group as fresh items, so they move to the end of
447 /// the text numbering (#527: the DOCX backend leaves such items in rich
448 /// table cells). A no-op for a well-formed tree.
449 ///
450 /// One deliberate divergence (#586): docling-core deletes each item *with
451 /// its children* and re-adds it from its text alone, so a mixed-format
452 /// item — an empty `list_item` over an `inline` group of text runs —
453 /// comes back empty and its text is gone from the JSON (and from the
454 /// Markdown/HTML docling serializes after that validator ran; the JSON
455 /// docling saves *before* exporting still has it). Here the children
456 /// follow the item into the group: the structure docling-core requires,
457 /// the content the document had.
458 pub fn wrap_misplaced_list_items(&mut self) {
459 let is_list_item = |t: &Self, id: usize| matches!(&t.items[id].kind, TreeKind::Text { label, .. } if label == "list_item");
460 let in_list_group = |t: &Self, id: usize| {
461 t.items[id].parent.is_some_and(
462 |p| matches!(&t.items[p].kind, TreeKind::Group { label, .. } if label == "list"),
463 )
464 };
465 let mut runs: Vec<Vec<usize>> = Vec::new();
466 // `None` = the body itself, which the walk yields first.
467 let mut prev: Option<usize> = None;
468 let mut stack: Vec<usize> = self.body.iter().rev().copied().collect();
469 while let Some(id) = stack.pop() {
470 if self.items[id].deleted {
471 continue;
472 }
473 if is_list_item(self, id) && !in_list_group(self, id) {
474 let continues =
475 prev.is_some_and(|p| is_list_item(self, p) && self.items[p].parent.is_none());
476 match runs.last_mut() {
477 Some(run) if continues => run.push(id),
478 _ => runs.push(vec![id]),
479 }
480 }
481 prev = Some(id);
482 stack.extend(self.items[id].children.iter().rev().copied());
483 }
484 for run in runs.into_iter().rev() {
485 let parent = self.items[run[0]].parent;
486 let group = self.add(
487 parent,
488 None,
489 TreeKind::Group {
490 label: "list".into(),
491 name: "group".into(),
492 },
493 );
494 let siblings = match parent {
495 Some(p) => &mut self.items[p].children,
496 None => &mut self.body,
497 };
498 siblings.pop();
499 let at = siblings
500 .iter()
501 .position(|&c| c == run[0])
502 .unwrap_or(siblings.len());
503 siblings.insert(at, group);
504 for &li in &run {
505 // Not `delete_subtree`: the children move to the copy (#586).
506 self.delete(li);
507 }
508 // `add_list_item` keeps the text, marker, formatting, hyperlink
509 // and first provenance — not comments or a source. The children
510 // (the inline group of a mixed-format item) are carried over.
511 for &li in &run {
512 let children = std::mem::take(&mut self.items[li].children);
513 let copy = TreeItem {
514 parent: Some(group),
515 children: children.clone(),
516 comments: Vec::new(),
517 source: None,
518 deleted: false,
519 ..self.items[li].clone()
520 };
521 let id = self.items.len();
522 self.items.push(copy);
523 for c in children {
524 self.items[c].parent = Some(id);
525 }
526 self.items[group].children.push(id);
527 }
528 }
529 }
530
531 /// The last live text-bucket item (docling's `doc.texts[-1]`).
532 pub fn last_text(&self) -> Option<usize> {
533 self.items.iter().rposition(|it| {
534 !it.deleted && matches!(it.kind, TreeKind::Text { .. } | TreeKind::Code { .. })
535 })
536 }
537
538 /// How many items of a bucket precede `id` — its `#/{bucket}/N` index.
539 pub fn bucket_index(&self, id: usize) -> usize {
540 let same = |k: &TreeKind| {
541 std::mem::discriminant(k) == std::mem::discriminant(&self.items[id].kind)
542 || matches!(
543 (k, &self.items[id].kind),
544 (TreeKind::Text { .. }, TreeKind::Code { .. })
545 | (TreeKind::Code { .. }, TreeKind::Text { .. })
546 )
547 };
548 self.items[..id]
549 .iter()
550 .filter(|it| !it.deleted && same(&it.kind))
551 .count()
552 }
553
554 /// The number of tables created so far (docling's `len(doc.tables)`).
555 pub fn table_count(&self) -> usize {
556 self.items
557 .iter()
558 .filter(|it| !it.deleted && matches!(it.kind, TreeKind::Table { .. }))
559 .count()
560 }
561}
562
563#[cfg(test)]
564mod tests {
565 use super::*;
566
567 fn text(t: &str) -> TreeKind {
568 TreeKind::Text {
569 label: "text".into(),
570 text: t.into(),
571 orig: None,
572 formatting: None,
573 hyperlink: None,
574 level: None,
575 list: None,
576 }
577 }
578
579 /// `add` registers the item as its parent's (or the body's) last child;
580 /// `reparent` moves it — a rich cell's items leave the heading they were
581 /// created under for the table's group.
582 #[test]
583 fn add_and_reparent_keep_docling_children_order() {
584 let mut t = ItemTree::default();
585 let title = t.add(None, None, text("Title"));
586 let a = t.add(Some(title), None, text("a"));
587 let b = t.add(Some(title), None, text("b"));
588 let table = t.add(
589 Some(title),
590 None,
591 TreeKind::Table {
592 table: Table::default(),
593 rich_cells: Vec::new(),
594 captions: Vec::new(),
595 },
596 );
597 let group = t.add(
598 Some(table),
599 None,
600 TreeKind::Group {
601 label: "unspecified".into(),
602 name: "rich_cell_group_1_0_0".into(),
603 },
604 );
605 assert_eq!(t.body, vec![title]);
606 assert_eq!(t.items[title].children, vec![a, b, table]);
607 t.reparent(a, Some(group));
608 assert_eq!(t.items[title].children, vec![b, table]);
609 assert_eq!(t.items[group].children, vec![a]);
610 assert_eq!(t.items[a].parent, Some(group));
611 assert_eq!(t.table_count(), 1);
612 // Text and code share the `texts` bucket.
613 let code = t.add(
614 None,
615 None,
616 TreeKind::Code {
617 text: "x".into(),
618 orig: None,
619 language: None,
620 formatting: None,
621 hyperlink: None,
622 },
623 );
624 assert_eq!(t.bucket_index(code), 3, "title, a, b precede it in `texts`");
625 assert_eq!(t.bucket_index(group), 0);
626 assert_eq!(t.body, vec![title, code]);
627 }
628
629 /// `append` renumbers a fragment built on its own (a slide converted in
630 /// parallel) so the merged tree reads as if built in one pass: parents,
631 /// children, comment back-refs and caption refs all shift together.
632 #[test]
633 fn append_renumbers_a_fragment_into_creation_order() {
634 let mut whole = ItemTree::default();
635 let slide0 = whole.add(
636 None,
637 None,
638 TreeKind::Group {
639 label: "chapter".into(),
640 name: "slide-0".into(),
641 },
642 );
643 whole.add(Some(slide0), None, text("first"));
644
645 let mut frag = ItemTree::default();
646 let slide1 = frag.add(
647 None,
648 None,
649 TreeKind::Group {
650 label: "chapter".into(),
651 name: "slide-1".into(),
652 },
653 );
654 let cap = frag.add_with_prov(
655 Some(slide1),
656 None,
657 TreeKind::Text {
658 label: "caption".into(),
659 text: "Title".into(),
660 orig: None,
661 formatting: None,
662 hyperlink: None,
663 level: None,
664 list: None,
665 },
666 TreeProv {
667 page_no: 2,
668 bbox: [1.0, 2.0, 3.0, 4.0],
669 bottom_left: true,
670 charspan: [0, 5],
671 },
672 );
673 let pic = frag.add(
674 Some(slide1),
675 None,
676 TreeKind::Picture {
677 captions: vec![cap],
678 image: None,
679 classification: Some("bar_chart".into()),
680 confidence: None,
681 chart: None,
682 dpi: None,
683 },
684 );
685 let note = frag.add(
686 None,
687 Some(ContentLayer::Notes),
688 TreeKind::Group {
689 label: "comment_section".into(),
690 name: "comment-slide2-1".into(),
691 },
692 );
693 frag.items[pic].comments.push(note);
694
695 whole.append(frag);
696 assert_eq!(whole.body, vec![slide0, 2, 5]);
697 assert_eq!(whole.items[2].children, vec![3, 4]);
698 assert_eq!(whole.items[3].parent, Some(2));
699 assert_eq!(whole.items[3].prov.as_ref().map(|p| p.page_no), Some(2));
700 assert!(
701 matches!(&whole.items[4].kind, TreeKind::Picture { captions, .. } if captions == &[3])
702 );
703 assert_eq!(whole.items[4].comments, vec![5]);
704 assert_eq!(whole.items[5].parent, None);
705 assert_eq!(
706 whole.bucket_index(4),
707 0,
708 "the fragment's picture is #/pictures/0"
709 );
710 assert_eq!(
711 whole.bucket_index(5),
712 2,
713 "slide-0, slide-1 precede it in `groups`"
714 );
715 }
716
717 /// `renumber_in_traversal_order`: docling's `concatenate` re-creates the
718 /// items as a pre-order walk meets them, so a group created after the
719 /// content it was wrapped around (a rich cell's group) comes first, and a
720 /// deleted item disappears; every cross-reference follows.
721 #[test]
722 fn renumbering_follows_the_traversal() {
723 let mut t = ItemTree::default();
724 let a = t.add(None, None, text("a"));
725 let table = t.add(
726 None,
727 None,
728 TreeKind::Table {
729 table: Table::default(),
730 rich_cells: Vec::new(),
731 captions: Vec::new(),
732 },
733 );
734 let cell_text = t.add(None, None, text("cell"));
735 let group = t.add(
736 Some(table),
737 None,
738 TreeKind::Group {
739 label: "unspecified".into(),
740 name: "rich_cell_group_1_0_0".into(),
741 },
742 );
743 t.reparent(cell_text, Some(group));
744 if let TreeKind::Table { rich_cells, .. } = &mut t.items[table].kind {
745 rich_cells.push((0, 0, group));
746 }
747 let gone = t.add(None, None, text("gone"));
748 t.delete(gone);
749 let z = t.add(None, None, text("z"));
750
751 let new_of = t.renumber_in_traversal_order();
752 assert_eq!(
753 new_of,
754 vec![Some(0), Some(1), Some(3), Some(2), None, Some(4)]
755 );
756 assert_eq!(t.items.len(), 5);
757 assert_eq!(t.body, vec![0, 1, 4]);
758 assert_eq!(t.items[1].children, vec![2], "the group follows its table");
759 assert_eq!(t.items[2].parent, Some(1));
760 assert_eq!(
761 t.items[3].parent,
762 Some(2),
763 "the cell text follows its group"
764 );
765 assert!(matches!(&t.items[3].kind, TreeKind::Text { text, .. } if text == "cell"));
766 assert!(
767 matches!(&t.items[1].kind, TreeKind::Table { rich_cells, .. } if rich_cells == &[(0, 0, 2)])
768 );
769 let _ = (a, z);
770 }
771
772 /// #586: a misplaced `list_item` keeps its children when it is re-homed.
773 /// docling-core's `validate_misplaced_list_items` re-adds the item from
774 /// its text alone, so a mixed-format item (empty `list_item` over an
775 /// `inline` group of runs) lost every run — a DOCX table cell whose list
776 /// paragraph followed a `numId 0` spacer came out as an empty bullet in
777 /// the JSON and the HTML. The group goes where the item stood, the copy
778 /// is numbered last, the inline group and its texts move under it.
779 #[test]
780 fn wrapping_a_misplaced_list_item_keeps_its_runs() {
781 let mut t = ItemTree::default();
782 let cell = t.add(
783 None,
784 None,
785 TreeKind::Group {
786 label: "unspecified".into(),
787 name: "rich_cell_group_1_0_1".into(),
788 },
789 );
790 let before = t.add(Some(cell), None, text("before"));
791 let item = t.add(
792 Some(cell),
793 None,
794 TreeKind::Text {
795 label: "list_item".into(),
796 text: String::new(),
797 orig: None,
798 formatting: None,
799 hyperlink: None,
800 level: None,
801 list: Some(ListMeta {
802 enumerated: false,
803 marker: String::new(),
804 }),
805 },
806 );
807 let inline = t.add(
808 Some(item),
809 None,
810 TreeKind::Group {
811 label: "inline".into(),
812 name: "group".into(),
813 },
814 );
815 let run_a = t.add(Some(inline), None, text("Second item text"));
816 let run_b = t.add(Some(inline), None, text("[Optional]"));
817 let after = t.add(Some(cell), None, text("after"));
818
819 t.wrap_misplaced_list_items();
820
821 let group = t.items[cell].children[1];
822 assert_eq!(t.items[cell].children, vec![before, group, after]);
823 assert!(
824 matches!(&t.items[group].kind, TreeKind::Group { label, name } if label == "list" && name == "group")
825 );
826 assert!(t.items[item].deleted, "the original item is deleted");
827 let copy = t.items[group].children[0];
828 assert_ne!(copy, item);
829 assert!(
830 copy > after,
831 "the copy is numbered after every existing item"
832 );
833 assert_eq!(t.items[copy].children, vec![inline]);
834 assert_eq!(t.items[inline].parent, Some(copy));
835 assert_eq!(t.items[inline].children, vec![run_a, run_b]);
836 for id in [inline, run_a, run_b] {
837 assert!(!t.items[id].deleted, "item {id} must survive the re-homing");
838 }
839 }
840}