Skip to main content

docling_core/
markdown.rs

1//! Markdown serializer for [`DoclingDocument`].
2
3use crate::document::{DoclingDocument, Node, Table};
4
5/// How pictures are rendered (mirrors docling-core's `ImageRefMode`).
6#[derive(Debug, Clone, Copy, PartialEq, Eq, Default)]
7pub enum ImageMode {
8    /// `<!-- image -->` (docling's default, and the only mode without image data).
9    #[default]
10    Placeholder,
11    /// `![Image](data:<mime>;base64,…)` — self-contained.
12    Embedded,
13    /// `![Image](<artifacts>/image_NNNNNN.<ext>)`; the bytes are returned for the
14    /// caller to write.
15    Referenced,
16}
17
18/// Serializer state threaded through the render walk.
19struct Ctx {
20    strict: bool,
21    /// Emit compact `| a | b |` tables instead of the padded GitHub serializer.
22    compact_tables: bool,
23    images: ImageMode,
24    artifacts_dir: String,
25    /// (relative path, bytes) for each referenced image — written by the caller.
26    artifacts: Vec<(String, Vec<u8>)>,
27    pic_index: usize,
28}
29
30/// Render a document to a Markdown string (pictures as placeholders).
31///
32/// `strict` selects the serializer-level behaviours that differ between
33/// docling-legacy output and cleaner Markdown — currently the code-fence
34/// language (legacy drops it, strict keeps it).
35pub fn to_markdown(doc: &DoclingDocument, strict: bool) -> String {
36    to_markdown_images(doc, strict, ImageMode::Placeholder, "artifacts").0
37}
38
39/// Render to Markdown with an explicit picture [`ImageMode`]. Returns the
40/// Markdown and, for [`ImageMode::Referenced`], the `(path, bytes)` of each image
41/// the caller should write (relative to the Markdown file).
42pub fn to_markdown_images(
43    doc: &DoclingDocument,
44    strict: bool,
45    images: ImageMode,
46    artifacts_dir: &str,
47) -> (String, Vec<(String, Vec<u8>)>) {
48    let mut ctx = Ctx {
49        strict,
50        compact_tables: doc.compact_tables,
51        images,
52        artifacts_dir: artifacts_dir.to_string(),
53        artifacts: Vec::new(),
54        pic_index: 0,
55    };
56    let mut blocks: Vec<String> = Vec::new();
57    render(&doc.nodes, &mut blocks, &mut ctx);
58    let mut body = blocks.join("\n\n");
59    // Strict mode only: turn recovered source hyperlinks into Markdown links.
60    // docling's standard pipeline drops them, so doing this in legacy mode would
61    // diverge from docling — hence strict-only, leaving conformance output intact.
62    if strict && !doc.links.is_empty() {
63        body = apply_links(&body, &doc.links);
64    }
65    let md = if body.is_empty() {
66        String::new()
67    } else {
68        format!("{body}\n")
69    };
70    (md, ctx.artifacts)
71}
72
73/// Wrap each recovered link's anchor text in Markdown `[anchor](href)`. Anchors
74/// arrive cleaned (curly quotes/dashes already normalized) but un-escaped, so we
75/// match against the body's HTML-escaped (`&`/`<`/`>`) form, the way prose nodes
76/// were serialized. Links are consumed in document order from a moving cursor, so
77/// a repeated anchor (e.g. two "issues") links its successive occurrences rather
78/// than all pointing at the first. An anchor that can't be located is skipped
79/// (its text may have been split across a line wrap or table cell).
80fn apply_links(body: &str, links: &[(String, String)]) -> String {
81    let mut out = body.to_string();
82    let mut cursor = 0usize;
83    for (anchor, href) in links {
84        let anchor = anchor
85            .replace('&', "&amp;")
86            .replace('<', "&lt;")
87            .replace('>', "&gt;");
88        if anchor.is_empty() {
89            continue;
90        }
91        if let Some(rel) = out[cursor..].find(&anchor) {
92            let at = cursor + rel;
93            // Don't relink inside an already-emitted `](` Markdown link target.
94            let replacement = format!("[{anchor}]({href})");
95            out.replace_range(at..at + anchor.len(), &replacement);
96            cursor = at + replacement.len();
97        }
98    }
99    out
100}
101
102/// Like [`apply_links`] but over a single chunk, consuming from a shared queue so
103/// the same `[anchor](href)` rewriting can be applied incrementally as Markdown is
104/// streamed out. Each queued link is matched (in document order) against `chunk`
105/// and rewritten in place; a link whose anchor is not in this chunk is carried
106/// forward in the queue for a later chunk. Anchors are recovered in document
107/// order and a chunk is always a contiguous run of whole blocks, so this
108/// reproduces [`apply_links`]' single moving cursor: the link lands in whichever
109/// chunk contains its anchor, identically to the buffered path. (A link whose
110/// anchor never appears is carried to the end and dropped — the same no-op
111/// `apply_links` performs for an unlocatable anchor.)
112fn apply_links_chunk(chunk: &str, queue: &mut Vec<(String, String)>) -> String {
113    let mut out = chunk.to_string();
114    let mut cursor = 0usize;
115    let mut carried: Vec<(String, String)> = Vec::new();
116    for (anchor_raw, href) in std::mem::take(queue) {
117        let anchor = anchor_raw
118            .replace('&', "&amp;")
119            .replace('<', "&lt;")
120            .replace('>', "&gt;");
121        if anchor.is_empty() {
122            continue;
123        }
124        if let Some(rel) = out[cursor..].find(&anchor) {
125            let at = cursor + rel;
126            let replacement = format!("[{anchor}]({href})");
127            out.replace_range(at..at + anchor.len(), &replacement);
128            cursor = at + replacement.len();
129        } else {
130            // Not in this chunk; try again when its block is flushed.
131            carried.push((anchor_raw, href));
132        }
133    }
134    *queue = carried;
135    out
136}
137
138/// Incremental Markdown serializer: feed finalized, in-document-order batches of
139/// [`Node`]s and receive Markdown chunks whose concatenation is **byte-identical**
140/// to [`to_markdown_images`] over the same nodes. This is the streaming
141/// counterpart of the buffered serializer — used to emit a document's Markdown in
142/// chunks (e.g. page by page, as the parallel PDF pipeline finishes pages) instead
143/// of building the whole string up front.
144///
145/// Only [`ImageMode::Placeholder`] and [`ImageMode::Embedded`] are streamable:
146/// [`ImageMode::Referenced`] needs a side-channel for the image bytes, which only
147/// the buffered [`to_markdown_images`] provides.
148///
149/// Each [`push`](Self::push) must contain whole blocks in reading order: a caller
150/// must not split a run of list items across two pushes (the run would render as
151/// two separate lists). Finalized PDF page batches already satisfy this.
152pub struct MarkdownStreamer {
153    strict: bool,
154    images: ImageMode,
155    compact_tables: bool,
156    /// Whether any non-empty chunk has been emitted yet (drives `\n\n` joins and
157    /// the trailing newline).
158    emitted_any: bool,
159    /// Recovered links not yet placed (strict mode), consumed in document order.
160    links: Vec<(String, String)>,
161}
162
163impl MarkdownStreamer {
164    /// Create a streamer. `compact_tables` mirrors [`DoclingDocument::compact_tables`].
165    pub fn new(strict: bool, images: ImageMode, compact_tables: bool) -> Self {
166        debug_assert!(
167            images != ImageMode::Referenced,
168            "referenced image mode is not streamable; use to_markdown_images"
169        );
170        Self {
171            strict,
172            images,
173            compact_tables,
174            emitted_any: false,
175            links: Vec::new(),
176        }
177    }
178
179    /// Render one finalized batch of nodes (plus any links recovered from the same
180    /// span, in document order) into the next Markdown chunk. Returns an empty
181    /// string when the batch produces no output (e.g. empty tables/pictures), in
182    /// which case nothing should be written.
183    pub fn push(&mut self, nodes: &[Node], links: &[(String, String)]) -> String {
184        self.links.extend(links.iter().cloned());
185        let mut ctx = Ctx {
186            strict: self.strict,
187            compact_tables: self.compact_tables,
188            images: self.images,
189            // Referenced mode is rejected at construction, so the artifact sink is
190            // never touched.
191            artifacts_dir: String::new(),
192            artifacts: Vec::new(),
193            pic_index: 0,
194        };
195        let mut blocks: Vec<String> = Vec::new();
196        render(nodes, &mut blocks, &mut ctx);
197        if blocks.is_empty() {
198            return String::new();
199        }
200        let mut body = blocks.join("\n\n");
201        if self.strict && !self.links.is_empty() {
202            body = apply_links_chunk(&body, &mut self.links);
203        }
204        let chunk = if self.emitted_any {
205            format!("\n\n{body}")
206        } else {
207            body
208        };
209        self.emitted_any = true;
210        chunk
211    }
212
213    /// Emit the trailing newline that finishes the document (empty if no content
214    /// was produced). Call exactly once, after the final [`push`](Self::push).
215    pub fn finish(self) -> String {
216        if self.emitted_any {
217            "\n".to_string()
218        } else {
219            String::new()
220        }
221    }
222}
223
224/// In `strict` mode, rewrite inline text for readability rather than byte-for-byte
225/// docling fidelity: undo the legacy `\_` underscore escaping, and tighten stray
226/// spaces around punctuation (`[ 37 , 36 ]` → `[37, 36]`, `( x )` → `(x)`). This
227/// cleans up both the PDF backend's glyph-split spacing and the space the legacy
228/// emphasis serialization leaves before punctuation (`*a* ,` → `*a*,`).
229/// Legacy/default output keeps docling's spacing untouched. Only inline text
230/// nodes pass through here — code blocks and table cells are left alone.
231fn strict_text(text: &str, strict: bool) -> String {
232    if !strict {
233        return text.to_string();
234    }
235    text.replace("\\_", "_")
236        .replace(" ,", ",")
237        .replace(" .", ".")
238        .replace(" ;", ";")
239        .replace(" )", ")")
240        .replace("( ", "(")
241        .replace(" ]", "]")
242        .replace("[ ", "[")
243}
244
245fn render(nodes: &[Node], blocks: &mut Vec<String>, ctx: &mut Ctx) {
246    let mut i = 0;
247    while i < nodes.len() {
248        match &nodes[i] {
249            Node::ListItem { .. } => {
250                let start = i;
251                i += 1;
252                loop {
253                    match nodes.get(i) {
254                        Some(Node::ListItem { .. }) => i += 1,
255                        // An empty paragraph between two list items is absorbed
256                        // into the run — docling keeps such a ListGroup
257                        // contiguous rather than splitting it.
258                        Some(Node::Paragraph { text })
259                            if text.is_empty()
260                                && matches!(nodes.get(i + 1), Some(Node::ListItem { .. })) =>
261                        {
262                            i += 1
263                        }
264                        _ => break,
265                    }
266                }
267                render_list_run(&nodes[start..i], blocks, ctx.strict);
268            }
269            other => {
270                render_one(other, blocks, ctx);
271                i += 1;
272            }
273        }
274    }
275}
276
277/// Render a contiguous run of list items.
278///
279/// Ordered items use their explicit `number`. A new sibling list (marked by
280/// `first_in_list`) at the same depth is separated by a blank line, matching
281/// docling-core's serializer.
282fn render_list_run(items: &[Node], blocks: &mut Vec<String>, strict: bool) {
283    let mut lines: Vec<String> = Vec::new();
284    // Per level, the previous item's (ordered, number) so we can detect a new
285    // sibling list.
286    let mut prev: Vec<Option<(bool, u64)>> = Vec::new();
287
288    for item in items {
289        let Node::ListItem {
290            ordered,
291            number,
292            first_in_list,
293            text,
294            level,
295            marker: _,
296            location: _,
297            dclx: _,
298            href: _,
299            layer,
300        } = item
301        else {
302            continue;
303        };
304        // A non-body (furniture) list item is omitted from Markdown, matching
305        // docling's content-layer filtering.
306        if layer.is_some() {
307            continue;
308        }
309        let level = *level as usize;
310
311        // Returning to a shallower level ends the deeper sibling lists.
312        prev.truncate(level + 1);
313        while prev.len() <= level {
314            prev.push(None);
315        }
316
317        // A new sibling list at the same depth gets a blank line: the kind flips
318        // (`<ul>`↔`<ol>`), an ordered run breaks (`1, 2` then `42`), or the
319        // backend flagged a fresh list (e.g. Markdown's bullet changing `-`→`*`).
320        // Only at the top level: nested sibling groups are children of a list
321        // item, and docling joins an item's children without blank lines.
322        if level == 0 {
323            if let Some((prev_ordered, prev_number)) = prev[level] {
324                let new_list = *first_in_list
325                    || prev_ordered != *ordered
326                    || (*ordered && *number != prev_number + 1);
327                if new_list {
328                    lines.push(String::new());
329                }
330            }
331        }
332
333        let indent = "    ".repeat(level);
334        let marker = if *ordered {
335            format!("{number}.")
336        } else {
337            "-".to_string()
338        };
339        lines.push(format!("{indent}{marker} {}", strict_text(text, strict)));
340        prev[level] = Some((*ordered, *number));
341    }
342
343    // A run consisting only of furniture (content-layer-filtered) items yields no
344    // lines; pushing an empty block here would surface as a stray blank line.
345    if !lines.is_empty() {
346        blocks.push(lines.join("\n"));
347    }
348}
349
350fn render_one(node: &Node, blocks: &mut Vec<String>, ctx: &mut Ctx) {
351    match node {
352        Node::Heading { level, text } => {
353            let hashes = "#".repeat((*level).clamp(1, 6) as usize);
354            blocks.push(format!("{hashes} {}", strict_text(text, ctx.strict)));
355        }
356        // An empty body paragraph (docling's blank-line text item) contributes
357        // nothing to Markdown — only DocLang/JSON keep it.
358        Node::Paragraph { text } if text.is_empty() => {}
359        Node::Paragraph { text } => blocks.push(strict_text(text, ctx.strict)),
360        Node::CheckboxItem { checked, text } => {
361            let mark = if *checked { "- [x] " } else { "- [ ] " };
362            blocks.push(strict_text(&format!("{mark}{text}"), ctx.strict));
363        }
364        Node::Code { language, text, .. } => {
365            // Legacy docling never emits a language on the fence; strict keeps it.
366            let lang = match language {
367                Some(l) if ctx.strict => l.as_str(),
368                _ => "",
369            };
370            blocks.push(format!("```{lang}\n{text}\n```"));
371        }
372        // A CodeFormula-decoded display formula renders as docling's `$$…$$`
373        // (the un-enriched pipeline emits a placeholder paragraph instead).
374        Node::Formula { latex, .. } => blocks.push(format!("$${latex}$$")),
375        Node::Table(table) => {
376            let rendered = render_table(table, ctx.compact_tables);
377            if !rendered.is_empty() {
378                blocks.push(rendered);
379            }
380        }
381        // Classification predictions don't affect docling's Markdown output.
382        Node::Picture { caption, image, .. } => {
383            if let Some(cap) = caption {
384                if !cap.is_empty() {
385                    blocks.push(cap.clone());
386                }
387            }
388            blocks.push(picture_marker(image.as_ref(), ctx));
389        }
390        // A chart renders as docling's picture-with-meta markdown: the caption,
391        // the placeholder, the humanized classification ("line_chart" ->
392        // "Line chart"), then the chart's data grid as a regular table.
393        Node::Chart {
394            kind,
395            table,
396            caption,
397            ..
398        } => {
399            if let Some(cap) = caption {
400                if !cap.is_empty() {
401                    blocks.push(cap.clone());
402                }
403            }
404            blocks.push(picture_marker(None, ctx));
405            blocks.push(humanize_label(kind));
406            let rendered = render_table(table, false);
407            if !rendered.is_empty() {
408                blocks.push(rendered);
409            }
410        }
411        // A DocLang-only node is omitted from Markdown.
412        Node::DoclangOnly(_) => {}
413        Node::Group { children, .. } => render(children, blocks, ctx),
414        Node::FieldRegion { items } => {
415            // docling renders the region container (which carries no text of its
416            // own) as a `<!-- missing-text -->` marker, then each field item the
417            // same way, followed by that item's marker/key/value as separate
418            // paragraphs.
419            blocks.push(MISSING_TEXT.to_string());
420            for item in items {
421                blocks.push(MISSING_TEXT.to_string());
422                for part in [&item.marker, &item.key, &item.value].into_iter().flatten() {
423                    blocks.push(strict_text(part, ctx.strict));
424                }
425            }
426        }
427        // A rich inline group renders exactly like a paragraph of its Markdown
428        // text — the structured runs are DocLang-only.
429        Node::InlineGroup { md_text, .. } => blocks.push(strict_text(md_text, ctx.strict)),
430        // A plain-text backend dump renders verbatim as a single block.
431        Node::TextDump(text) => {
432            if !text.is_empty() {
433                blocks.push(text.clone());
434            }
435        }
436        // Furniture (page headers/footers, HTML `<title>`) is excluded from
437        // Markdown by default, mirroring docling.
438        Node::Furniture { .. } => {}
439        Node::PageFurniture { .. } => {}
440        // Layout provenance is DocLang-only; render the wrapped node.
441        Node::Located { inner, .. } => render_one(inner, blocks, ctx),
442        // Page breaks are DocLang-only; docling omits them from Markdown.
443        Node::PageBreak => {}
444        // Handled by the run-merging branch in `render`.
445        Node::ListItem { .. } => unreachable!("list items are rendered in runs"),
446    }
447}
448
449/// docling's placeholder for a structural node (a field region / item) that has
450/// no text of its own.
451const MISSING_TEXT: &str = "<!-- missing-text -->";
452
453/// The Markdown for a picture under the active [`ImageMode`]; Referenced mode also
454/// records the bytes in `ctx.artifacts` for the caller to write.
455/// docling-core's `_humanize_text`: underscores to spaces, first letter
456/// capitalized ("line_chart" -> "Line chart").
457fn humanize_label(label: &str) -> String {
458    let text = label.replace('_', " ");
459    let mut chars = text.chars();
460    match chars.next() {
461        Some(f) => f.to_uppercase().collect::<String>() + chars.as_str(),
462        None => text,
463    }
464}
465
466fn picture_marker(image: Option<&crate::PictureImage>, ctx: &mut Ctx) -> String {
467    match (ctx.images, image) {
468        (ImageMode::Embedded, Some(img)) => format!("![Image]({})", img.data_uri()),
469        (ImageMode::Referenced, Some(img)) => {
470            let path = format!(
471                "{}/image_{:06}.{}",
472                ctx.artifacts_dir,
473                ctx.pic_index,
474                ext_for(&img.mimetype)
475            );
476            ctx.pic_index += 1;
477            ctx.artifacts.push((path.clone(), img.data.clone()));
478            format!("![Image]({path})")
479        }
480        // Placeholder, or any mode with no extracted image.
481        _ => "<!-- image -->".to_string(),
482    }
483}
484
485fn ext_for(mimetype: &str) -> &str {
486    match mimetype {
487        "image/jpeg" => "jpg",
488        "image/gif" => "gif",
489        "image/webp" => "webp",
490        "image/bmp" => "bmp",
491        "image/tiff" => "tif",
492        _ => "png",
493    }
494}
495
496/// Render a table. `compact` selects between two serializers:
497///
498/// - **padded** (default) — docling-core's `tabulate(tablefmt="github")`: columns
499///   are padded to a fixed width (header width + a minimum padding of 2, or the
500///   widest data cell); numeric columns (every data cell parses as a number) are
501///   right-aligned, others left-aligned; separators are plain dashes of
502///   `width + 2`. Matches current published docling (DOCX/HTML conformance).
503/// - **compact** — `| a | b |` cells with single-dash `| - | - |` separators, no
504///   width padding. Matches the committed PDF groundtruth corpus, which predates
505///   the padded serializer.
506///
507/// Each cell is first escaped (`\n` → space, `|` → `&#124;`) so it can't break the
508/// table. Row 0 is the header.
509/// Whether a table cell counts as a number for column alignment, matching
510/// `tabulate`'s detection: an ordinary float/int (`f64`-parseable, covering
511/// `1e2`/`inf`/`+1.5`) **or** a thousands-separated number like `7,015`.
512fn is_number_cell(t: &str) -> bool {
513    t.parse::<f64>().is_ok() || is_thousands_number(t)
514}
515
516/// A number with comma thousands-separators, per `tabulate`'s
517/// `_float_with_thousands_separators` regex
518/// (`^(([+-]?[0-9]{1,3})(?:,([0-9]{3}))*)?(?(1)\.[0-9]*|\.[0-9]+)?$`): the
519/// integer part is 1–3 digits then any number of `,ddd` groups; the fraction is
520/// optional (and, without an integer part, must have at least one digit).
521fn is_thousands_number(t: &str) -> bool {
522    let b = t.as_bytes();
523    let mut i = 0;
524    let start = i;
525    if i < b.len() && (b[i] == b'+' || b[i] == b'-') {
526        i += 1;
527    }
528    // First digit chunk: 1–3 digits.
529    let d0 = i;
530    while i < b.len() && b[i].is_ascii_digit() && i - d0 < 3 {
531        i += 1;
532    }
533    let has_int = i > d0;
534    if has_int {
535        // Subsequent `,ddd` groups (exactly three digits each).
536        while i + 3 < b.len() + 1
537            && b.get(i) == Some(&b',')
538            && b.get(i + 1).is_some_and(u8::is_ascii_digit)
539            && b.get(i + 2).is_some_and(u8::is_ascii_digit)
540            && b.get(i + 3).is_some_and(u8::is_ascii_digit)
541        {
542            i += 4;
543        }
544    } else {
545        // A sign only counts with an integer part.
546        i = start;
547    }
548    // Optional fraction.
549    if i < b.len() && b[i] == b'.' {
550        i += 1;
551        let f0 = i;
552        while i < b.len() && b[i].is_ascii_digit() {
553            i += 1;
554        }
555        if !has_int && i == f0 {
556            return false; // `.` with no digits and no integer part
557        }
558    } else if !has_int {
559        return false; // neither integer nor fractional part
560    }
561    i == b.len()
562}
563
564pub(crate) fn render_table(table: &Table, compact: bool) -> String {
565    if table.rows.is_empty() {
566        return String::new();
567    }
568    let num_cols = table.rows.iter().map(Vec::len).max().unwrap_or(0);
569    if num_cols == 0 {
570        return String::new();
571    }
572
573    // Escaped, rectangular grid (ragged rows padded with empty cells). `tabulate`
574    // strips data cells of surrounding whitespace but leaves the header row as-is.
575    let grid: Vec<Vec<String>> = table
576        .rows
577        .iter()
578        .enumerate()
579        .map(|(r, row)| {
580            (0..num_cols)
581                .map(|c| {
582                    let cell = escape_cell(row.get(c).map(String::as_str).unwrap_or(""));
583                    if r == 0 {
584                        cell
585                    } else {
586                        cell.trim().to_string()
587                    }
588                })
589                .collect()
590        })
591        .collect();
592
593    if compact {
594        // Compact: cells joined by " | ", no padding, single-dash separators.
595        let render_row = |r: usize| -> String { format!("| {} |", grid[r].join(" | ")) };
596        let mut lines = Vec::with_capacity(grid.len() + 1);
597        lines.push(render_row(0));
598        let sep: Vec<&str> = (0..num_cols).map(|_| "-").collect();
599        lines.push(format!("| {} |", sep.join(" | ")));
600        for r in 1..grid.len() {
601            lines.push(render_row(r));
602        }
603        return lines.join("\n");
604    }
605
606    // Display width (Unicode scalar count — good enough for now).
607    let dw = |s: &str| s.chars().count();
608    let data_rows = 1..grid.len();
609
610    // A column is right-aligned when at least one data cell is numeric and every
611    // non-empty data cell is numeric — matching `tabulate`'s column typing, where
612    // empty cells are "missing" (ignored) and a number may carry thousands
613    // separators (`7,015`), which a plain `f64` parse rejects.
614    let right: Vec<bool> = (0..num_cols)
615        .map(|c| {
616            let mut any = false;
617            for r in data_rows.clone() {
618                let t = grid[r][c].trim();
619                if t.is_empty() {
620                    continue;
621                }
622                if !is_number_cell(t) {
623                    return false;
624                }
625                any = true;
626            }
627            any
628        })
629        .collect();
630
631    // Column width = max(header_width + MIN_PADDING(2), max data-cell width).
632    let width: Vec<usize> = (0..num_cols)
633        .map(|c| {
634            let mut w = dw(&grid[0][c]) + 2;
635            for r in data_rows.clone() {
636                w = w.max(dw(&grid[r][c]));
637            }
638            w
639        })
640        .collect();
641
642    let fmt_cell = |s: &str, c: usize| -> String {
643        let pad = " ".repeat(width[c].saturating_sub(dw(s)));
644        let body = if right[c] {
645            format!("{pad}{s}")
646        } else {
647            format!("{s}{pad}")
648        };
649        format!(" {body} ")
650    };
651    let render_row = |r: usize| -> String {
652        let cells: Vec<String> = (0..num_cols).map(|c| fmt_cell(&grid[r][c], c)).collect();
653        format!("|{}|", cells.join("|"))
654    };
655
656    let mut lines = Vec::with_capacity(grid.len() + 1);
657    lines.push(render_row(0));
658    let sep: Vec<String> = (0..num_cols).map(|c| "-".repeat(width[c] + 2)).collect();
659    lines.push(format!("|{}|", sep.join("|")));
660    for r in data_rows {
661        lines.push(render_row(r));
662    }
663    lines.join("\n")
664}
665
666/// Escape a table cell so it can't break the markdown table: newlines become
667/// spaces and pipes become the `&#124;` HTML entity (matches docling-core).
668fn escape_cell(s: &str) -> String {
669    s.replace('\n', " ").replace('|', "&#124;")
670}
671
672#[cfg(test)]
673mod tests {
674    use super::*;
675
676    #[test]
677    fn renders_headings_paragraphs_and_lists() {
678        let mut doc = DoclingDocument::new("demo");
679        doc.add_heading(1, "Title");
680        doc.add_paragraph("Hello world.");
681        doc.push(Node::ListItem {
682            ordered: false,
683            number: 1,
684            first_in_list: true,
685            text: "first".into(),
686            level: 0,
687            marker: None,
688            location: None,
689            dclx: None,
690            href: None,
691            layer: None,
692        });
693        doc.push(Node::ListItem {
694            ordered: false,
695            number: 2,
696            first_in_list: false,
697            text: "second".into(),
698            level: 0,
699            marker: None,
700            location: None,
701            dclx: None,
702            href: None,
703            layer: None,
704        });
705        let md = doc.export_to_markdown();
706        assert_eq!(md, "# Title\n\nHello world.\n\n- first\n- second\n");
707    }
708
709    #[test]
710    fn strict_renders_recovered_links_legacy_does_not() {
711        let mut doc = DoclingDocument::new("cv");
712        doc.add_paragraph("Find me on LinkedIn or GitHub.");
713        doc.links = vec![
714            ("LinkedIn".into(), "https://www.linkedin.com/in/x/".into()),
715            ("GitHub".into(), "https://github.com/x/".into()),
716        ];
717        // Legacy/docling mode: links are left untouched (conformance preserved).
718        assert_eq!(doc.export_to_markdown(), "Find me on LinkedIn or GitHub.\n");
719        // Strict mode: anchors become Markdown links.
720        assert_eq!(
721            doc.export_to_markdown_with(true),
722            "Find me on [LinkedIn](https://www.linkedin.com/in/x/) or [GitHub](https://github.com/x/).\n"
723        );
724    }
725
726    #[test]
727    fn strict_links_match_escaped_anchor_and_consume_in_order() {
728        let mut doc = DoclingDocument::new("d");
729        // The PDF assembler HTML-escapes prose, so by serialization time the body
730        // already carries `&amp;`; the anchor is stored un-escaped. The matcher must
731        // escape the anchor to find it. Two identical anchors link in document order.
732        doc.add_paragraph("AI &amp; ML here, and issues here, then issues there.");
733        doc.links = vec![
734            ("AI & ML".into(), "https://a/".into()),
735            ("issues".into(), "https://first/".into()),
736            ("issues".into(), "https://second/".into()),
737        ];
738        assert_eq!(
739            doc.export_to_markdown_with(true),
740            "[AI &amp; ML](https://a/) here, and [issues](https://first/) here, then [issues](https://second/) there.\n"
741        );
742    }
743
744    #[test]
745    fn renders_compact_table() {
746        let mut doc = DoclingDocument::new("t");
747        // The compact form is opt-in (the PDF backend sets it); default output uses
748        // the padded GitHub serializer (covered by the regression fixtures).
749        doc.compact_tables = true;
750        doc.push(Node::Table(Table {
751            rows: vec![vec!["a".into(), "b".into()], vec!["1".into(), "2".into()]],
752            location: None,
753            structure: None,
754            cell_blocks: None,
755        }));
756        let md = doc.export_to_markdown();
757        assert_eq!(md, "| a | b |\n| - | - |\n| 1 | 2 |\n");
758    }
759
760    #[test]
761    fn renders_padded_github_table_by_default() {
762        let mut doc = DoclingDocument::new("t");
763        doc.push(Node::Table(Table {
764            rows: vec![vec!["a".into(), "b".into()], vec!["1".into(), "2".into()]],
765            location: None,
766            structure: None,
767            cell_blocks: None,
768        }));
769        let md = doc.export_to_markdown();
770        // Numeric data columns are right-aligned; columns padded to header+2.
771        assert_eq!(md, "|   a |   b |\n|-----|-----|\n|   1 |   2 |\n");
772    }
773
774    #[test]
775    fn strict_unescapes_inline_underscores_legacy_keeps_them() {
776        let mut doc = DoclingDocument::new("t");
777        doc.add_heading(1, "a\\_b");
778        doc.add_paragraph("x\\_y");
779        doc.push(Node::ListItem {
780            ordered: false,
781            number: 1,
782            first_in_list: true,
783            text: "i\\_j".into(),
784            level: 0,
785            marker: None,
786            location: None,
787            dclx: None,
788            href: None,
789            layer: None,
790        });
791        // Legacy reproduces docling's `\_` escaping byte-for-byte.
792        assert_eq!(doc.export_to_markdown(), "# a\\_b\n\nx\\_y\n\n- i\\_j\n");
793        // Strict prefers literal underscores (Rust-only readability mode).
794        assert_eq!(doc.export_to_markdown_with(true), "# a_b\n\nx_y\n\n- i_j\n");
795    }
796
797    /// Drive a document's nodes through [`MarkdownStreamer`] in the given page
798    /// splits and assert the concatenated chunks equal the buffered serializer.
799    fn assert_stream_matches(
800        doc: &DoclingDocument,
801        strict: bool,
802        images: ImageMode,
803        splits: &[usize],
804    ) {
805        let want = to_markdown_images(doc, strict, images, "artifacts").0;
806        let mut streamer = MarkdownStreamer::new(strict, images, doc.compact_tables);
807        let mut got = String::new();
808        let mut start = 0;
809        for &end in splits {
810            // Links only matter in strict mode; feed them all with the first batch
811            // that has content (document order is preserved by the queue).
812            let links = if start == 0 {
813                doc.links.as_slice()
814            } else {
815                &[]
816            };
817            got.push_str(&streamer.push(&doc.nodes[start..end], links));
818            start = end;
819        }
820        got.push_str(&streamer.push(
821            &doc.nodes[start..],
822            if start == 0 {
823                doc.links.as_slice()
824            } else {
825                &[]
826            },
827        ));
828        got.push_str(&streamer.finish());
829        assert_eq!(
830            got, want,
831            "streamed output diverged (splits={splits:?}, strict={strict})"
832        );
833    }
834
835    #[test]
836    fn streaming_is_byte_identical_to_buffered() {
837        let mut doc = DoclingDocument::new("d");
838        doc.add_heading(1, "Title");
839        doc.add_paragraph("First paragraph.");
840        doc.push(Node::ListItem {
841            ordered: false,
842            number: 1,
843            first_in_list: true,
844            text: "a".into(),
845            level: 0,
846            marker: None,
847            location: None,
848            dclx: None,
849            href: None,
850            layer: None,
851        });
852        doc.push(Node::ListItem {
853            ordered: false,
854            number: 2,
855            first_in_list: false,
856            text: "b".into(),
857            level: 0,
858            marker: None,
859            location: None,
860            dclx: None,
861            href: None,
862            layer: None,
863        });
864        doc.push(Node::Code {
865            language: Some("rust".into()),
866            text: "let x = 1;".into(),
867            orig: None,
868        });
869        doc.push(Node::Table(Table {
870            rows: vec![vec!["a".into(), "b".into()], vec!["1".into(), "2".into()]],
871            location: None,
872            structure: None,
873            cell_blocks: None,
874        }));
875        doc.push(Node::Picture {
876            caption: Some("Fig 1".into()),
877            image: None,
878            classification: None,
879        });
880        doc.add_paragraph("Last paragraph.");
881
882        // A run of list items must never straddle a split, so try splits that fall
883        // on safe block boundaries (the streaming PDF assembler guarantees this).
884        for &strict in &[false, true] {
885            for &images in &[ImageMode::Placeholder, ImageMode::Embedded] {
886                for splits in [&[][..], &[1][..], &[2][..], &[4][..], &[1, 4, 6][..]] {
887                    assert_stream_matches(&doc, strict, images, splits);
888                }
889            }
890        }
891    }
892
893    #[test]
894    fn streaming_applies_recovered_links_in_strict_mode() {
895        let mut doc = DoclingDocument::new("d");
896        doc.add_paragraph("See LinkedIn for details.");
897        doc.add_paragraph("And GitHub too.");
898        doc.links = vec![
899            ("LinkedIn".into(), "https://lnkd/".into()),
900            ("GitHub".into(), "https://gh/".into()),
901        ];
902        // The second anchor lives in the second block, so it must be carried across
903        // the page boundary and placed when that block streams out.
904        assert_stream_matches(&doc, true, ImageMode::Placeholder, &[1]);
905    }
906
907    #[test]
908    fn strict_tightens_punctuation_spacing_legacy_keeps_it() {
909        let mut doc = DoclingDocument::new("t");
910        doc.add_paragraph("see [ 37 , 36 ] and ( x ) .");
911        // Legacy keeps docling's spacing byte-for-byte.
912        assert_eq!(doc.export_to_markdown(), "see [ 37 , 36 ] and ( x ) .\n");
913        // Strict tightens punctuation for readable Markdown.
914        assert_eq!(doc.export_to_markdown_with(true), "see [37, 36] and (x).\n");
915    }
916}