docling_pdf/lib.rs
1//! PDF backend for docling.rs.
2//!
3//! A port of docling's standard PDF pipeline: pdfium extracts the text layer
4//! (cells with bounding boxes) and renders page images; a discriminative ONNX
5//! stack (layout detection, table structure, OCR) classifies regions; the cells
6//! are assembled in reading order into a [`DoclingDocument`].
7//!
8//! Current stages: pdfium text-cell extraction + page rendering ([`pdfium_backend`])
9//! and the deterministic text/reading-order assembly ([`assemble`]). The layout,
10//! table-structure and OCR ONNX stages land behind [`Pipeline`] next.
11
12// Without `ml` only the text-layer path runs; the shared assembly/label
13// helpers it doesn't exercise stay compiled for API stability (the full
14// build still flags genuinely dead code).
15#![cfg_attr(not(feature = "ml"), allow(dead_code))]
16
17// Reading-order assembly. Public under `ocr-prep` so the browser pipeline can
18// reuse the geometric table reconstruction and its reliability gate (#157).
19#[cfg(feature = "ocr-prep")]
20pub mod assemble;
21#[cfg(not(feature = "ocr-prep"))]
22mod assemble;
23// Vector checkbox squares from the text parser's path walk (#609).
24pub mod checkbox;
25mod dp_lines;
26#[cfg(feature = "ml")]
27pub mod enrich;
28// Public so sibling crates (e.g. docling-rag's ONNX embedder) can route their
29// own `ort` sessions through the same `DOCLING_RS_EP` selection. Kept as a
30// re-export after the logic moved to the shared `docling-onnx` crate —
31// `docling_pdf::ep::…` remains the stable path downstream crates code against.
32#[cfg(feature = "ml")]
33pub use docling_onnx as ep;
34// Heading-hierarchy stage (#302): PDF font-name style parsing, outline
35// extraction (pure lopdf), and the level-assignment pass. No feature gate on
36// the logic itself — only the glyph style pass needs pdfium (`ml`).
37#[cfg(feature = "render")]
38pub mod dparse_render;
39mod font_style;
40mod heading_hierarchy;
41pub mod layout;
42#[cfg(feature = "ml")]
43mod mets;
44#[cfg(feature = "ner")]
45pub mod ner;
46#[cfg(feature = "ml")]
47mod ocr;
48#[cfg(feature = "ml")]
49mod ocr_cls;
50#[cfg(any(feature = "ml", feature = "ocr-prep"))]
51pub mod ocr_det;
52#[cfg(feature = "ocr-prep")]
53pub mod ocr_prep;
54#[cfg(feature = "ml")]
55mod orient;
56pub mod outline;
57pub mod pdfium_backend;
58#[cfg(feature = "ml")]
59pub mod picture;
60#[cfg(feature = "ml")]
61pub mod quality;
62mod reading_order;
63#[cfg(feature = "ml")]
64mod redact_ocr;
65// Pure-Rust region resampling (page→1024px box-average, crop→448 bilinear) —
66// available to the browser TableFormer path (#157 stage 3), not just `ml`.
67#[cfg(feature = "ocr-prep")]
68pub mod resample;
69#[cfg(feature = "ocr-prep")]
70pub mod scanned;
71// Built-in standard-14 font metrics for the pure-Rust text parser (#187) —
72// no feature gate: the wasm/pdf-text path needs them like the native one.
73pub mod info;
74pub mod pdf_meta;
75// What the reader repaired to read a file (#934) — typed provenance, on
76// demand; pure Rust like `pdf_meta`.
77pub mod repairs;
78pub use repairs::{pdf_repairs, PdfRepair};
79#[cfg(feature = "render")]
80pub mod raster;
81#[cfg(feature = "render")]
82pub mod render;
83mod std14;
84#[cfg(feature = "ml")]
85pub mod tableformer;
86#[cfg(feature = "ml")]
87mod tesseract;
88pub mod textparse;
89#[cfg(feature = "ocr-prep")]
90pub mod tf_core;
91// docling's TableFormer cell matcher — pure Rust, shared with the browser
92// TableFormer path (#157 stage 3).
93#[cfg(feature = "ocr-prep")]
94pub mod tf_match;
95pub mod timing;
96
97#[cfg(feature = "ml")]
98use std::collections::BTreeMap;
99use std::fmt;
100#[cfg(feature = "ml")]
101use std::sync::mpsc::{sync_channel, Receiver};
102#[cfg(feature = "ml")]
103use std::sync::{Arc, Mutex};
104
105// An execution provider only exists on its OS, and ort's prebuilt ONNX
106// Runtime binaries follow suit — requesting an impossible pairing otherwise
107// surfaces as a cryptic ort-sys linker error ("no builds available that
108// satisfy the requested feature set"). Catch it at type-check time with an
109// actionable message instead.
110#[cfg(all(feature = "coreml", not(target_vendor = "apple")))]
111compile_error!(
112 "the `coreml` execution provider exists only on Apple targets (macOS/iOS). \
113 On Linux use `--features cuda` or `--features tensorrt` (NVIDIA), on \
114 Windows also `--features directml`, or build without EP features for CPU."
115);
116#[cfg(all(feature = "directml", not(target_os = "windows")))]
117compile_error!(
118 "the `directml` execution provider exists only on Windows. On Linux use \
119 `--features cuda` or `--features tensorrt` (NVIDIA), on macOS \
120 `--features coreml`, or build without EP features for CPU."
121);
122#[cfg(all(any(feature = "cuda", feature = "tensorrt"), target_vendor = "apple"))]
123compile_error!(
124 "the `cuda`/`tensorrt` execution providers have no Apple builds (no NVIDIA \
125 support on macOS). Use `--features coreml` there, or build without EP \
126 features for CPU."
127);
128
129use docling_core::DoclingDocument;
130pub use docling_core::EncryptionError;
131// The env-knob helpers only gate ML-pipeline diagnostics and tuning; the
132// pure text-layer (wasm) build has no call sites.
133#[cfg(feature = "ml")]
134use docling_core::Node;
135#[cfg(feature = "ml")]
136use docling_core::{debug_log, env};
137
138pub use dp_lines::WordGap;
139pub use heading_hierarchy::HeadingHierarchyOptions;
140#[cfg(feature = "ml")]
141pub use mets::{convert_mets_gbs, convert_mets_gbs_with_options, convert_mets_gbs_with_pipeline};
142#[cfg(feature = "ml")]
143pub use ocr::{OcrEngine, OcrLang, OcrMode};
144#[cfg(feature = "render")]
145pub use pdfium_backend::PdfDocument;
146pub use pdfium_backend::TextOptions;
147pub use pdfium_backend::{PdfPage, TextCell};
148#[cfg(feature = "ml")]
149pub use tesseract::{
150 installed_languages as tesseract_languages, lang_arg as tesseract_lang_arg, TesseractOptions,
151};
152// Plain page rasterization (#243) — pdfium only, no models.
153#[cfg(feature = "render")]
154pub use pdfium_backend::{render_pages, RenderedPage};
155#[cfg(feature = "ml")]
156pub use render::Omission;
157
158/// Errors from the PDF backend. Detailed and surfaced (never silently skipped).
159#[derive(Debug)]
160pub enum PdfError {
161 /// pdfium failed to bind, open, or read the document (and, historically,
162 /// any pipeline error outside the models).
163 Pdfium(String),
164 /// The document itself: an object model lopdf cannot read, a page no
165 /// renderer can draw, a page range outside it.
166 Document(String),
167 /// The layout ONNX model failed to load or run.
168 Layout(String),
169 /// The OCR ONNX model failed to load or run.
170 Ocr(String),
171 /// The document budget ([`Pipeline::document_timeout`]) ran out between
172 /// two pages. Internally the sentinel that stops the page walk; a caller
173 /// never sees it from [`Pipeline::convert`] — the conversion returns the
174 /// pages it finished and reports the cut through [`Completion`].
175 Timeout(String),
176 /// The document is encrypted and the password given (or none) does not
177 /// open it (#636) — the typed value a caller prompts for a password on,
178 /// reachable through [`std::error::Error::source`] too. Its text is the
179 /// one docling raises.
180 Encrypted(EncryptionError),
181}
182
183impl fmt::Display for PdfError {
184 fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
185 match self {
186 PdfError::Pdfium(m) => write!(f, "pdf: pdfium error: {m}"),
187 PdfError::Document(m) => write!(f, "pdf: {m}"),
188 PdfError::Layout(m) => write!(f, "pdf: {m}"),
189 PdfError::Ocr(m) => write!(f, "pdf: {m}"),
190 PdfError::Timeout(m) => write!(f, "pdf: {m}"),
191 PdfError::Encrypted(EncryptionError::WrongPassword) => {
192 f.write_str("pdf: the PDF is encrypted and the password is wrong")
193 }
194 PdfError::Encrypted(EncryptionError::NeedPassword) => {
195 f.write_str("pdf: the PDF is encrypted: a password is required")
196 }
197 PdfError::Encrypted(e) => write!(f, "pdf: the PDF is encrypted: {e}"),
198 }
199 }
200}
201
202/// How a conversion ended: every selected page, or the pages that fit the
203/// document budget (docling's `document_timeout`, `PARTIAL_SUCCESS`).
204#[derive(Debug, Clone, Copy, PartialEq, Eq)]
205pub enum Completion {
206 /// Every selected page was processed.
207 Complete,
208 /// The budget ran out: `pages_done` of `pages_selected` pages are in the
209 /// document, the rest were never rendered (or never processed).
210 TimedOut {
211 pages_done: usize,
212 pages_selected: usize,
213 budget: std::time::Duration,
214 },
215}
216
217/// One finished page, as the per-page progress hook
218/// ([`Pipeline::set_progress`]) reports it — docling's `PageCompletedProgress`
219/// minus the document fields the caller adds (#783).
220#[derive(Debug, Clone, Copy, PartialEq, Eq)]
221pub struct PageProgress {
222 /// The page's 1-based number in the document: a page window keeps the
223 /// source numbering, so a `2-3` window reports pages 2 and 3.
224 pub page_no: usize,
225 /// Pages finished so far in this conversion, this one included. On the
226 /// parallel walks pages finish out of order, so this — not `page_no` —
227 /// drives a progress bar.
228 pub pages_done: usize,
229 /// Pages the conversion selected: the window's size, else the whole
230 /// document.
231 pub pages_selected: usize,
232}
233
234/// The per-page progress hook: shared with the parallel workers, so it must
235/// be cheap and thread-safe.
236pub type PageProgressHook = std::sync::Arc<dyn Fn(PageProgress) + Send + Sync>;
237
238impl Completion {
239 pub fn timed_out(&self) -> bool {
240 matches!(self, Completion::TimedOut { .. })
241 }
242
243 /// docling's `ErrorItem.error_message` for the cut, `None` when complete.
244 pub fn message(&self) -> Option<String> {
245 match self {
246 Completion::Complete => None,
247 Completion::TimedOut {
248 pages_done,
249 pages_selected,
250 budget,
251 } => Some(format!(
252 "document timeout of {:.3}s exceeded after {pages_done} of {pages_selected} \
253 pages; the output holds the pages processed",
254 budget.as_secs_f64()
255 )),
256 }
257 }
258}
259
260/// A converted document and how its conversion ended ([`Completion`]).
261#[derive(Debug, Clone)]
262pub struct Converted {
263 pub document: DoclingDocument,
264 pub completion: Completion,
265}
266
267/// Has the document budget run out?
268fn expired(deadline: Option<std::time::Instant>) -> bool {
269 deadline.is_some_and(|d| std::time::Instant::now() >= d)
270}
271
272/// The sentinel the page walks stop on once the budget is gone.
273fn timeout_sentinel() -> PdfError {
274 PdfError::Timeout("document timeout exceeded".into())
275}
276
277impl std::error::Error for PdfError {
278 fn source(&self) -> Option<&(dyn std::error::Error + 'static)> {
279 match self {
280 PdfError::Encrypted(e) => Some(e),
281 _ => None,
282 }
283 }
284}
285
286#[cfg(feature = "pdfium")]
287impl From<pdfium_render::prelude::PdfiumError> for PdfError {
288 fn from(e: pdfium_render::prelude::PdfiumError) -> Self {
289 // A failed dlopen means the library this build was asked to use
290 // (`pdfium` feature) is not there. Say what to do instead of leaking
291 // the raw loader error.
292 if matches!(e, pdfium_render::prelude::PdfiumError::LoadLibraryError(_)) {
293 // The loader error pretty-prints over several lines; compact it.
294 let detail = e
295 .to_string()
296 .split_whitespace()
297 .collect::<Vec<_>>()
298 .join(" ");
299 return PdfError::Pdfium(format!(
300 "the pdfium library is not installed. This build's `pdfium` feature \
301 was asked for it (DOCLING_RS_RENDERER=pdfium, or a file the pure-Rust \
302 object model cannot read): point PDFIUM_DYNAMIC_LIB_PATH at a directory \
303 containing the pdfium library, or unset DOCLING_RS_RENDERER — the \
304 default renderer needs no library. Declarative formats (DOCX, HTML, \
305 Markdown, …) never need it. [{detail}]"
306 ));
307 }
308 PdfError::Pdfium(e.to_string())
309 }
310}
311
312/// Convert a PDF's **embedded text layer only** — no ONNX, no
313/// threads: the pure-Rust content-stream parser ([`textparse`]) feeds the same
314/// orphan-region assembly the `no_ocr` pipeline flag uses, so text-layer PDFs
315/// come out identical to `--text-layer-only` (flat, line-grouped paragraphs in reading
316/// order; no headings/lists/tables/pictures, and no hyperlink recovery).
317///
318/// This is the only conversion entry compiled without the `ml` feature (it is
319/// what a wasm32 build runs). A scanned/image-only PDF (no embedded text
320/// layer) yields an empty document rather than an error, same as `no_ocr` —
321/// callers can detect that and fall back to an OCR-capable build.
322pub fn convert_text_layer(bytes: &[u8], name: &str) -> Result<DoclingDocument, PdfError> {
323 convert_text_layer_pages(bytes, name, None)
324}
325
326/// The error for bytes no reader here can parse as a PDF — not a PDF at
327/// all, or one damaged past the parser's repairs. Said about the file, not
328/// about a missing renderer or text layer: no library would read it either.
329pub(crate) const UNREADABLE: &str =
330 "not a readable PDF: the file is damaged or not a PDF (its structure could not be parsed)";
331
332/// [`convert_text_layer`] restricted to a **1-based inclusive** page window
333/// (issue #80's `--pages`); `None` converts everything. The window is
334/// validated the same way as [`Pipeline::pages`]: `first <= last`, 1-based,
335/// and it must select at least one existing page.
336pub fn convert_text_layer_pages(
337 bytes: &[u8],
338 name: &str,
339 pages: Option<(usize, usize)>,
340) -> Result<DoclingDocument, PdfError> {
341 convert_text_layer_pages_with(bytes, name, pages, None)
342}
343
344/// [`convert_text_layer_pages`] with the word-gap rule of the text layer's
345/// line and word contraction (#921, [`WordGap`]); `None` = the process
346/// default (`DOCLING_RS_PDF_WORD_GAP`, else [`WordGap::Adaptive`]).
347pub fn convert_text_layer_pages_with(
348 bytes: &[u8],
349 name: &str,
350 pages: Option<(usize, usize)>,
351 word_gap: Option<WordGap>,
352) -> Result<DoclingDocument, PdfError> {
353 convert_text_layer_pages_opts(
354 bytes,
355 name,
356 pages,
357 TextOptions {
358 word_gap,
359 ..TextOptions::from_env()
360 },
361 )
362}
363
364/// [`convert_text_layer_pages`] with every text-layer setting (#921): the
365/// word-gap rule and whether typographic punctuation is kept as drawn.
366pub fn convert_text_layer_pages_opts(
367 bytes: &[u8],
368 name: &str,
369 pages: Option<(usize, usize)>,
370 text: TextOptions,
371) -> Result<DoclingDocument, PdfError> {
372 if let Some((first, last)) = pages {
373 if first == 0 || last < first {
374 return Err(PdfError::Pdfium(format!(
375 "invalid page range {first}-{last} (pages are 1-based, first <= last)"
376 )));
377 }
378 }
379 let mut doc = DoclingDocument::new(name);
380 // A file nothing can open is an error, not a PDF that merely has no
381 // text layer (the caller's "scanned? needs OCR" hint).
382 let Some(mut parser) = textparse::PageTextParser::open(bytes) else {
383 return Err(PdfError::Document(UNREADABLE.into()));
384 };
385 parser.set_word_gap(text.word_gap);
386 parser.set_keep_typography(Some(text.keep_typography));
387 // Pages are parsed, assembled and dropped one at a time: holding every
388 // page's cells until the end made the peak scale with the whole file —
389 // and a page window still paid for every page outside it.
390 let total = parser.page_count();
391 let mut tally = textparse::TextLayerTally::default();
392 for i in 0..total {
393 let selected = pages.is_none_or(|(first, last)| (first..=last).contains(&(i + 1)));
394 // A page outside the window only feeds the vestigial verdict below,
395 // so it is parsed only until that verdict is settled (on any real
396 // text layer, by the first page or two).
397 if !selected && tally.proves_text(total) {
398 continue;
399 }
400 let page = parser.text_page(i);
401 tally.add(&page);
402 if !selected {
403 continue;
404 }
405 let mut regions = Vec::new();
406 assemble::add_orphan_regions(&mut regions, &page.cells);
407 let table_rows = vec![None; regions.len()];
408 let enrich_out = vec![None; regions.len()];
409 let (mut nodes, links) =
410 assemble::assemble_page(&page, regions, &table_rows, &enrich_out, None);
411 assemble::stamp_page_no(&mut nodes, i + 1);
412 doc.nodes.extend(nodes);
413 doc.links.extend(links);
414 }
415 // A vestigial layer (a few typed-in form fields over scanned pages) is not
416 // the document's text: return the empty document, which callers already
417 // report as "no text layer" — so an OCR-capable caller falls back to OCR
418 // instead of proudly extracting thirteen characters.
419 if tally.is_vestigial() {
420 return Ok(DoclingDocument::new(name));
421 }
422 if let Some((first, last)) = pages {
423 if first > total {
424 return Err(PdfError::Pdfium(format!(
425 "page range {first}-{last} is outside the document ({total} page(s))"
426 )));
427 }
428 }
429 assemble::merge_continuations(&mut doc.nodes);
430 // Grown a page at a time, the node vector can be up to half spare
431 // capacity — hundreds of MB on a long document, held for its lifetime.
432 doc.nodes.shrink_to_fit();
433 Ok(doc)
434}
435
436/// Whether a PDF's embedded text layer is the document's text — the verdict
437/// behind [`Pipeline::text_layer_auto`]: `true` when the pages carry real
438/// text, `false` for a scanned / image-only file or a vestigial layer (a few
439/// typed-in form fields over scans; [`textparse::TextLayerTally`]'s rule,
440/// the one [`convert_text_layer`] answers with an empty document). Pure
441/// Rust, no render, no model: the pages are parsed only until the verdict
442/// is settled — on a digital document, by the first page or two; a scanned
443/// one is walked to the end, which is cheap (its content streams hold no
444/// text). Every page counts, not just a page window: the window is the
445/// caller's output selection, the document's nature is the whole file's.
446/// An unreadable file is an error, as on every other path.
447pub fn text_layer_has_text(bytes: &[u8], password: Option<&str>) -> Result<bool, PdfError> {
448 let Some(mut parser) = textparse::PageTextParser::open_with_password(bytes, password) else {
449 return Err(PdfError::Document(UNREADABLE.into()));
450 };
451 let total = parser.page_count();
452 let mut tally = textparse::TextLayerTally::default();
453 for i in 0..total {
454 if tally.proves_text(total) {
455 return Ok(true);
456 }
457 tally.add(&parser.text_page(i));
458 }
459 Ok(!tally.is_vestigial())
460}
461
462/// Threads ONNX inference may use, capped by `DOCLING_RS_PDF_THREADS` if set.
463/// Defaults to the available parallelism (ort otherwise picks a low number).
464#[cfg(feature = "ml")]
465pub(crate) fn intra_threads() -> usize {
466 if let Some(n) = env::parse::<usize>("DOCLING_RS_PDF_THREADS").filter(|&n| n > 0) {
467 return n;
468 }
469 env::cpu_budget()
470}
471
472#[cfg(feature = "ml")]
473/// TableFormer's intra-op width (#262): `DOCLING_RS_TF_INTRA` explicitly,
474/// else the shared [`intra_threads`] budget. The shared TF session used to
475/// take the raw host width on top of the already-sized worker pools —
476/// under a cgroup CPU limit that oversubscription showed up as constant
477/// throttling and ~66% higher peak memory (each intra thread carries its own
478/// arena slab); the reporter's 4-CPU/8-core case dropped from 2.2 GB to
479/// 1.3 GB peak by capping this pool.
480pub(crate) fn tf_intra() -> usize {
481 if let Some(n) = env::parse::<usize>("DOCLING_RS_TF_INTRA").filter(|&n| n > 0) {
482 return n;
483 }
484 intra_threads()
485}
486
487#[cfg(feature = "ml")]
488/// True when `DOCLING_RS_FP32` forces the full-precision models even where
489/// an INT8 variant sits next to the fp32 default.
490pub(crate) fn fp32_forced() -> bool {
491 env::flag("DOCLING_RS_FP32")
492}
493
494#[cfg(feature = "ml")]
495/// Should the int8 model defaults be skipped in favor of fp32? Either the
496/// user said so (`DOCLING_RS_FP32`), or a GPU execution provider is selected
497/// (#74) — the int8 exports are QDQ graphs calibrated for CPU kernels and
498/// only conformance-validated there. An explicit `DOCLING_*_ONNX` path
499/// override still wins over this at every call site.
500pub(crate) fn prefer_fp32() -> bool {
501 fp32_forced() || docling_onnx::prefers_fp32()
502}
503
504#[cfg(feature = "render")]
505/// Resolve a default (CWD-relative) asset path — the shared chain in
506/// [`docling_core::assets`]: CWD, then next to the executable and one level
507/// above it (the `scripts/install/install.sh` layout).
508pub(crate) fn resolve_asset(rel: &str) -> String {
509 docling_core::assets::resolve(rel)
510}
511
512/// One resolved runtime asset — which file a stage would load right now,
513/// given the CWD, the env overrides and the int8/fp32 preference.
514#[cfg(feature = "ml")]
515#[derive(Debug, Clone)]
516pub struct ModelEntry {
517 /// Pipeline stage, e.g. `layout`, `tableformer.decoder`, `ocr.rec`.
518 pub stage: &'static str,
519 /// The resolved path (absolute or CWD-relative, as it will be opened).
520 pub path: String,
521 /// Whether the file exists right now.
522 pub found: bool,
523 /// File size in bytes (0 when missing) — enough to tell an int8 quant
524 /// from an fp32 graph, or a stale model from a re-published one, at a
525 /// glance without hashing gigabytes per request.
526 pub bytes: u64,
527}
528
529/// Resolve the whole runtime model set **without loading anything** — the
530/// exact selection each stage performs at load time (layout honors the
531/// int8/fp32 preference, TableFormer its decoder ranking, OCR the language
532/// pair), plus the pdfium library. docling-serve exposes this at
533/// `/v1/config` and logs it at startup, so "the server picked up different
534/// models" is one `curl` away instead of a mystery of dissolved tables.
535/// Resolution is CWD-relative with an exe-dir fallback, so the answer can
536/// legitimately differ between two working directories.
537#[cfg(feature = "ml")]
538pub fn model_inventory() -> Vec<ModelEntry> {
539 fn entry(stage: &'static str, path: String) -> ModelEntry {
540 let meta = std::fs::metadata(&path).ok();
541 ModelEntry {
542 stage,
543 found: meta.is_some(),
544 bytes: meta.map(|m| m.len()).unwrap_or(0),
545 path,
546 }
547 }
548 let (enc, dec, bbx) = tableformer::resolved_paths();
549 let (rec, dict) = ocr::resolve_rec_pair(ocr::OcrLang::from_env());
550 let pdfium =
551 env::nonempty("PDFIUM_DYNAMIC_LIB_PATH").unwrap_or_else(|| resolve_asset(".pdfium/lib"));
552 vec![
553 entry(
554 "layout",
555 model_path(
556 "DOCLING_LAYOUT_ONNX",
557 ".models/layout_heron.onnx",
558 ".models/layout_heron_int8.onnx",
559 ),
560 ),
561 entry("tableformer.encoder", enc),
562 entry("tableformer.decoder", dec),
563 entry("tableformer.bbox", bbx),
564 entry("ocr.rec", rec),
565 entry("ocr.dict", dict),
566 entry("ocr.textline_ori", ocr_cls::resolve_cls_path()),
567 entry("pdfium", pdfium),
568 ]
569}
570
571/// Resolve a model path: an explicit env override always wins; otherwise the
572/// INT8 variant of the default path when it exists on disk (the quantized
573/// models are conformance-validated — see docs/PDF_CONFORMANCE.md — and load/run
574/// markedly faster on CPU), unless `DOCLING_RS_FP32` opts back into full
575/// precision; else the fp32 default.
576#[cfg(feature = "ml")]
577pub(crate) fn model_path(key: &str, fp32_default: &str, int8_default: &str) -> String {
578 if let Some(p) = env::nonempty(key) {
579 return p;
580 }
581 if !prefer_fp32() {
582 let p = resolve_asset(int8_default);
583 if std::path::Path::new(&p).exists() {
584 return p;
585 }
586 }
587 resolve_asset(fp32_default)
588}
589
590/// Decode a standalone image with hard resource limits. A crafted image can
591/// declare enormous dimensions in a few-KB file; `image::load_from_memory`
592/// then tries to allocate the full pixel buffer (e.g. 60000×60000 → ~10 GB),
593/// and allocation failure aborts the whole process, bypassing the per-request
594/// panic catch. The 256 MiB alloc / 30000-px caps below turn that into a
595/// recoverable decode error instead. `DOCLING_RS_MAX_IMAGE_PIXELS` overrides
596/// the per-side pixel cap for the rare legitimately-huge scan.
597///
598/// Gated on `ml`: the only callers (`convert_image`, the METS backend) are
599/// ML-only, and the `image` crate is an `ml`-feature dependency — the
600/// text-layer wasm build has neither.
601#[cfg(feature = "ml")]
602pub(crate) fn decode_image_limited(bytes: &[u8]) -> Result<image::RgbImage, PdfError> {
603 let max_side: u32 = env::parse("DOCLING_RS_MAX_IMAGE_PIXELS").unwrap_or(30_000);
604 decode_image_with_max_side(bytes, max_side)
605}
606
607/// Whether `bytes` is an ISOBMFF HEIF/HEIC container (the `ftyp` brands
608/// iPhones write). Checked by content, not extension — HEIC regularly
609/// arrives misnamed `.jpg`.
610#[cfg(feature = "ml")]
611fn is_heif(bytes: &[u8]) -> bool {
612 bytes.len() >= 12
613 && &bytes[4..8] == b"ftyp"
614 && matches!(
615 &bytes[8..12],
616 b"heic" | b"heix" | b"hevc" | b"heim" | b"heis" | b"hevm" | b"hevs" | b"mif1" | b"msf1"
617 )
618}
619
620/// Decode a HEIF/HEIC primary image via libheif (#211). Behind the opt-in
621/// `heif` feature — libheif is a native dependency the default build (and
622/// wasm) must not carry.
623#[cfg(all(feature = "ml", feature = "heif"))]
624fn decode_heif(bytes: &[u8], max_side: u32) -> Result<image::RgbImage, PdfError> {
625 use libheif_rs::{ColorSpace, HeifContext, LibHeif, RgbChroma};
626 let err = |e: String| PdfError::Pdfium(format!("heif: {e}"));
627 let ctx = HeifContext::read_from_bytes(bytes).map_err(|e| err(e.to_string()))?;
628 let handle = ctx.primary_image_handle().map_err(|e| err(e.to_string()))?;
629 if handle.width() > max_side || handle.height() > max_side {
630 return Err(err(format!(
631 "image dimensions {}x{} exceed the {max_side}px per-side cap \
632 (DOCLING_RS_MAX_IMAGE_PIXELS overrides)",
633 handle.width(),
634 handle.height()
635 )));
636 }
637 let lib = LibHeif::new();
638 let img = lib
639 .decode(&handle, ColorSpace::Rgb(RgbChroma::Rgb), None)
640 .map_err(|e| err(e.to_string()))?;
641 let (w, h) = (img.width(), img.height());
642 let planes = img.planes();
643 let plane = planes
644 .interleaved
645 .ok_or_else(|| err("no RGB plane".into()))?;
646 let stride = plane.stride;
647 let mut out = image::RgbImage::new(w, h);
648 for (y, row) in out.rows_mut().enumerate() {
649 let src = &plane.data[y * stride..y * stride + w as usize * 3];
650 for (x, px) in row.enumerate() {
651 px.0 = [src[x * 3], src[x * 3 + 1], src[x * 3 + 2]];
652 }
653 }
654 Ok(out)
655}
656
657#[cfg(feature = "ml")]
658fn decode_image_with_max_side(bytes: &[u8], max_side: u32) -> Result<image::RgbImage, PdfError> {
659 use image::{ImageDecoder, ImageReader};
660 use std::io::Cursor;
661
662 if is_heif(bytes) {
663 #[cfg(feature = "heif")]
664 return decode_heif(bytes, max_side);
665 #[cfg(not(feature = "heif"))]
666 return Err(PdfError::Pdfium(
667 "HEIC/HEIF input needs a build with the `heif` cargo feature \
668 (rebuild with --features heif; links the system libheif)"
669 .into(),
670 ));
671 }
672
673 let mut limits = image::Limits::default();
674 limits.max_image_width = Some(max_side);
675 limits.max_image_height = Some(max_side);
676 limits.max_alloc = Some(256 * 1024 * 1024);
677
678 let mut reader = ImageReader::new(Cursor::new(bytes))
679 .with_guessed_format()
680 .map_err(|e| PdfError::Pdfium(format!("image: {e}")))?;
681 reader.limits(limits);
682 let mut decoder = reader
683 .into_decoder()
684 .map_err(|e| PdfError::Pdfium(format!("image: {e}")))?;
685 // docling#4247 (2.128, `ImageOps.exif_transpose`): a camera stores the
686 // sensor readout plus an orientation tag rather than rotated pixels, so
687 // a portrait photo would reach layout and OCR on its side unless the tag
688 // is honoured. An unreadable tag is no orientation, not an error.
689 let orientation = decoder
690 .orientation()
691 .unwrap_or(image::metadata::Orientation::NoTransforms);
692 let mut img = image::DynamicImage::from_decoder(decoder)
693 .map_err(|e| PdfError::Pdfium(format!("image: {e}")))?;
694 img.apply_orientation(orientation);
695 Ok(img.into_rgb8())
696}
697
698/// Image outputs of the PDF/image pipeline (#519/#520): the scale picture
699/// crops are delivered at and whether each page's render is kept as the
700/// document's page image. See [`Pipeline::images_scale`] /
701/// [`Pipeline::generate_page_images`].
702#[derive(Debug, Clone, Copy, Default, PartialEq)]
703pub struct ImageOutput {
704 /// Pixels per PDF point for picture crops and page images — docling's
705 /// `images_scale`. `None` keeps the pipeline's own page render (2.0
706 /// px/pt), resampling nothing.
707 pub scale: Option<f32>,
708 /// Keep every page's render as the document's page image — docling's
709 /// `generate_page_images`.
710 pub page_images: bool,
711}
712
713#[cfg(feature = "ml")]
714/// The layout regions of a page with a text layer that no programmatic text
715/// touches — docling's "clusters without any visible programmatic text"
716/// (`_find_pdf_aware_layout_ocr_rects`: any text cell intersecting the
717/// cluster counts), so their text, if any, is drawn as outlines (#931).
718/// Text, furniture and table regions only: pictures have their own pass.
719/// Furniture is relabeled `text`, the label region OCR reads.
720fn text_less_regions(regions: &[layout::Region], cells: &[TextCell]) -> Vec<layout::Region> {
721 regions
722 .iter()
723 .filter(|r| {
724 ocr_prep::is_text_label(r.label)
725 || assemble::is_table_like(r.label)
726 || matches!(r.label, "page_header" | "page_footer")
727 })
728 .filter(|r| {
729 !cells.iter().any(|c| {
730 !c.text.trim().is_empty()
731 && c.r.min(r.r) > c.l.max(r.l)
732 && c.b.min(r.b) > c.t.max(r.t)
733 })
734 })
735 .map(|r| layout::Region {
736 label: match r.label {
737 "page_header" | "page_footer" => "text",
738 l => l,
739 },
740 ..r.clone()
741 })
742 .collect()
743}
744
745#[cfg(feature = "ml")]
746/// The OCR'd cells with their recognition confidence, for the per-cell report
747/// (#915): a copy of what feeds the page `ocr_score`, kept beside it.
748fn cell_confs(scored: &[(TextCell, f32)]) -> Vec<docling_core::OcrCellConfidence> {
749 scored
750 .iter()
751 .map(|(c, conf)| docling_core::OcrCellConfidence {
752 text: c.text.clone(),
753 l: c.l,
754 t: c.t,
755 r: c.r,
756 b: c.b,
757 confidence: *conf,
758 })
759 .collect()
760}
761
762#[cfg(feature = "ml")]
763/// The document's report: page scores plus the per-cell OCR confidences (#915).
764fn confidence_report(
765 confs: std::collections::BTreeMap<usize, docling_core::confidence::PageConfidence>,
766 cells: std::collections::BTreeMap<usize, Vec<docling_core::OcrCellConfidence>>,
767) -> docling_core::ConfidenceReport {
768 cells.into_iter().fold(
769 docling_core::ConfidenceReport::from_pages(confs),
770 |report, (n, cells)| report.with_ocr_cells(n, cells),
771 )
772}
773
774#[cfg(feature = "ml")]
775/// One page's assembled output: typed nodes plus the page's hyperlinks (kept
776/// separate so pages processed out of order can be stitched back in page
777/// order) and its confidence scores (#183).
778type PageOut = (
779 Vec<Node>,
780 Vec<(String, String)>,
781 docling_core::confidence::PageConfidence,
782 // The page image, when page images are requested (#520).
783 Option<docling_core::PictureImage>,
784 // The page's OCR cells with their recognition confidence (#915).
785 Vec<docling_core::OcrCellConfidence>,
786);
787
788/// One page from its text layer alone — the fastest path: no layout, OCR
789/// or TableFormer inference at all. The embedded text cells (if any) become
790/// flat, line-grouped paragraphs in reading order via the same
791/// orphan-region machinery that normally rescues text the detector missed —
792/// here it rescues *all* of it. A page with no text layer (scanned /
793/// image-only) yields nothing. Shared by the `no_ocr` worker branch
794/// (images, METS) and the worker-less PDF loop of
795/// [`Pipeline::convert_text_only`], so `text_layer_only=true` and an
796/// `auto` document with text come out byte-identical.
797#[cfg(feature = "ml")]
798fn text_only_page(
799 page: &PdfPage,
800 images_scale: Option<f32>,
801) -> (
802 Vec<Node>,
803 Vec<(String, String)>,
804 docling_core::confidence::PageConfidence,
805) {
806 let parse = quality::parse_score(&page.cells);
807 let mut regions = Vec::new();
808 assemble::add_orphan_regions(&mut regions, &page.cells);
809 let table_rows = vec![None; regions.len()];
810 let enrich_out = vec![None; regions.len()];
811 let conf = quality::page_confidence(parse, ®ions, &[]);
812 let (nodes, links) = timing::timed("assemble_page", || {
813 assemble::assemble_page(page, regions, &table_rows, &enrich_out, images_scale)
814 });
815 (nodes, links, conf)
816}
817
818/// The page's image for the document's page images (#520) under `images`:
819/// `None` unless page images were requested, or when the page has no bitmap
820/// (the text-layer-only paths never render one).
821#[cfg(feature = "ml")]
822fn page_image_under(
823 images: ImageOutput,
824 n: usize,
825 page: &PdfPage,
826) -> Option<docling_core::PictureImage> {
827 if !images.page_images {
828 return None;
829 }
830 let image = assemble::page_image(page, images.scale);
831 if image.is_none() {
832 docling_core::debug_log!("page {}: no page bitmap for the page image", n + 1);
833 }
834 image
835}
836
837#[cfg(feature = "ml")]
838/// A page between region resolution and TableFormer: what `prepare_page`
839/// produced and `complete_page` still needs, so a pool worker can park the
840/// page while another worker holds the shared TableFormer (see `Staged`).
841struct Prepared {
842 regions: Vec<layout::Region>,
843 ocr_confs: Vec<f32>,
844 ocr_cells: Vec<docling_core::OcrCellConfidence>,
845 parse: Option<f64>,
846}
847
848#[cfg(feature = "ml")]
849/// A pool worker's per-page outcome. `NeedsTables` is a page whose only
850/// remaining stage is the shared TableFormer, which was busy when the worker
851/// got there: rather than block on the mutex — one worker idle for the
852/// whole of another worker's table decode, ~9.5 s of the 60-page .NET slice
853/// on a 2-worker pool — the worker keeps the page aside, pulls the next one
854/// off the render channel, and comes back once the slot is free. Results
855/// are reassembled by page index anyway, so completion order is free.
856enum Staged {
857 Done(PageOut),
858 NeedsTables(Prepared),
859}
860
861#[cfg(feature = "ml")]
862/// How many pages a pool worker keeps parked on the TableFormer before it
863/// falls back to waiting: each carries its ~5 MB bitmaps, so this bounds the
864/// extra residency to two pages per worker on top of the render channel.
865const MAX_DEFERRED_PAGES: usize = 2;
866
867#[cfg(feature = "ml")]
868/// The pool-wide TableFormer slot: one instance shared by every worker, loaded
869/// lazily on the first table region any worker sees. Tables appear on a
870/// minority of pages, so per-worker copies mostly multiplied ~0.4 GB of
871/// weights+arenas by the pool size for nothing; a single shared instance keeps
872/// the peak flat regardless of pool width, and a table's structure prediction
873/// is independent of which worker runs it, so output is byte-identical. The
874/// mutex serialises concurrent tables — the shared instance is loaded with the
875/// full intra-op thread budget to compensate (one wide TableFormer instead of
876/// several narrow ones).
877enum TfSlot {
878 /// Not attempted yet (no table seen so far).
879 Unloaded,
880 /// Load attempted, graphs absent — geometric fallback (warned once).
881 Missing,
882 Ready(tableformer::TableFormer),
883}
884
885#[cfg(feature = "ml")]
886impl TfSlot {
887 /// Load TableFormer on the first call: `Ready`, or `Missing` when its
888 /// graphs are absent (the geometric fallback — degradation, not an
889 /// error). Later calls are no-ops either way.
890 fn load(&mut self) {
891 if matches!(self, TfSlot::Unloaded) {
892 *self = match tableformer::TableFormer::load_with(tf_intra()) {
893 Some(tf) => TfSlot::Ready(tf),
894 None => TfSlot::Missing,
895 };
896 }
897 }
898}
899
900#[cfg(feature = "ml")]
901type SharedTables = Arc<Mutex<TfSlot>>;
902
903#[cfg(feature = "ml")]
904/// The same lazy shared-slot pattern for the (rarer still) enrichment models:
905/// one instance per pipeline, loaded on the first region that needs it.
906enum EnrichSlot<T> {
907 Unloaded,
908 /// Load attempted, model files absent — enrichment skipped (warned once).
909 Missing,
910 Ready(T),
911}
912
913#[cfg(feature = "ml")]
914type SharedClassifier = Arc<Mutex<EnrichSlot<enrich::PictureClassifier>>>;
915#[cfg(feature = "ml")]
916type SharedCodeFormula = Arc<Mutex<EnrichSlot<enrich::CodeFormula>>>;
917
918#[cfg(feature = "ml")]
919/// The opt-in enrichment passes, mirroring docling's `PdfPipelineOptions`
920/// flags (`do_picture_classification`, `do_code_enrichment`,
921/// `do_formula_enrichment`). All off by default.
922#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
923pub struct EnrichmentOptions {
924 /// Classify each picture with DocumentFigureClassifier (26 classes).
925 pub picture_classification: bool,
926 /// Rewrite code blocks (and detect their language) with CodeFormulaV2.
927 pub code: bool,
928 /// Decode display formulas to LaTeX with CodeFormulaV2.
929 pub formula: bool,
930}
931
932#[cfg(feature = "ml")]
933impl EnrichmentOptions {
934 fn any(&self) -> bool {
935 self.picture_classification || self.code || self.formula
936 }
937}
938
939#[cfg(feature = "ml")]
940/// The layout model's input for a page: the pypdfium2-exact scale-1.0 page
941/// image (#478) when the renderer produced one, else the legacy stretch of the 2×
942/// bitmap (browser / METS paths) — see [`layout::LayoutSrc`]. Public so the
943/// diagnostic examples feed [`layout::LayoutModel::predict`] the same input
944/// the pipeline does.
945pub fn layout_src(page: &PdfPage) -> layout::LayoutSrc<'_> {
946 match &page.image_layout {
947 Some(img) => layout::LayoutSrc::PageImage(img),
948 None => layout::LayoutSrc::Raw(&page.image),
949 }
950}
951
952#[cfg(feature = "ml")]
953/// RapidOCR's longest side (`Global.max_side_len`, 2000): docling renders the
954/// OCR input at `OcrOptions.scale` = 3 and RapidOCR then shrinks what exceeds
955/// this before detecting and recognizing, so that is the resolution its
956/// models actually see.
957pub(crate) const RAPIDOCR_MAX_SIDE: f32 = 2000.0;
958
959#[cfg(feature = "ml")]
960/// The px/pt the OCR reads a page at: the caller's `ocr_scale` when set; else,
961/// for a standalone bitmap input — its own scale-1.0 page, a resolution the
962/// pipeline did not choose — docling's rule (#570): three pixels per point
963/// (`OcrOptions.scale`), shrunk so the longer side stays within RapidOCR's
964/// 2000 px (a 754 × 1000 scanned form reads at 2.0; a 3000 px photo at 0.67,
965/// as RapidOCR would downsample it). `None` for a rendered page (PDF, hOCR),
966/// which keeps its own render — the 2.0 px/pt the conformance baselines are
967/// pinned to. Measured on FUNSD at 1.0 / 2.0 / 3.0 px/pt: 0.825 / 0.856 /
968/// 0.836 word recall — the recognizer likes its crops at the resolution
969/// RapidOCR hands it, no more.
970fn page_ocr_scale(ocr_scale: Option<f32>, width: f32, height: f32, scale: f32) -> Option<f32> {
971 if ocr_scale.is_some() || scale != 1.0 {
972 return ocr_scale;
973 }
974 let longest = width.max(height);
975 if longest <= 0.0 {
976 return None;
977 }
978 let s = 3.0f32.min(RAPIDOCR_MAX_SIDE / longest);
979 ((s - 1.0).abs() > 1e-3).then_some(s)
980}
981
982#[cfg(feature = "ml")]
983/// The bitmap + px/pt scale the OCR reads (#254, docling#3877's
984/// `OcrOptions.scale`): the page's own render unless `ocr_scale` asks for a
985/// different resolution, where a PIL-bicubic resample of that render is built
986/// once per page (cached in `cache`) and shared by every OCR pass. Resampling
987/// — rather than a second native pdfium render — keeps one code path across
988/// PDF, image, and hOCR inputs and leaves the layout/TableFormer pixels (and
989/// with them the conformance baseline) untouched; the 144-dpi base render is
990/// itself supersampled down from 216 dpi, so an upscaled OCR view loses
991/// little against a native render.
992fn ocr_input<'a>(
993 cache: &'a mut Option<image::RgbImage>,
994 image: &'a image::RgbImage,
995 scale: f32,
996 ocr_scale: Option<f32>,
997) -> (&'a image::RgbImage, f32) {
998 match ocr_scale {
999 Some(s) if (s - scale).abs() > 1e-3 && image.width() > 1 => {
1000 let f = s / scale;
1001 let img = cache.get_or_insert_with(|| {
1002 let dw = ((image.width() as f32 * f).round() as u32).max(1);
1003 let dh = ((image.height() as f32 * f).round() as u32).max(1);
1004 resample::pil_resize(image, dw, dh, resample::PilFilter::Bicubic)
1005 });
1006 (img, s)
1007 }
1008 _ => (image, scale),
1009 }
1010}
1011
1012#[cfg(feature = "ml")]
1013/// A self-contained set of the per-page models (layout, OCR). Each parallel
1014/// page-worker owns its own `Worker` so inference runs concurrently without
1015/// sharing an ONNX session (`ort`'s `Session::run` is `&mut self`); only the
1016/// rarely-hit TableFormer is shared (see [`TfSlot`]).
1017struct Worker {
1018 /// `None` when `no_ocr` skips layout entirely — no model load, no inference.
1019 layout: Option<layout::LayoutModel>,
1020 ocr: OcrSlot,
1021 det: DetSlot,
1022 /// Text-detector results computed ahead of a page's OCR pass (#429), keyed
1023 /// by page index: the detector reads nothing but the page image, so it
1024 /// runs on its own thread while the layout model predicts — see
1025 /// [`Self::detect_alongside`] — and the OCR block collects the boxes here
1026 /// instead of paying for detection serially.
1027 pending_det: BTreeMap<usize, Result<Vec<ocr_det::DetBox>, String>>,
1028 /// This worker's intra-op thread budget — also the OCR lane count (see
1029 /// [`ocr::OcrModel::load_with`]): a pool worker with two threads runs two
1030 /// single-thread recognisers, the primary as many as its cores.
1031 intra: usize,
1032 /// Shared TableFormer slot; `None` when `no_table_former`/`no_ocr` skip it.
1033 tables: Option<SharedTables>,
1034 /// Shared enrichment slots; `None` unless the corresponding flag is on.
1035 classifier: Option<SharedClassifier>,
1036 code_formula: Option<SharedCodeFormula>,
1037 enrich: EnrichmentOptions,
1038 /// Skip layout, OCR, and TableFormer; reconstruct text purely from the PDF's
1039 /// embedded text layer. See [`Pipeline::no_ocr`].
1040 no_ocr: bool,
1041 /// Discard the embedded text layer and OCR every page. See
1042 /// [`Pipeline::force_full_page_ocr`].
1043 force_full_page_ocr: bool,
1044 /// Keep text-panel pictures as pictures instead of demoting them to
1045 /// paragraphs. See [`Pipeline::no_text_panels`].
1046 no_text_panels: bool,
1047 /// Never run OCR, but keep layout + TableFormer (#244) — docling's
1048 /// `do_ocr=False`. See [`Pipeline::skip_ocr`].
1049 skip_ocr: bool,
1050 /// Which recognition model [`Self::ocr`] loads. See [`Pipeline::ocr_lang`].
1051 ocr_lang: ocr::OcrLang,
1052 /// Which engine [`Self::ocr`] is (#460). See [`Pipeline::ocr_engine`].
1053 ocr_engine: ocr::OcrEngine,
1054 /// Tesseract's `-l` argument (#460). See [`Pipeline::tesseract_lang`].
1055 tesseract_lang: Option<String>,
1056 /// OCR render scale override (px/pt, #254). See [`Pipeline::ocr_scale`].
1057 ocr_scale: Option<f32>,
1058 /// RapidOCR's line orientation handling (#666), handed to the PP-OCR
1059 /// recognizer. See [`Pipeline::ocr_textline_orientation`].
1060 textline_orientation: bool,
1061 /// Picture-crop scale and page images (#519/#520). See [`ImageOutput`].
1062 images: ImageOutput,
1063}
1064
1065#[cfg(feature = "ml")]
1066/// The worker's lazily-loaded OCR recognition model. `Missing` records a
1067/// failed load (#244: degradation over failure — a deployment without the OCR
1068/// model still gets layout + TableFormer, and OCR-dependent regions stay
1069/// empty) so the load isn't retried per page.
1070enum OcrSlot {
1071 Unloaded,
1072 Ready(Recognizer),
1073 Missing,
1074}
1075
1076#[cfg(feature = "ml")]
1077/// The loaded OCR engine (#460): PP-OCRv3 via ONNX Runtime or the
1078/// `tesseract` binary. Both answer the two questions the page assembly asks
1079/// — line cells for text regions, word cells for table regions — in page
1080/// points with a 0–1 confidence, so the call sites never know which runs.
1081pub(crate) enum Recognizer {
1082 PpOcr(ocr::OcrModel),
1083 Tesseract(tesseract::TesseractOcr),
1084}
1085
1086#[cfg(feature = "ml")]
1087impl Recognizer {
1088 /// See [`ocr::OcrModel::ocr_page_with`].
1089 fn ocr_page(
1090 &mut self,
1091 img: &image::RgbImage,
1092 regions: &[layout::Region],
1093 scale: f32,
1094 ) -> Result<Vec<(TextCell, f32)>, String> {
1095 self.ocr_page_with(img, regions, scale, None)
1096 }
1097
1098 /// [`Self::ocr_page`] with the detector's boxes (#570). Tesseract segments its
1099 /// own lines and ignores the detector's boxes.
1100 fn ocr_page_with(
1101 &mut self,
1102 img: &image::RgbImage,
1103 regions: &[layout::Region],
1104 scale: f32,
1105 detected: Option<&[ocr_det::DetBox]>,
1106 ) -> Result<Vec<(TextCell, f32)>, String> {
1107 match self {
1108 Self::PpOcr(m) => m.ocr_page_with(img, regions, scale, detected),
1109 Self::Tesseract(t) => t.ocr_page(img, regions, scale),
1110 }
1111 }
1112
1113 /// See [`ocr::OcrModel::ocr_table_words`].
1114 fn ocr_table_words(
1115 &mut self,
1116 img: &image::RgbImage,
1117 regions: &[layout::Region],
1118 scale: f32,
1119 detected: Option<&[ocr_det::DetBox]>,
1120 ) -> Result<Vec<(TextCell, f32)>, String> {
1121 match self {
1122 Self::PpOcr(m) => m.ocr_table_words(img, regions, scale, detected),
1123 Self::Tesseract(t) => t.ocr_table_words(img, regions, scale),
1124 }
1125 }
1126}
1127
1128#[cfg(feature = "ml")]
1129/// The text detector (#429), loaded on the first OCR'd page like the
1130/// recognizer; `Missing` keeps the pipeline recognition-only.
1131enum DetSlot {
1132 Unloaded,
1133 Ready(ocr_det::DetModel),
1134 Missing,
1135}
1136
1137#[cfg(feature = "ml")]
1138impl Worker {
1139 #[allow(clippy::too_many_arguments)] // mirrors the Pipeline's option set
1140 fn load(
1141 intra: usize,
1142 tables: Option<SharedTables>,
1143 enrich_slots: (Option<SharedClassifier>, Option<SharedCodeFormula>),
1144 enrich: EnrichmentOptions,
1145 no_ocr: bool,
1146 skip_ocr: bool,
1147 force_full_page_ocr: bool,
1148 no_text_panels: bool,
1149 ocr_lang: ocr::OcrLang,
1150 ocr_engine: ocr::OcrEngine,
1151 tesseract_lang: Option<String>,
1152 ocr_scale: Option<f32>,
1153 textline_orientation: bool,
1154 images: ImageOutput,
1155 ) -> Result<Self, PdfError> {
1156 Ok(Self {
1157 images,
1158 layout: if no_ocr {
1159 None
1160 } else {
1161 Some(layout::LayoutModel::load_with(intra).map_err(PdfError::Layout)?)
1162 },
1163 ocr: OcrSlot::Unloaded,
1164 det: DetSlot::Unloaded,
1165 pending_det: BTreeMap::new(),
1166 intra,
1167 tables,
1168 classifier: enrich_slots.0,
1169 code_formula: enrich_slots.1,
1170 enrich,
1171 no_ocr,
1172 skip_ocr,
1173 force_full_page_ocr,
1174 no_text_panels,
1175 ocr_lang,
1176 ocr_engine,
1177 tesseract_lang,
1178 ocr_scale,
1179 textline_orientation,
1180 })
1181 }
1182
1183 /// Set [`Pipeline::ocr_textline_orientation`] here and on a loaded
1184 /// PP-OCR recognizer.
1185 fn set_textline_orientation(&mut self, enable: bool) {
1186 self.textline_orientation = enable;
1187 if let OcrSlot::Ready(Recognizer::PpOcr(m)) = &mut self.ocr {
1188 m.set_textline_orientation(enable);
1189 }
1190 }
1191
1192 /// The OCR model, or `None` when this conversion must not (or cannot) OCR:
1193 /// `skip_ocr` short-circuits, and a failed model load degrades to `None`
1194 /// with a one-time warning instead of failing the conversion (#244) —
1195 /// unless `force_full_page_ocr` demanded OCR explicitly, where a missing
1196 /// model stays a hard error (the text layer was deliberately discarded, so
1197 /// degrading would silently emit an empty document).
1198 fn ocr_model(&mut self) -> Result<Option<&mut Recognizer>, PdfError> {
1199 if self.skip_ocr {
1200 return Ok(None);
1201 }
1202 if matches!(self.ocr, OcrSlot::Unloaded) {
1203 let loaded = match self.ocr_engine {
1204 ocr::OcrEngine::PpOcr => {
1205 ocr::OcrModel::load_with(self.ocr_lang, self.intra).map(|mut m| {
1206 m.set_textline_orientation(self.textline_orientation);
1207 Recognizer::PpOcr(m)
1208 })
1209 }
1210 ocr::OcrEngine::Tesseract => tesseract::TesseractOcr::load(
1211 tesseract::TesseractOptions::from_env(self.tesseract_lang.clone()),
1212 self.intra,
1213 )
1214 .map(Recognizer::Tesseract),
1215 };
1216 match loaded {
1217 Ok(model) => self.ocr = OcrSlot::Ready(model),
1218 Err(e) if self.force_full_page_ocr => return Err(PdfError::Ocr(e)),
1219 Err(e) => {
1220 static WARNED: std::sync::Once = std::sync::Once::new();
1221 let hint = match self.ocr_engine {
1222 ocr::OcrEngine::PpOcr => {
1223 "run scripts/install/download_dependencies.sh for the model"
1224 }
1225 ocr::OcrEngine::Tesseract => {
1226 "install tesseract-ocr with the language pack, or use --ocr-engine ppocr"
1227 }
1228 };
1229 WARNED.call_once(|| {
1230 eprintln!(
1231 "warning: OCR model unavailable ({e}); continuing without OCR — \
1232 scanned pages and text inside images will come back empty \
1233 ({hint})"
1234 );
1235 });
1236 self.ocr = OcrSlot::Missing;
1237 }
1238 }
1239 }
1240 Ok(match &mut self.ocr {
1241 OcrSlot::Ready(model) => Some(model),
1242 _ => None,
1243 })
1244 }
1245
1246 /// The text detector (#429): loaded on first use, `None` with `skip_ocr`
1247 /// or when the model is not installed — recognition stays region-scoped
1248 /// then, exactly the pre-#429 behavior, and `DOCLING_RS_DEBUG` says why.
1249 fn det_model(&mut self) -> Option<&mut ocr_det::DetModel> {
1250 if self.skip_ocr {
1251 return None;
1252 }
1253 if matches!(self.det, DetSlot::Unloaded) {
1254 match ocr_det::DetModel::load(self.intra) {
1255 Ok(model) => self.det = DetSlot::Ready(model),
1256 Err(e) => {
1257 docling_core::debug_log!(
1258 "docling-pdf: text detection unavailable ({e}); OCR stays region-scoped"
1259 );
1260 self.det = DetSlot::Missing;
1261 }
1262 }
1263 }
1264 match &mut self.det {
1265 DetSlot::Ready(model) => Some(model),
1266 _ => None,
1267 }
1268 }
1269
1270 /// Run layout (+ OCR for cell-less pages) + TableFormer and assemble page `n`
1271 /// into its nodes and links. Pure given the page (mutates only the worker's
1272 /// lazily-loaded OCR model), so it is safe to run concurrently across pages.
1273 fn process(&mut self, n: usize, page: &mut PdfPage) -> Result<PageOut, PdfError> {
1274 if self.no_ocr {
1275 // Fastest path (`text_only_page`): the text layer alone, nothing
1276 // inferred. Pages with no embedded text layer yield nothing;
1277 // convert those without `no_ocr`.
1278 let (nodes, links, conf) = text_only_page(page, self.images.scale);
1279 return Ok((nodes, links, conf, self.page_image(n, page), Vec::new()));
1280 }
1281 self.normalize_orientation(n, page)?;
1282 let det_pages = self.det_candidates(std::slice::from_ref(&(n, &*page)))?;
1283 let regions = {
1284 let Self {
1285 layout,
1286 det,
1287 pending_det,
1288 ocr_scale,
1289 ..
1290 } = self;
1291 let layout = layout.as_mut().expect("layout model loaded unless no_ocr");
1292 let page: &PdfPage = &*page;
1293 Self::detect_alongside(det, pending_det, *ocr_scale, &det_pages, || {
1294 timing::timed("layout.predict", || {
1295 layout.predict(layout_src(page), page.width, page.height)
1296 })
1297 })
1298 }
1299 .map_err(|e| PdfError::Layout(format!("page {}: {e}", n + 1)))?;
1300 // The 2x bitmap (#682), when the detections read it (#897).
1301 self.ensure_bitmap(page, ®ions)?;
1302 self.finish_page(n, page, regions)
1303 }
1304
1305 /// The page's render as docling's `PageItem.image` (#520), at
1306 /// [`ImageOutput::scale`] — `None` unless page images were requested, or
1307 /// when the page has no bitmap (the text-layer-only `no_ocr` path never
1308 /// renders one).
1309 fn page_image(&self, n: usize, page: &PdfPage) -> Option<docling_core::PictureImage> {
1310 page_image_under(self.images, n, page)
1311 }
1312
1313 /// The pages of a batch the text detector should sweep (#429): bitmap
1314 /// pages (no text layer) on a run with OCR and the detector available —
1315 /// loading both lazily here so the concurrent run below needs no `self`.
1316 fn det_candidates<'p>(
1317 &mut self,
1318 pages: &[(usize, &'p PdfPage)],
1319 ) -> Result<Vec<(usize, &'p PdfPage)>, PdfError> {
1320 // Pages the orientation pass already detected (#571) are skipped:
1321 // their boxes wait in `pending_det`.
1322 let scanned: Vec<(usize, &PdfPage)> = pages
1323 .iter()
1324 .filter(|(n, p)| {
1325 p.cells.is_empty() && p.image.width() > 1 && !self.pending_det.contains_key(n)
1326 })
1327 .copied()
1328 .collect();
1329 if scanned.is_empty() || self.ocr_model()?.is_none() || self.det_model().is_none() {
1330 return Ok(Vec::new());
1331 }
1332 Ok(scanned)
1333 }
1334
1335 /// Run `layout` on the calling thread while the text detector sweeps
1336 /// `det_pages` on another (#429): the detector only reads each page's
1337 /// image, layout only reads the pages too, and the two sessions are
1338 /// independent, so on a scanned page the ~0.7 s detection hides behind
1339 /// layout instead of adding to it. Results land in `pending_det` for the
1340 /// OCR block; a detector failure is recorded per page and surfaces there.
1341 fn detect_alongside<T>(
1342 det: &mut DetSlot,
1343 pending_det: &mut BTreeMap<usize, Result<Vec<ocr_det::DetBox>, String>>,
1344 ocr_scale: Option<f32>,
1345 det_pages: &[(usize, &PdfPage)],
1346 layout: impl FnOnce() -> T,
1347 ) -> T {
1348 let DetSlot::Ready(det) = det else {
1349 return layout();
1350 };
1351 if det_pages.is_empty() {
1352 return layout();
1353 }
1354 std::thread::scope(|s| {
1355 let handle = s.spawn(move || {
1356 det_pages
1357 .iter()
1358 .map(|&(n, page)| {
1359 let mut view = None;
1360 let (img, _) = ocr_input(
1361 &mut view,
1362 &page.image,
1363 page.scale,
1364 page_ocr_scale(ocr_scale, page.width, page.height, page.scale),
1365 );
1366 (n, timing::timed("ocr.det", || det.detect(img)))
1367 })
1368 .collect::<Vec<_>>()
1369 });
1370 let out = layout();
1371 match handle.join() {
1372 Ok(results) => pending_det.extend(results),
1373 Err(_) => {
1374 for &(n, _) in det_pages {
1375 pending_det.insert(n, Err("ocr-det: detection thread panicked".into()));
1376 }
1377 }
1378 }
1379 out
1380 })
1381 }
1382
1383 /// Content-based orientation normalization (#225), before any inference:
1384 /// a physically rotated scan (sideways phone photo, landscape-fed sheet)
1385 /// has `/Rotate 0`, so the metadata pass in `extract_page` never fires and
1386 /// layout+OCR would read a sideways raster. Only pages with no text layer
1387 /// at all are probed (a digital page's raster is upright by construction,
1388 /// and its cells — not its pixels — carry the text); the detected angle
1389 /// composes with any `/Rotate` normalization through the same
1390 /// [`PdfPage::unrotate`] + display-space assembly mapping. Detection is
1391 /// evidence-gated and degrades to a no-op — see [`orient`].
1392 fn normalize_orientation(&mut self, n: usize, page: &mut PdfPage) -> Result<(), PdfError> {
1393 let scanned =
1394 page.cells.is_empty() && page.word_cells.is_empty() && page.code_cells.is_empty();
1395 if self.no_ocr || self.skip_ocr || !scanned || page.image.width() <= 1 || !orient::enabled()
1396 {
1397 return Ok(());
1398 }
1399 // The probe reads text through the OCR model; without one (missing —
1400 // #244 degradation) the page stays as rendered.
1401 if self.ocr_model()?.is_none() {
1402 return Ok(());
1403 }
1404 // The text detector's lines are the probe's crops (#571). Detected on
1405 // the page image as rendered; an upright page keeps them for the OCR
1406 // pass (`pending_det`) when that pass reads the same bitmap, so the
1407 // detector runs once per page as before — a rotated page detects
1408 // again on the un-rotated image, alongside layout.
1409 let boxes: Vec<ocr_det::DetBox> = match self.det_model() {
1410 Some(det) => timing::timed("orient.det", || det.detect(&page.image)).unwrap_or_else(|e| {
1411 debug_log!("docling-pdf: page {}: text detection failed ({e}); probing projection strips", n + 1);
1412 Vec::new()
1413 }),
1414 None => Vec::new(),
1415 };
1416 let ocr_scale = page_ocr_scale(self.ocr_scale, page.width, page.height, page.scale);
1417 let Some(ocr) = self.ocr_model()? else {
1418 return Ok(());
1419 };
1420 let deg = timing::timed("orient.detect", || {
1421 orient::detect(&page.image, ocr, page.scale, &boxes)
1422 });
1423 if deg != 0 {
1424 debug_log!(
1425 "docling-pdf: page {}: content rotated {deg}° in the raster; \
1426 un-rotating before layout/OCR",
1427 n + 1
1428 );
1429 page.unrotate(deg);
1430 } else if !boxes.is_empty() && ocr_scale.is_none_or(|s| (s - page.scale).abs() <= 1e-3) {
1431 self.pending_det.insert(n, Ok(boxes));
1432 }
1433 Ok(())
1434 }
1435
1436 /// Layout-detect a whole batch of pages with one inference call (issue #73),
1437 /// then run each page's remaining stages (OCR / TableFormer / enrichment /
1438 /// assembly) per page. Index-aligned with `items`; a layout failure fails
1439 /// every page in the batch (they shared the one inference call).
1440 fn process_batch(&mut self, items: &mut [(usize, PdfPage)]) -> Vec<Result<Staged, PdfError>> {
1441 if self.no_ocr {
1442 // No layout model to batch — the text-layer-only path is per page.
1443 return items
1444 .iter_mut()
1445 .map(|(n, page)| {
1446 let n = *n;
1447 self.process(n, page).map(Staged::Done)
1448 })
1449 .collect();
1450 }
1451 // Orientation-normalize every scanned page before the shared layout
1452 // call — the batched inference must see upright bitmaps too (#225).
1453 for (n, page) in items.iter_mut() {
1454 let n = *n;
1455 if let Err(e) = self.normalize_orientation(n, page) {
1456 // Model-load failure — every page in the batch needs the same
1457 // model, so they all fail alike (mirrors the layout-error arm).
1458 let msg = e.to_string();
1459 return items
1460 .iter()
1461 .map(|_| Err(PdfError::Ocr(msg.clone())))
1462 .collect();
1463 }
1464 }
1465 let inputs: Vec<(layout::LayoutSrc<'_>, f32, f32)> = items
1466 .iter()
1467 .map(|(_, page)| (layout_src(page), page.width, page.height))
1468 .collect();
1469 let refs: Vec<(usize, &PdfPage)> = items.iter().map(|(n, page)| (*n, page)).collect();
1470 let det_pages = match self.det_candidates(&refs) {
1471 Ok(p) => p,
1472 Err(e) => {
1473 let msg = e.to_string();
1474 return items
1475 .iter()
1476 .map(|_| Err(PdfError::Ocr(msg.clone())))
1477 .collect();
1478 }
1479 };
1480 let batched = {
1481 let Self {
1482 layout,
1483 det,
1484 pending_det,
1485 ocr_scale,
1486 ..
1487 } = self;
1488 let layout = layout.as_mut().expect("layout model loaded unless no_ocr");
1489 Self::detect_alongside(det, pending_det, *ocr_scale, &det_pages, || {
1490 timing::timed("layout.predict", || layout.predict_batch(&inputs))
1491 })
1492 };
1493 match batched {
1494 Ok(all) => items
1495 .iter_mut()
1496 .zip(all)
1497 .map(|((n, page), regions)| {
1498 // The 2x bitmap (#682), when the detections read it (#897).
1499 self.ensure_bitmap(page, ®ions)?;
1500 self.stage_page(*n, page, regions)
1501 })
1502 .collect(),
1503 Err(e) => items
1504 .iter()
1505 .map(|(n, _)| Err(PdfError::Layout(format!("page {}: {e}", n + 1))))
1506 .collect(),
1507 }
1508 }
1509
1510 /// A pool worker's main loop: pull rendered pages off the shared channel
1511 /// (whatever is already there, up to the layout batch size), process them,
1512 /// and hand each finished page — or its error — to `deliver`, which returns
1513 /// `false` to stop early (the streaming consumer went away). Returns when
1514 /// the channel is closed and every page this worker took is delivered.
1515 ///
1516 /// Pages whose TableFormer turn would have to wait are parked (see
1517 /// `Staged`) and retried before every new pull; while any are parked the
1518 /// pull is non-blocking, so an empty channel means the worker waits for
1519 /// the TableFormer rather than for the renderer. The parking budget
1520 /// (`MAX_DEFERRED_PAGES`) bounds resident bitmaps; past it the worker waits
1521 /// like the serial path. Output is independent of completion order — the
1522 /// callers reassemble by page index.
1523 fn run_pool(
1524 &mut self,
1525 work_rx: &Mutex<Receiver<(usize, PdfPage)>>,
1526 layout_batch: usize,
1527 deadline: Option<std::time::Instant>,
1528 mut deliver: impl FnMut(usize, Result<PageOut, PdfError>) -> bool,
1529 ) {
1530 use std::sync::mpsc::TryRecvError;
1531 let mut deferred: std::collections::VecDeque<(usize, PdfPage, Prepared)> =
1532 std::collections::VecDeque::new();
1533 loop {
1534 // The document budget (#497), checked between pages: once it is
1535 // spent, every page still queued or parked is reported as timed
1536 // out instead of processed — the renderer stops feeding at the
1537 // same check, so the queue drains quickly.
1538 if expired(deadline) {
1539 while let Some((idx, _, _)) = deferred.pop_front() {
1540 if !deliver(idx, Err(timeout_sentinel())) {
1541 return;
1542 }
1543 }
1544 let rx = work_rx.lock().unwrap();
1545 loop {
1546 match rx.recv() {
1547 Ok((idx, _page)) => {
1548 if !deliver(idx, Err(timeout_sentinel())) {
1549 return;
1550 }
1551 }
1552 Err(_) => return,
1553 }
1554 }
1555 }
1556 // Any parked page the slot has since freed up for, oldest first.
1557 let mut i = 0;
1558 while i < deferred.len() {
1559 let (_, page, prepared) = &deferred[i];
1560 match self.table_rows_try(page, &prepared.regions) {
1561 Some(rows) => {
1562 let (idx, mut page, prepared) = deferred.remove(i).expect("index in range");
1563 if !deliver(idx, self.complete_page(idx, &mut page, prepared, rows)) {
1564 return;
1565 }
1566 }
1567 None => i += 1,
1568 }
1569 }
1570 // Hold the receiver lock only for the recv (plus a non-blocking drain
1571 // up to the layout batch size); release before the (long) per-page
1572 // work so other workers can pull concurrently.
1573 let mut batch: Vec<(usize, PdfPage)> = Vec::new();
1574 let mut closed = false;
1575 {
1576 let rx = work_rx.lock().unwrap();
1577 let first = if deferred.is_empty() {
1578 rx.recv().map_err(|_| TryRecvError::Disconnected)
1579 } else {
1580 rx.try_recv()
1581 };
1582 match first {
1583 Ok(item) => {
1584 batch.push(item);
1585 while batch.len() < layout_batch {
1586 match rx.try_recv() {
1587 Ok(item) => batch.push(item),
1588 Err(_) => break,
1589 }
1590 }
1591 }
1592 Err(TryRecvError::Empty) => {}
1593 Err(TryRecvError::Disconnected) => closed = true,
1594 }
1595 }
1596 if batch.is_empty() {
1597 // Nothing new rendered (or the channel is closed): wait our turn
1598 // on the oldest parked page instead.
1599 match deferred.pop_front() {
1600 Some((idx, mut page, prepared)) => {
1601 let rows = self.table_rows_blocking(&page, &prepared.regions);
1602 if !deliver(idx, self.complete_page(idx, &mut page, prepared, rows)) {
1603 return;
1604 }
1605 continue;
1606 }
1607 None if closed => return,
1608 None => continue,
1609 }
1610 }
1611 let outs = self.process_batch(&mut batch);
1612 for ((idx, mut page), out) in batch.into_iter().zip(outs) {
1613 let delivered = match out {
1614 Ok(Staged::Done(out)) => deliver(idx, Ok(out)),
1615 Ok(Staged::NeedsTables(prepared)) => {
1616 if deferred.len() < MAX_DEFERRED_PAGES {
1617 deferred.push_back((idx, page, prepared));
1618 true
1619 } else {
1620 let rows = self.table_rows_blocking(&page, &prepared.regions);
1621 deliver(idx, self.complete_page(idx, &mut page, prepared, rows))
1622 }
1623 }
1624 Err(e) => deliver(idx, Err(e)),
1625 };
1626 if !delivered {
1627 return;
1628 }
1629 }
1630 }
1631 }
1632
1633 /// Everything after layout detection: per-label confidence thresholds,
1634 /// overlap resolution, orphan-text recovery, OCR for cell-less pages,
1635 /// TableFormer, enrichment, and page assembly. The serial path: waits
1636 /// for the shared TableFormer when a table needs it.
1637 fn finish_page(
1638 &mut self,
1639 n: usize,
1640 page: &mut PdfPage,
1641 regions: Vec<layout::Region>,
1642 ) -> Result<PageOut, PdfError> {
1643 let prepared = self.prepare_page(n, page, regions)?;
1644 self.ensure_bitmap(page, &prepared.regions)?;
1645 let table_rows = self.table_rows_blocking(page, &prepared.regions);
1646 self.complete_page(n, page, prepared, table_rows)
1647 }
1648
1649 /// The pool path: like [`finish_page`](Self::finish_page), except that a
1650 /// page whose TableFormer turn would have to wait comes back as
1651 /// [`Staged::NeedsTables`] for the worker loop to park (see `Staged`).
1652 fn stage_page(
1653 &mut self,
1654 n: usize,
1655 page: &mut PdfPage,
1656 regions: Vec<layout::Region>,
1657 ) -> Result<Staged, PdfError> {
1658 let prepared = self.prepare_page(n, page, regions)?;
1659 self.ensure_bitmap(page, &prepared.regions)?;
1660 match self.table_rows_try(page, &prepared.regions) {
1661 Some(rows) => Ok(Staged::Done(self.complete_page(n, page, prepared, rows)?)),
1662 None => Ok(Staged::NeedsTables(prepared)),
1663 }
1664 }
1665
1666 /// Does this page read its 2x bitmap (#897)? Layout reads the separate
1667 /// 1x render; after it, the bitmap feeds TableFormer, picture crops (and
1668 /// the picture classifier, in-picture OCR and the image outputs on
1669 /// them), the code/formula enrichment crops, page images and OCR. A
1670 /// page without any of those — on a text-heavy document most pages —
1671 /// never renders it. Decided on the detections as they stand (a
1672 /// superset of what survives postprocessing, which only drops or
1673 /// demotes pictures and tables), and asked again wherever the regions
1674 /// change.
1675 fn needs_bitmap(&self, page: &PdfPage, regions: &[layout::Region]) -> bool {
1676 self.images.page_images
1677 || self.force_full_page_ocr
1678 || page.cells.is_empty()
1679 || regions.iter().any(|r| {
1680 r.label == "picture"
1681 || assemble::is_table_like(r.label)
1682 || (r.label == "code" && self.enrich.code)
1683 || (r.label == "formula" && self.enrich.formula)
1684 })
1685 }
1686
1687 /// Install the page's deferred 2x bitmap (#682) if [`Self::needs_bitmap`]
1688 /// says it is read — rendering it now when it was not rendered already —
1689 /// and otherwise decline it, so the render helper skips it when it has
1690 /// not got to it yet (#897). Idempotent, and a declined page can still
1691 /// be resolved later (an fp32 layout retry with new detections).
1692 fn ensure_bitmap(
1693 &self,
1694 page: &mut PdfPage,
1695 regions: &[layout::Region],
1696 ) -> Result<(), PdfError> {
1697 if page.image_pending() {
1698 if self.needs_bitmap(page, regions) {
1699 page.resolve_image()?;
1700 } else {
1701 page.decline_image();
1702 }
1703 }
1704 Ok(())
1705 }
1706
1707 /// Does this page need the shared TableFormer at all? Table-free pages
1708 /// never touch (or load) it.
1709 fn needs_tables(&self, regions: &[layout::Region]) -> bool {
1710 self.tables.is_some() && regions.iter().any(|r| assemble::is_table_like(r.label))
1711 }
1712
1713 /// TableFormer structure for every table region of the page, on an
1714 /// already-locked slot (loading the model on first use). Tables serialise
1715 /// on this mutex, so the one instance gets the shared thread budget
1716 /// (quota-aware, #262) — DOCLING_RS_TF_INTRA narrows it further where the
1717 /// memory-per-thread tradeoff matters more than table latency.
1718 fn predict_tables(
1719 guard: &mut TfSlot,
1720 page: &PdfPage,
1721 regions: &[layout::Region],
1722 ) -> Vec<Option<tf_core::TableGrid>> {
1723 let mut table_rows: Vec<Option<tf_core::TableGrid>> = vec![None; regions.len()];
1724 debug_assert!(
1725 !page.image_pending(),
1726 "TableFormer reached an unrendered page bitmap"
1727 );
1728 guard.load();
1729 if let TfSlot::Ready(tf) = guard {
1730 // One 1024-px frame per page, shared by all of its tables, and one
1731 // call for all of them: with the dynamic-batch decoder their
1732 // decode steps are shared (each step costs about the same for B
1733 // tables as for one).
1734 let page1024 = tableformer::TableFormer::page_1024(&page.image);
1735 let (idx, boxes): (Vec<usize>, Vec<[f32; 4]>) = regions
1736 .iter()
1737 .enumerate()
1738 .filter(|(_, r)| assemble::is_table_like(r.label))
1739 .map(|(i, r)| (i, [r.l, r.t, r.r, r.b]))
1740 .unzip();
1741 let rows =
1742 tf.predict_tables_on(page.image.height(), &page1024, &boxes, &page.word_cells);
1743 for (i, grid) in idx.into_iter().zip(rows) {
1744 table_rows[i] = grid;
1745 }
1746 }
1747 table_rows
1748 }
1749
1750 /// Table structure for the page, waiting for the shared slot if another
1751 /// worker holds it (else geometric fallback downstream when there is no
1752 /// TableFormer at all). The `tableformer` timing stage here includes any
1753 /// wait.
1754 fn table_rows_blocking(
1755 &self,
1756 page: &PdfPage,
1757 regions: &[layout::Region],
1758 ) -> Vec<Option<tf_core::TableGrid>> {
1759 match self.tables.as_ref().filter(|_| self.needs_tables(regions)) {
1760 Some(slot) => timing::timed("tableformer", || {
1761 Self::predict_tables(&mut slot.lock().unwrap(), page, regions)
1762 }),
1763 None => vec![None; regions.len()],
1764 }
1765 }
1766
1767 /// Non-blocking variant: `None` when the slot is held by another worker
1768 /// right now — the caller parks the page and tries again later.
1769 fn table_rows_try(
1770 &self,
1771 page: &PdfPage,
1772 regions: &[layout::Region],
1773 ) -> Option<Vec<Option<tf_core::TableGrid>>> {
1774 let Some(slot) = self.tables.as_ref().filter(|_| self.needs_tables(regions)) else {
1775 return Some(vec![None; regions.len()]);
1776 };
1777 match slot.try_lock() {
1778 Ok(mut guard) => Some(timing::timed("tableformer", || {
1779 Self::predict_tables(&mut guard, page, regions)
1780 })),
1781 Err(std::sync::TryLockError::WouldBlock) => None,
1782 Err(std::sync::TryLockError::Poisoned(e)) => panic!("TableFormer slot poisoned: {e}"),
1783 }
1784 }
1785
1786 /// The stages before TableFormer: fp32 escalation, per-label confidence
1787 /// thresholds, overlap resolution, orphan-text recovery, OCR for cell-less
1788 /// pages, in-picture text and table-word recognition.
1789 fn prepare_page(
1790 &mut self,
1791 n: usize,
1792 page: &mut PdfPage,
1793 regions: Vec<layout::Region>,
1794 ) -> Result<Prepared, PdfError> {
1795 // Force-OCR is exactly "pretend the text layer is not there": clear
1796 // every cell kind the extractors produced before anything reads them,
1797 // and the ordinary no-text-layer machinery below — full-page OCR,
1798 // OCR-fed TableFormer matching — takes over unchanged. (`no_ocr` wins
1799 // when both are set, mirroring docling, where `force_full_page_ocr`
1800 // is a sub-option of `do_ocr`; the no-ocr path never reaches here.)
1801 // Done here rather than in `process` so the batched layout path
1802 // (`process_batch` → `finish_page`) honors the flag too.
1803 // Parse quality is scored on the extracted text layer before force-OCR
1804 // discards it (docling's page-preprocessing stage runs before OCR too,
1805 // so its parse_score also reflects the original text layer).
1806 let parse = quality::parse_score(&page.cells);
1807 // Recognition confidences of every OCR'd cell on this page → ocr_score.
1808 let mut ocr_confs: Vec<f32> = Vec::new();
1809 let mut ocr_cells: Vec<docling_core::OcrCellConfidence> = Vec::new();
1810 // The bitmap the OCR reads (#254): with `ocr_scale` set, a resample of
1811 // the page render at the requested px/pt, built lazily on the first
1812 // OCR use so non-OCR pages never pay for it. Copied out of `self` up
1813 // front — the OCR sites hold `self.ocr_model()`'s mutable borrow.
1814 let ocr_scale = page_ocr_scale(self.ocr_scale, page.width, page.height, page.scale);
1815 let mut ocr_view: Option<image::RgbImage> = None;
1816 if self.force_full_page_ocr {
1817 page.cells.clear();
1818 page.code_cells.clear();
1819 page.word_cells.clear();
1820 }
1821 // Quant-robustness guard: the default int8 layout graph keeps its
1822 // confidences near the 0.5 label thresholds, and a different CPU's
1823 // quantized kernels can flip a whole page's detections under them —
1824 // tables and paragraphs then dissolve into orphan one-liners while the
1825 // same build converts the page perfectly elsewhere. When a dense
1826 // digital page ends up with detections covering almost none of its
1827 // text cells, re-run that one page on the fp32 graph (lazy-loaded,
1828 // auto-int8 selection only) and keep whichever detections cover more.
1829 let mut regions = regions;
1830 if !page.cells.is_empty() {
1831 let thresholded = |rs: &[layout::Region]| -> Vec<layout::Region> {
1832 rs.iter()
1833 .filter(|r| r.score >= layout::label_threshold(r.label))
1834 .cloned()
1835 .collect()
1836 };
1837 let text_cells = page
1838 .cells
1839 .iter()
1840 .filter(|c| !c.text.trim().is_empty())
1841 .count();
1842 let cov = assemble::layout_cell_coverage(&thresholded(®ions), &page.cells);
1843 if text_cells >= 15 && cov < 0.5 {
1844 let retry = self
1845 .layout
1846 .as_mut()
1847 .expect("layout model loaded unless no_ocr")
1848 .predict_fp32_fallback(layout_src(page), page.width, page.height)
1849 .map_err(|e| PdfError::Layout(format!("page {}: {e}", n + 1)))?;
1850 if let Some(retry) = retry {
1851 let cov2 = assemble::layout_cell_coverage(&thresholded(&retry), &page.cells);
1852 if cov2 > cov {
1853 debug_log!(
1854 "docling-pdf: page {}: int8 layout covered {:.0}% of the text \
1855 cells; the fp32 retry covers {:.0}% — using it",
1856 n + 1,
1857 cov * 100.0,
1858 cov2 * 100.0
1859 );
1860 regions = retry;
1861 // New detections may read the bitmap the first ones
1862 // did not (#897).
1863 self.ensure_bitmap(page, ®ions)?;
1864 }
1865 }
1866 }
1867 }
1868 // docling's LayoutPostprocessor drops each detection below its label's
1869 // confidence threshold (stricter than the 0.3 base the predictor keeps),
1870 // before any overlap resolution. This removes the low-confidence tables /
1871 // pictures / list-items that otherwise double-emit or mis-classify.
1872 if env::flag("DOCLING_RS_DEBUG_REGIONS") {
1873 for r in ®ions {
1874 eprintln!(
1875 "DBG raw {} {:.2} [{:.0},{:.0},{:.0},{:.0}]",
1876 r.label, r.score, r.l, r.t, r.r, r.b
1877 );
1878 }
1879 }
1880 regions.retain(|r| r.score >= layout::label_threshold(r.label));
1881 // docling's full-page picture filter and same-label picture dedup run
1882 // on the thresholded detections, before overlap resolution: a picture
1883 // that is the whole page goes (its text reads out as text), and a
1884 // figure proposed both whole and as sub-panels collapses to one box
1885 // (see `dedup_pictures`).
1886 assemble::drop_full_page_pictures(&mut regions, page.width, page.height);
1887 assemble::dedup_pictures(&mut regions);
1888 // Resolve overlapping detections once, before OCR.
1889 let mut regions = assemble::resolve(regions);
1890 // Emit text the detector missed as orphan text regions (docling parity).
1891 assemble::add_orphan_regions(&mut regions, &page.cells);
1892 // Drop phantom empty low-confidence picture boxes (docling parity).
1893 assemble::drop_false_pictures(&mut regions, &page.cells, page.width, page.height);
1894 // A regular region fully inside a surviving table/index/picture is that
1895 // special's child (a cell / in-figure label), not a separate block —
1896 // remove it so it isn't emitted twice (docling parity).
1897 assemble::drop_contained_regulars(&mut regions);
1898 // A one-line paragraph in the bottom margin under a body-less heading
1899 // is that heading's text, not furniture (deliberate deviation).
1900 assemble::reclaim_heading_body_footers(&mut regions, page.width);
1901 // No text layer → recognise text from the page image via OCR.
1902 let ocred = page.cells.is_empty();
1903 // Lines the text detector found outside every layout region (#429).
1904 let mut det_cells: Vec<pdfium_backend::TextCell> = Vec::new();
1905 if ocred {
1906 // Region-scoped OCR recognizes each region's crop, so two regular
1907 // regions over the same ink would read it twice — a low-score
1908 // paragraph box over the high-score line boxes `greedy` keeps.
1909 // Collapse such groups to one region first; a digital
1910 // page resolves the same overlap through cell ownership in
1911 // `fit_regions_to_cells` and needs nothing here.
1912 assemble::merge_overlapping_regulars(&mut regions);
1913 // The text detector's lines (#429), fetched before the region
1914 // pass: with `det` line mode (#570, the default) they are the
1915 // recognizer's crops inside the regions too — RapidOCR's own line
1916 // source — not only the supplement outside them. Usually computed
1917 // already, alongside layout (see `detect_alongside`); a page that
1918 // reached OCR another way detects here. Degradation over failure:
1919 // a detector that loaded but cannot run (a damaged file, an
1920 // allocation failure) costs the page its detected lines, not its
1921 // conversion — the projection segmentation is complete on its own.
1922 let detected: Vec<ocr_det::DetBox> = if self.ocr_model()?.is_some() {
1923 let (img, _) = ocr_input(&mut ocr_view, &page.image, page.scale, ocr_scale);
1924 match self.pending_det.remove(&n) {
1925 Some(result) => result,
1926 None => match self.det_model() {
1927 Some(det) => timing::timed("ocr.det", || det.detect(img)),
1928 None => Ok(Vec::new()),
1929 },
1930 }
1931 .unwrap_or_else(|e| {
1932 docling_core::debug_log!(
1933 "docling-pdf: page {}: text detection failed ({e}); keeping the \
1934 region-scoped OCR only",
1935 n + 1
1936 );
1937 Vec::new()
1938 })
1939 } else {
1940 Vec::new()
1941 };
1942 docling_core::debug_log!(
1943 "docling-pdf: page {}: text detector found {} line(s): {:?}",
1944 n + 1,
1945 detected.len(),
1946 detected
1947 .iter()
1948 .map(|d| (
1949 d.l.round(),
1950 d.t.round(),
1951 d.r.round(),
1952 d.b.round(),
1953 (d.score * 100.0).round() / 100.0
1954 ))
1955 .collect::<Vec<_>>()
1956 );
1957 let det_lines =
1958 (ocr::det_lines() && !detected.is_empty()).then_some(detected.as_slice());
1959 // `None` = `skip_ocr` or a missing model (#244): the page keeps
1960 // its layout regions (and TableFormer structure below) with no
1961 // recognized text, instead of failing the conversion.
1962 if let Some(ocr) = self.ocr_model()? {
1963 let (img, scl) = ocr_input(&mut ocr_view, &page.image, page.scale, ocr_scale);
1964 let cells = timing::timed("ocr.page", || {
1965 ocr.ocr_page_with(img, ®ions, scl, det_lines)
1966 })
1967 .map_err(|e| PdfError::Ocr(format!("page {}: {e}", n + 1)))?;
1968 ocr_confs.extend(cells.iter().map(|(_, conf)| conf));
1969 ocr_cells.extend(cell_confs(&cells));
1970 page.cells = cells.into_iter().map(|(cell, _)| cell).collect();
1971 // Table interiors carry no words yet: region-scoped OCR skips
1972 // table labels, and a scanned page has no pdfium text layer — so
1973 // TableFormer's cell matcher got an empty word list and the table
1974 // dissolved (#173). Recognize the table regions' word crops
1975 // (mirroring the browser scanned path): `word_cells` feeds the
1976 // matcher, and the same cells join `cells` so the geometric
1977 // fallback and the table's region text see them too.
1978 if regions.iter().any(|r| assemble::is_table_like(r.label)) {
1979 let (img, scl) = ocr_input(&mut ocr_view, &page.image, page.scale, ocr_scale);
1980 let words = timing::timed("ocr.table_words", || {
1981 ocr.ocr_table_words(img, ®ions, scl, det_lines)
1982 })
1983 .map_err(|e| PdfError::Ocr(format!("page {}: {e}", n + 1)))?;
1984 ocr_confs.extend(words.iter().map(|(_, conf)| conf));
1985 ocr_cells.extend(cell_confs(&words));
1986 let words: Vec<_> = words.into_iter().map(|(cell, _)| cell).collect();
1987 page.cells.extend(words.iter().cloned());
1988 page.word_cells = words;
1989 }
1990 }
1991 // docling's OCR engines detect text lines over the whole bitmap
1992 // and every line becomes a cell, so text the layout model gave no
1993 // region — a diagram's labels, a stamp, a margin note — still
1994 // reads out as orphan text. Region-scoped recognition above stays
1995 // the source inside layout regions; the detector (#429) adds only
1996 // the lines no recognized cell already covers, and the orphan pass
1997 // below places them (those inside a kept picture or table become
1998 // that special's silent children, as upstream).
1999 if self.ocr_model()?.is_some() {
2000 let (img, scl) = ocr_input(&mut ocr_view, &page.image, page.scale, ocr_scale);
2001 let uncovered = ocr_det::uncovered_lines(&detected, scl, ®ions, &page.cells);
2002 docling_core::debug_log!(
2003 "docling-pdf: page {}: {} of {} detected line(s) not covered by the region pass",
2004 n + 1,
2005 uncovered.len(),
2006 detected.len()
2007 );
2008 if let (false, Some(ocr)) = (uncovered.is_empty(), self.ocr_model()?) {
2009 let scored =
2010 timing::timed("ocr.det_lines", || ocr.ocr_page(img, &uncovered, scl))
2011 .map_err(|e| PdfError::Ocr(format!("page {}: {e}", n + 1)))?;
2012 ocr_confs.extend(scored.iter().map(|(_, conf)| conf));
2013 ocr_cells.extend(cell_confs(&scored));
2014 det_cells = scored.into_iter().map(|(cell, _)| cell).collect();
2015 page.cells.extend(det_cells.iter().cloned());
2016 }
2017 }
2018 } else {
2019 // A page with a text layer can still draw part of its text as
2020 // vector outlines (`gs -dNoOutputFonts`, flattened print PDFs):
2021 // the layout model finds the paragraph, the text layer holds
2022 // nothing there, and the region dropped empty — one native `Page
2023 // 1` footer was enough to switch the page's OCR off (#931).
2024 // docling's default `pdf_aware_layout_regions` OCRs every layout
2025 // cluster without any programmatic text in it
2026 // (`_find_pdf_aware_layout_ocr_rects`); do the same for the text,
2027 // furniture and table regions (pictures go through the bare-
2028 // picture pass below).
2029 let empty = text_less_regions(®ions, &page.cells);
2030 if let (false, Some(_)) = (empty.is_empty(), self.ocr_model()?) {
2031 // A page with a text layer and no picture or table had its
2032 // bitmap declined (#897); render it now — rare, and only
2033 // for the page that needs it.
2034 page.resolve_image()?;
2035 let (img, scl) = ocr_input(&mut ocr_view, &page.image, page.scale, ocr_scale);
2036 let detected = match self.det_model() {
2037 Some(det) => timing::timed("ocr.det", || det.detect(img)).unwrap_or_else(|e| {
2038 docling_core::debug_log!(
2039 "docling-pdf: page {}: text detection failed ({e}); keeping the \
2040 region-scoped OCR only",
2041 n + 1
2042 );
2043 Vec::new()
2044 }),
2045 None => Vec::new(),
2046 };
2047 let det_lines =
2048 (ocr::det_lines() && !detected.is_empty()).then_some(detected.as_slice());
2049 let ocr = self.ocr_model()?.expect("checked above");
2050 let scored = timing::timed("ocr.page", || {
2051 ocr.ocr_page_with(img, &empty, scl, det_lines)
2052 })
2053 .map_err(|e| PdfError::Ocr(format!("page {}: {e}", n + 1)))?;
2054 docling_core::debug_log!(
2055 "docling-pdf: page {}: {} text-less layout region(s) OCR'd, {} line(s)",
2056 n + 1,
2057 empty.len(),
2058 scored.len()
2059 );
2060 ocr_confs.extend(scored.iter().map(|(_, conf)| conf));
2061 ocr_cells.extend(cell_confs(&scored));
2062 page.cells.extend(scored.into_iter().map(|(cell, _)| cell));
2063 if empty.iter().any(|r| assemble::is_table_like(r.label)) {
2064 let words = timing::timed("ocr.table_words", || {
2065 ocr.ocr_table_words(img, &empty, scl, det_lines)
2066 })
2067 .map_err(|e| PdfError::Ocr(format!("page {}: {e}", n + 1)))?;
2068 ocr_confs.extend(words.iter().map(|(_, conf)| conf));
2069 ocr_cells.extend(cell_confs(&words));
2070 let words: Vec<_> = words.into_iter().map(|(cell, _)| cell).collect();
2071 page.cells.extend(words.iter().cloned());
2072 page.word_cells.extend(words);
2073 }
2074 }
2075 }
2076 // Region-scoped OCR skips `picture` interiors, and a digital page's
2077 // text layer cannot see into an embedded raster either — so a figure
2078 // that is really a text box (terms-and-conditions exported as an
2079 // image) lost its words on every page kind. Python docling OCRs the
2080 // bitmap-covered areas of *every* page — even digital ones — once they
2081 // exceed `bitmap_area_threshold` (5 % of the page); the browser paths
2082 // already do. Recognize the big text-less crops here too; the panel
2083 // demotion / orphan recovery below place the lines.
2084 let mut pic_cells: Vec<pdfium_backend::TextCell> = Vec::new();
2085 {
2086 let page_area = (page.width * page.height).max(1.0);
2087 let has_text = |r: &layout::Region| {
2088 page.cells.iter().any(|c| {
2089 let ca = ((c.r - c.l) * (c.b - c.t)).max(1.0);
2090 let ix = (r.r.min(c.r) - r.l.max(c.l)).max(0.0);
2091 let iy = (r.b.min(c.b) - r.t.max(c.t)).max(0.0);
2092 !c.text.trim().is_empty() && ix * iy / ca > 0.5
2093 })
2094 };
2095 // A captioned picture can never demote to a text panel (see
2096 // recover_text_panels), and on digital pages its speculative OCR
2097 // would be discarded anyway — don't pay for it.
2098 let captioned = |r: &layout::Region| {
2099 regions.iter().any(|c| {
2100 c.label == "caption"
2101 && c.r.min(r.r) - c.l.max(r.l) > 0.0
2102 && ((c.t >= r.b && c.t - r.b <= 25.0) || (r.t >= c.b && r.t - c.b <= 25.0))
2103 })
2104 };
2105 let bare: Vec<layout::Region> = regions
2106 .iter()
2107 .filter(|r| {
2108 r.label == "picture"
2109 && (r.r - r.l) * (r.b - r.t) / page_area >= 0.05
2110 && !has_text(r)
2111 && (ocred || !captioned(r))
2112 })
2113 .map(|r| layout::Region {
2114 label: "text",
2115 ..r.clone()
2116 })
2117 .collect();
2118 // Speculative OCR (#244): with `skip_ocr` or no model, big bare
2119 // pictures simply stay pictures.
2120 if let (false, Some(ocr)) = (bare.is_empty(), self.ocr_model()?) {
2121 let (img, scl) = ocr_input(&mut ocr_view, &page.image, page.scale, ocr_scale);
2122 let scored = timing::timed("ocr.pictures", || ocr.ocr_page(img, &bare, scl))
2123 .map_err(|e| PdfError::Ocr(format!("page {}: {e}", n + 1)))?;
2124 // Speculative in-picture OCR counts toward ocr_score only on
2125 // OCR'd pages, where the recognized lines actually join the
2126 // output; on a digital page they may be discarded below.
2127 if ocred {
2128 ocr_confs.extend(scored.iter().map(|(_, conf)| conf));
2129 ocr_cells.extend(cell_confs(&scored));
2130 }
2131 pic_cells = scored.into_iter().map(|(cell, _)| cell).collect();
2132 page.cells.extend(pic_cells.iter().cloned());
2133 }
2134 }
2135 let cells_before_pic_ocr = page.cells.len() - pic_cells.len();
2136 // A "picture" that is really a colored text panel — dense, wide,
2137 // multi-line — reads out as paragraphs instead of shipping as pixels;
2138 // sparse in-picture text (a chart's labels) keeps the crop and stays
2139 // inside it as the picture's silent children (docling parity, #200).
2140 // `no_text_panels` (#173) opts out entirely for image-extraction
2141 // workflows.
2142 if !self.no_text_panels {
2143 assemble::recover_text_panels(&mut regions, &page.cells);
2144 }
2145 // On an OCR'd page, in-picture text that did NOT demote its picture
2146 // mostly stays silent, exactly as in docling: its postprocess step
2147 // "Remove regular clusters that are included in wrappers" walks
2148 // SPECIAL_TYPES — which includes PICTURE — so an orphan text cluster
2149 // >80 % contained in a kept picture becomes that picture's child and
2150 // never reaches the serializer. Only border-straddlers (≤80 %
2151 // containment) survive as text. Emitting *everything* here used to
2152 // splice a chart's OCR'd axis ticks into the body text right next to
2153 // the image chunk (#200) — so the orphan pass places the recognized
2154 // lines, the containment drop re-runs for the in-table ones, and
2155 // `assemble_page` nests the in-picture ones under their picture
2156 // (JSON-only children, `assemble::picture_parents`).
2157 if ocred && (!pic_cells.is_empty() || !det_cells.is_empty()) {
2158 // Pictures (and wrappers) no longer count as claimers (#165), so
2159 // the plain orphan pass places the recognized lines directly —
2160 // the detector's lines (#429) the same way.
2161 assemble::add_orphan_regions(&mut regions, &pic_cells);
2162 assemble::add_orphan_regions(&mut regions, &det_cells);
2163 assemble::drop_contained_regulars(&mut regions);
2164 } else if !ocred && !pic_cells.is_empty() {
2165 // Digital page, picture kept: its speculative OCR cells must not
2166 // linger in the text-cell set (they were appended at the tail).
2167 let kept: Vec<layout::Region> = regions
2168 .iter()
2169 .filter(|r| r.label == "picture")
2170 .cloned()
2171 .collect();
2172 let tail = page.cells.split_off(cells_before_pic_ocr);
2173 page.cells.extend(tail.into_iter().filter(|c| {
2174 !kept.iter().any(|r| {
2175 let ca = ((c.r - c.l) * (c.b - c.t)).max(1.0);
2176 let ix = (r.r.min(c.r) - r.l.max(c.l)).max(0.0);
2177 let iy = (r.b.min(c.b) - r.t.max(c.t)).max(0.0);
2178 ix * iy / ca > 0.5
2179 })
2180 }));
2181 }
2182 // A text-less *table* detected inside a picture on a digital page — a
2183 // screenshot of a table (2203's Figure 10) — has no text layer and no
2184 // scanned-path OCR to feed it, so its grid used to serialize empty and
2185 // the whole element vanished. docling OCRs bitmap-covered areas on
2186 // every page kind and its table cluster collects those cells; mirror
2187 // the scanned path for exactly these tables: recognize word crops and
2188 // feed them to the TableFormer matcher and the cell set.
2189 if !ocred {
2190 let has_text = |t: &layout::Region| {
2191 page.cells.iter().any(|c| {
2192 let ca = ((c.r - c.l) * (c.b - c.t)).max(1.0);
2193 let ix = (t.r.min(c.r) - t.l.max(c.l)).max(0.0);
2194 let iy = (t.b.min(c.b) - t.t.max(c.t)).max(0.0);
2195 !c.text.trim().is_empty() && ix * iy / ca > 0.5
2196 })
2197 };
2198 let in_picture = |t: &layout::Region| {
2199 regions.iter().any(|r| {
2200 r.label == "picture" && {
2201 let ta = ((t.r - t.l) * (t.b - t.t)).max(1.0);
2202 let ix = (r.r.min(t.r) - r.l.max(t.l)).max(0.0);
2203 let iy = (r.b.min(t.b) - r.t.max(t.t)).max(0.0);
2204 ix * iy / ta > 0.5
2205 }
2206 })
2207 };
2208 let pic_tables: Vec<layout::Region> = regions
2209 .iter()
2210 .filter(|t| assemble::is_table_like(t.label) && !has_text(t) && in_picture(t))
2211 .cloned()
2212 .collect();
2213 // Same degradation as above: without OCR the in-picture table
2214 // keeps its structure (TableFormer is geometry-driven) minus text.
2215 if let (false, Some(ocr)) = (pic_tables.is_empty(), self.ocr_model()?) {
2216 let (img, scl) = ocr_input(&mut ocr_view, &page.image, page.scale, ocr_scale);
2217 let words = timing::timed("ocr.table_words", || {
2218 ocr.ocr_table_words(img, &pic_tables, scl, None)
2219 })
2220 .map_err(|e| PdfError::Ocr(format!("page {}: {e}", n + 1)))?;
2221 ocr_confs.extend(words.iter().map(|(_, conf)| conf));
2222 ocr_cells.extend(cell_confs(&words));
2223 let words: Vec<_> = words.into_iter().map(|(cell, _)| cell).collect();
2224 page.cells.extend(words.iter().cloned());
2225 page.word_cells.extend(words);
2226 }
2227 }
2228 // The cells are final: fit every regular region to the cells it
2229 // claims and fold the orphans it now surrounds (#419), before
2230 // TableFormer and the reading order see the boxes.
2231 assemble::fit_regions_to_cells(&mut regions, &page.cells);
2232 Ok(Prepared {
2233 regions,
2234 ocr_confs,
2235 ocr_cells,
2236 parse,
2237 })
2238 }
2239
2240 /// The stages after TableFormer: enrichment, the page confidence report and
2241 /// assembly into typed nodes.
2242 fn complete_page(
2243 &mut self,
2244 n: usize,
2245 page: &mut PdfPage,
2246 prepared: Prepared,
2247 table_rows: Vec<Option<tf_core::TableGrid>>,
2248 ) -> Result<PageOut, PdfError> {
2249 let Prepared {
2250 regions,
2251 ocr_confs,
2252 ocr_cells,
2253 parse,
2254 } = prepared;
2255 // The crops and the page image below read the bitmap (#897).
2256 self.ensure_bitmap(page, ®ions)?;
2257 if env::flag("DOCLING_RS_DEBUG_REGIONS") {
2258 for (i, r) in regions.iter().enumerate() {
2259 eprintln!(
2260 "DBG final {} {:.2} [{:.0},{:.0},{:.0},{:.0}] rows={:?}",
2261 r.label,
2262 r.score,
2263 r.l,
2264 r.t,
2265 r.r,
2266 r.b,
2267 table_rows[i]
2268 .as_ref()
2269 .map(|t| (t.rows.len(), t.rows.first().map(|r| r.len())))
2270 );
2271 }
2272 eprintln!(
2273 "DBG cells={} words={}",
2274 page.cells.len(),
2275 page.word_cells.len()
2276 );
2277 }
2278 // Enrichment passes (opt-in): DocumentPictureClassifier over picture
2279 // regions, CodeFormulaV2 over code/formula regions. Same shared-slot
2280 // shape as TableFormer — one lazily-loaded instance per pipeline, only
2281 // ever locked when a page actually has a matching region.
2282 let mut enrich_out: Vec<Option<assemble::Enrichment>> = vec![None; regions.len()];
2283 if let Some(slot) = self.classifier.as_ref() {
2284 if regions.iter().any(|r| r.label == "picture") {
2285 timing::timed("picture_classifier", || {
2286 let mut guard = slot.lock().unwrap();
2287 if matches!(*guard, EnrichSlot::Unloaded) {
2288 *guard = match enrich::PictureClassifier::load_with(intra_threads()) {
2289 Some(m) => EnrichSlot::Ready(m),
2290 None => EnrichSlot::Missing,
2291 };
2292 }
2293 if let EnrichSlot::Ready(model) = &mut *guard {
2294 for (i, r) in regions.iter().enumerate() {
2295 if r.label != "picture" {
2296 continue;
2297 }
2298 let Some(crop) = assemble::crop_region_scaled(
2299 page,
2300 [r.l, r.t, r.r, r.b],
2301 enrich::CLASSIFIER_SCALE,
2302 ) else {
2303 continue;
2304 };
2305 match model.classify(&crop) {
2306 Ok(classes) => {
2307 enrich_out[i] =
2308 Some(assemble::Enrichment::PictureClasses(classes));
2309 }
2310 Err(e) => eprintln!("docling-pdf: page {}: {e}", n + 1),
2311 }
2312 }
2313 }
2314 });
2315 }
2316 }
2317 if let Some(slot) = self.code_formula.as_ref() {
2318 let wants = |label: &str| {
2319 (label == "code" && self.enrich.code) || (label == "formula" && self.enrich.formula)
2320 };
2321 if regions.iter().any(|r| wants(r.label)) {
2322 timing::timed("code_formula", || {
2323 let mut guard = slot.lock().unwrap();
2324 if matches!(*guard, EnrichSlot::Unloaded) {
2325 *guard = match enrich::CodeFormula::load_with(intra_threads()) {
2326 Some(m) => EnrichSlot::Ready(m),
2327 None => EnrichSlot::Missing,
2328 };
2329 }
2330 if let EnrichSlot::Ready(model) = &mut *guard {
2331 for (i, r) in regions.iter().enumerate() {
2332 if !wants(r.label) {
2333 continue;
2334 }
2335 // docling crops the postprocessed cluster box — the
2336 // union of the region's text cells, not the raw
2337 // detector box — expanded by 18% per side, at
2338 // ~120 dpi.
2339 let [bl, bt, br, bb] = assemble::region_cell_bbox(r, &page.cells)
2340 .unwrap_or([r.l, r.t, r.r, r.b]);
2341 let (w, h) = (br - bl, bb - bt);
2342 let ex = enrich::CODE_FORMULA_EXPANSION;
2343 let bbox = [bl - w * ex, bt - h * ex, br + w * ex, bb + h * ex];
2344 let Some(crop) = assemble::crop_region_scaled(
2345 page,
2346 bbox,
2347 enrich::CODE_FORMULA_SCALE,
2348 ) else {
2349 continue;
2350 };
2351 let kind = if r.label == "code" {
2352 enrich::CodeFormulaKind::Code
2353 } else {
2354 enrich::CodeFormulaKind::Formula
2355 };
2356 match model.predict(&crop, kind) {
2357 Ok(text) => {
2358 enrich_out[i] = Some(match kind {
2359 enrich::CodeFormulaKind::Code => {
2360 let (code, language) =
2361 enrich::extract_code_language(&text);
2362 assemble::Enrichment::Code {
2363 language,
2364 text: code,
2365 }
2366 }
2367 enrich::CodeFormulaKind::Formula => {
2368 assemble::Enrichment::Formula { latex: text }
2369 }
2370 });
2371 }
2372 Err(e) => eprintln!("docling-pdf: page {}: {e}", n + 1),
2373 }
2374 }
2375 }
2376 });
2377 }
2378 }
2379 // Score the final region set (docling assigns layout_score over the
2380 // postprocessed clusters — the page elements assemble_page emits,
2381 // which leave out the regulars nested in a picture as its children).
2382 let parents = assemble::picture_parents(®ions);
2383 let elements: Vec<layout::Region> = regions
2384 .iter()
2385 .zip(&parents)
2386 .filter(|(_, p)| p.is_none())
2387 .map(|(r, _)| r.clone())
2388 .collect();
2389 let conf = quality::page_confidence(parse, &elements, &ocr_confs);
2390 let (nodes, links) = timing::timed("assemble_page", || {
2391 assemble::assemble_page(page, regions, &table_rows, &enrich_out, self.images.scale)
2392 });
2393 Ok((nodes, links, conf, self.page_image(n, page), ocr_cells))
2394 }
2395}
2396
2397#[cfg(feature = "ml")]
2398/// Per-worker ONNX intra-op threads. The layout model is memory-bandwidth bound,
2399/// so on a typical machine two threads per worker (sharing one in-cache copy of
2400/// the weights) extracts more throughput than one fat model or many single-thread
2401/// workers. `DOCLING_RS_PDF_INTRA` overrides for per-machine tuning.
2402fn pdf_intra() -> usize {
2403 if let Some(n) = env::parse::<usize>("DOCLING_RS_PDF_INTRA").filter(|&n| n > 0) {
2404 return n;
2405 }
2406 if intra_threads() >= 2 {
2407 2
2408 } else {
2409 1
2410 }
2411}
2412
2413#[cfg(feature = "ml")]
2414/// How many page-workers to spin up for a multi-page PDF. `DOCLING_RS_PDF_WORKERS`
2415/// overrides; otherwise size the pool so `workers × intra ≈ cores`.
2416///
2417/// The pool scales with the machine (#324 follow-up testing): the old hard cap
2418/// of 4 left most of a many-core box idle — on a 16-core M4 Max, 10 workers
2419/// measured ~1.2× over the capped pool (10.0 → 8.5 s on a 130-page document,
2420/// byte-identical output). The ceiling of 16 is a memory bound, not a
2421/// performance one: each worker holds its own layout/OCR sessions (~0.4 GB),
2422/// so a worst-case pool stays under ~6.5 GB even on a ≥32-core host — and
2423/// docling-serve's per-request pools sit behind its `DOCLING_RS_MAX_MEMORY_MB`
2424/// admission control besides. Machines with 4 or fewer effective threads keep
2425/// the exact old sizing (`threads / intra`, min 1).
2426fn pdf_worker_count() -> usize {
2427 if let Some(n) = env::parse::<usize>("DOCLING_RS_PDF_WORKERS").filter(|&n| n > 0) {
2428 return n;
2429 }
2430 // One GPU behind every worker (#842): the pool multiplies the sessions,
2431 // and on the device that is memory, not speed — the serial primary's
2432 // render-ahead already overlaps rasterization with inference.
2433 if docling_onnx::prefers_single_worker() {
2434 return 1;
2435 }
2436 (intra_threads() / pdf_intra()).clamp(1, 16)
2437}
2438
2439#[cfg(feature = "ml")]
2440/// Max pages a worker layout-detects with one batched inference call (issue
2441/// #73). Workers drain the work channel opportunistically up to this size —
2442/// whatever is already rendered gets batched, so batching never *waits* for
2443/// pages and adds no latency when rendering is the bottleneck.
2444///
2445/// Default: per-page (1) on the CPU provider, 4 when a CUDA-class GPU
2446/// provider is selected (#338) — CoreML stays per-page (#602,
2447/// [`docling_onnx::prefers_batching`]): only the pinned batch=1 graph is
2448/// static, and static partitions are all CoreML is allowed to take. The old
2449/// "4 on 8+ cores" CPU default was a hypothesis — that single-session
2450/// amortization pays off with a wider thread budget —
2451/// and every actual CPU measurement lands the other way: a 4-core x86 box
2452/// runs the 9-page 2206.01062 fixture in 8.5 s/conv at batch=1 vs 9.3 s at
2453/// batch=4 (re-measured for #338; the original 8.1 vs 9.3 agrees), and the
2454/// issue-#338 report measured batch=1 ~2× faster on a 16-core M4 Max at
2455/// every worker count — batching only adds cache pressure once workers
2456/// saturate the cores. On a GPU the per-call dispatch overhead is real and
2457/// batching amortizes it, so the GPU default stays. Output is bit-identical
2458/// at every batch size, so this is purely a throughput knob.
2459/// `DOCLING_RS_PDF_LAYOUT_BATCH` overrides either way; `1` = per-page.
2460pub(crate) fn pdf_layout_batch() -> usize {
2461 env::parse::<usize>("DOCLING_RS_PDF_LAYOUT_BATCH")
2462 .filter(|&n| n > 0)
2463 .unwrap_or_else(|| {
2464 if docling_onnx::prefers_batching() {
2465 4
2466 } else {
2467 1
2468 }
2469 })
2470}
2471
2472#[cfg(feature = "ml")]
2473/// Pages the serial path renders ahead of the one being processed
2474/// ([`Pipeline::render_ahead`], #682): one in flight on the worker, two
2475/// rendered and waiting — a bound on resident page bitmaps, like the
2476/// pool's channel.
2477const RENDER_AHEAD: usize = 2;
2478
2479#[cfg(feature = "ml")]
2480/// Minimum page count before a PDF goes to the parallel worker pool; below
2481/// it the serial primary runs its model on every core. Four since #682 (was
2482/// six, from the days of the pool's uncached model load): the layout model's
2483/// intra-op scaling is weak (on a 4-core x86 box 4 threads run a page 2.35×
2484/// faster than one, 2 threads 1.6×), so two workers of two threads convert
2485/// pages faster than one worker of four converts them in turn — 4–5-page
2486/// digital documents measured 18% faster than the serial path with its
2487/// render-ahead, the pool's load included (the pool is sized to the page
2488/// count, [`Pipeline::ensure_pool_for`]). Not lower: the text detector a
2489/// scanned page needs scales well with threads, so a 2–3-page scan runs
2490/// 15–20% *slower* split across two-thread workers than on the four-thread
2491/// primary. `DOCLING_RS_PDF_PARALLEL_MIN` overrides.
2492fn pdf_parallel_min() -> usize {
2493 env::parse::<usize>("DOCLING_RS_PDF_PARALLEL_MIN")
2494 .filter(|&n| n > 0)
2495 .unwrap_or(4)
2496}
2497
2498#[cfg(feature = "ml")]
2499/// A reusable PDF pipeline. The **primary** worker runs its models on every core,
2500/// so a single-page / small / image / METS input is converted at full intra-op
2501/// speed with no pool to load. A document with enough pages instead fans out
2502/// across a **pool** of narrower workers processed concurrently. Both load lazily
2503/// and are cached for reuse, so a one-shot conversion only pays for what it uses.
2504pub struct Pipeline {
2505 /// Full-intra worker for the serial path; loaded on first serial use.
2506 primary: Option<Worker>,
2507 /// Narrower workers (≈cores/`target_workers` threads each) for the parallel
2508 /// path; loaded on first multi-page use and cached.
2509 pool: Vec<Worker>,
2510 /// The single TableFormer instance every worker shares (see [`TfSlot`]).
2511 tables: SharedTables,
2512 /// The shared enrichment-model slots (same pattern as [`TfSlot`]).
2513 classifier: SharedClassifier,
2514 code_formula: SharedCodeFormula,
2515 /// Desired pool size for multi-page documents.
2516 target_workers: usize,
2517 /// Page count at/above which the parallel pool is worth its load cost.
2518 parallel_min: usize,
2519 /// Skip loading/running TableFormer; table regions fall back to geometric
2520 /// reconstruction. See [`Pipeline::no_table_former`].
2521 no_table_former: bool,
2522 /// Skip layout, OCR, and TableFormer entirely. See [`Pipeline::no_ocr`].
2523 no_ocr: bool,
2524 /// Per document: the text layer alone when it carries the text, the
2525 /// full pipeline otherwise. See [`Pipeline::text_layer_auto`].
2526 text_layer_auto: bool,
2527 /// Keep layout + TableFormer, never OCR (#244). See [`Pipeline::skip_ocr`].
2528 skip_ocr: bool,
2529 /// OCR every page even when it carries a text layer. See
2530 /// [`Pipeline::force_full_page_ocr`].
2531 force_full_page_ocr: bool,
2532 /// Never demote text-panel pictures. See [`Pipeline::no_text_panels`].
2533 no_text_panels: bool,
2534 /// Opt-in enrichment passes. See [`Pipeline::enrichments`].
2535 enrich: EnrichmentOptions,
2536 /// 1-based inclusive page window to convert. See [`Pipeline::pages`].
2537 page_range: Option<(usize, usize)>,
2538 /// OCR recognition language. See [`Pipeline::ocr_lang`].
2539 ocr_lang: ocr::OcrLang,
2540 /// Which OCR engine runs (#460). See [`Pipeline::ocr_engine`].
2541 ocr_engine: ocr::OcrEngine,
2542 /// Tesseract's languages (#460). See [`Pipeline::tesseract_lang`].
2543 tesseract_lang: Option<String>,
2544 /// Which regions feed the OCR (#254). See [`Pipeline::ocr_mode`].
2545 ocr_mode: ocr::OcrMode,
2546 /// OCR render scale override in px/pt (#254). See [`Pipeline::ocr_scale`].
2547 ocr_scale: Option<f32>,
2548 /// Picture OCR scale (#666). See [`Pipeline::picture_ocr_scale`].
2549 picture_ocr_scale: Option<f32>,
2550 /// The text layer's word-gap rule (#921); `None` = the environment's
2551 /// (`DOCLING_RS_PDF_WORD_GAP`), else [`WordGap::Adaptive`].
2552 word_gap: Option<WordGap>,
2553 /// Keep typographic punctuation as drawn (#921); `None` = the
2554 /// environment's (`DOCLING_RS_PDF_KEEP_TYPOGRAPHY`), else folding.
2555 keep_typography: Option<bool>,
2556 /// RapidOCR's line orientation handling (#666). See
2557 /// [`Pipeline::ocr_textline_orientation`].
2558 textline_orientation: bool,
2559 /// Picture-crop scale and page images (#519/#520). See
2560 /// [`Pipeline::images_scale`] / [`Pipeline::generate_page_images`].
2561 images: ImageOutput,
2562 /// Heading-level inference (#302). See [`Pipeline::heading_hierarchy`].
2563 heading_hierarchy: HeadingHierarchyOptions,
2564 /// Optional per-page progress hook, invoked after each page finishes on
2565 /// every walk — the serial and parallel buffered paths, both streaming
2566 /// paths and the pre-rendered page path (images, METS). The converter
2567 /// turns it into its progress events (#783); `None` costs nothing.
2568 progress: Option<PageProgressHook>,
2569 /// Per-document wall-clock budget (docling's `document_timeout`, #497).
2570 /// See [`Pipeline::document_timeout`].
2571 document_timeout: Option<std::time::Duration>,
2572}
2573
2574#[cfg(feature = "ml")]
2575impl Pipeline {
2576 /// Construct the pipeline. Models load lazily on first use (full-intra primary
2577 /// for serial inputs, the helper pool for multi-page PDFs), so nothing is
2578 /// loaded that a given document doesn't need.
2579 pub fn new() -> Result<Self, PdfError> {
2580 Ok(Self {
2581 primary: None,
2582 pool: Vec::new(),
2583 tables: Arc::new(Mutex::new(TfSlot::Unloaded)),
2584 classifier: Arc::new(Mutex::new(EnrichSlot::Unloaded)),
2585 code_formula: Arc::new(Mutex::new(EnrichSlot::Unloaded)),
2586 target_workers: pdf_worker_count(),
2587 parallel_min: pdf_parallel_min(),
2588 no_table_former: false,
2589 no_ocr: false,
2590 text_layer_auto: false,
2591 skip_ocr: false,
2592 force_full_page_ocr: false,
2593 no_text_panels: false,
2594 enrich: EnrichmentOptions::default(),
2595 page_range: None,
2596 ocr_lang: ocr::OcrLang::from_env(),
2597 ocr_engine: ocr::OcrEngine::from_env(),
2598 tesseract_lang: None,
2599 ocr_mode: ocr::OcrMode::from_env(),
2600 ocr_scale: ocr::scale_from_env(),
2601 picture_ocr_scale: ocr::picture_scale_from_env(),
2602 word_gap: None,
2603 keep_typography: None,
2604 textline_orientation: ocr::textline_orientation_from_env(),
2605 images: ImageOutput::default(),
2606 heading_hierarchy: HeadingHierarchyOptions::default(),
2607 progress: None,
2608 document_timeout: None,
2609 })
2610 }
2611
2612 /// A wall-clock budget for each document (docling's
2613 /// `PipelineOptions.document_timeout`, #497; `None` = unlimited, the
2614 /// default). The budget starts when a conversion starts and is checked
2615 /// cooperatively between pages: once it is spent, no further page is
2616 /// rendered or processed, the pages already finished are assembled into
2617 /// the document, and the outcome says so (`Completion::TimedOut`, which
2618 /// the converter reports as `PartialSuccess` with a timeout error, the
2619 /// way docling does). A page in flight finishes — the check costs
2620 /// nothing inside a page, and a single-page document (an image) is never
2621 /// cut. For a long-lived pipeline use
2622 /// [`set_document_timeout`](Self::set_document_timeout) before each
2623 /// conversion.
2624 pub fn document_timeout(mut self, timeout: Option<std::time::Duration>) -> Self {
2625 self.document_timeout = timeout;
2626 self
2627 }
2628
2629 /// In-place variant of [`document_timeout`](Self::document_timeout) for a
2630 /// warm pipeline — set it before every conversion so no request inherits
2631 /// a previous one's budget.
2632 pub fn set_document_timeout(&mut self, timeout: Option<std::time::Duration>) {
2633 self.document_timeout = timeout;
2634 }
2635
2636 /// The deadline of a conversion starting now, from the configured budget.
2637 fn deadline(&self) -> Option<std::time::Instant> {
2638 self.document_timeout.map(|t| std::time::Instant::now() + t)
2639 }
2640
2641 /// The [`Completion`] of a walk that processed `done` of `selected` pages.
2642 fn completion(&self, timed_out: bool, done: usize, selected: usize) -> Completion {
2643 match (timed_out, self.document_timeout) {
2644 (true, Some(budget)) => Completion::TimedOut {
2645 pages_done: done,
2646 pages_selected: selected,
2647 budget,
2648 },
2649 _ => Completion::Complete,
2650 }
2651 }
2652
2653 /// Infer section-header levels after assembly (#302, docling's
2654 /// `HeadingHierarchyModel`): PDF bookmarks > legal/outline numbering >
2655 /// font style, off by default — see [`HeadingHierarchyOptions`]. Pure
2656 /// post-processing configuration; for a warm pipeline use
2657 /// [`set_heading_hierarchy`](Self::set_heading_hierarchy).
2658 pub fn heading_hierarchy(mut self, opts: HeadingHierarchyOptions) -> Self {
2659 self.heading_hierarchy = opts;
2660 self
2661 }
2662
2663 /// In-place variant of [`heading_hierarchy`](Self::heading_hierarchy) for
2664 /// a long-lived pipeline (docling-serve's warm instance) — like
2665 /// [`set_pages`](Self::set_pages), set it before every conversion so no
2666 /// request inherits a previous one's choice.
2667 pub fn set_heading_hierarchy(&mut self, opts: HeadingHierarchyOptions) {
2668 self.heading_hierarchy = opts;
2669 }
2670
2671 /// Run the enabled heading-hierarchy stage (#302) on an assembled
2672 /// document: gather the outline (bookmarks) and the per-page glyph styles
2673 /// on demand, then assign levels in place. `bytes` is `None` on paths
2674 /// with no PDF behind them (standalone images, METS) — those degrade to
2675 /// the numbering signal, exactly like docling without parsed pages.
2676 fn apply_heading_hierarchy(&self, nodes: &mut [Node], bytes: Option<&[u8]>) {
2677 let opts = &self.heading_hierarchy;
2678 if !opts.enabled {
2679 return;
2680 }
2681 let outline = match bytes {
2682 Some(bytes) if opts.use_bookmarks => outline::extract_outline(bytes),
2683 _ => Vec::new(),
2684 };
2685 let styles = match bytes {
2686 Some(bytes) if opts.use_style => {
2687 let pages = heading_hierarchy::heading_pages(nodes);
2688 textparse::glyph_styles(bytes, &pages)
2689 }
2690 _ => Default::default(),
2691 };
2692 heading_hierarchy::apply(nodes, &outline, &styles, opts);
2693 }
2694
2695 /// Install (or clear) the per-page progress hook: called with a
2696 /// [`PageProgress`] after each page completes, on every conversion entry
2697 /// point ([`convert`](Self::convert), the streaming variants,
2698 /// [`convert_image`](Self::convert_image), [`process_pages`](Self::process_pages)).
2699 /// Shared with the parallel workers, so the callback must be cheap and
2700 /// thread-safe. A page the budget cut off is never reported here — the
2701 /// walk's [`Completion`] says how many were; the converter reports those
2702 /// as failed pages (#783).
2703 pub fn set_progress(&mut self, cb: Option<PageProgressHook>) {
2704 self.progress = cb;
2705 }
2706
2707 /// Convert only pages `first..=last` (**1-based**, like the page numbers a
2708 /// PDF viewer shows — issue #80's `--pages A-B`). Out-of-range pages are
2709 /// skipped before rasterization, so the cost is proportional to the window,
2710 /// not the document. `last` past the end of the document clamps; a window
2711 /// that selects no pages at all (`first` beyond the last page) is an error
2712 /// at convert time. `None` (the default) converts everything.
2713 pub fn pages(mut self, range: Option<(usize, usize)>) -> Self {
2714 self.page_range = range;
2715 self
2716 }
2717
2718 /// In-place variant of [`pages`](Self::pages) for a long-lived pipeline
2719 /// (e.g. docling-serve's warm instance) that applies a per-request window
2720 /// without rebuilding — unlike the model switches, the window is pure
2721 /// configuration. Set it before every conversion; it stays until changed.
2722 pub fn set_pages(&mut self, range: Option<(usize, usize)>) {
2723 self.page_range = range;
2724 }
2725
2726 /// OCR recognition language (see [`OcrLang`]): English by default, `ch`
2727 /// for the multilingual docling-conformance model. `None` keeps the
2728 /// process default (`DOCLING_RS_OCR_LANG`, else English). Set before the
2729 /// first conversion; for a warm pipeline use
2730 /// [`set_ocr_lang`](Self::set_ocr_lang).
2731 pub fn ocr_lang(mut self, lang: Option<ocr::OcrLang>) -> Self {
2732 self.set_ocr_lang(lang);
2733 self
2734 }
2735
2736 /// In-place variant of [`ocr_lang`](Self::ocr_lang) for a long-lived
2737 /// pipeline (docling-serve's warm instance). Unlike the page window this
2738 /// is a *model* switch: any worker whose cached recognition model was
2739 /// loaded for a different language drops it, to be lazily reloaded on the
2740 /// next OCR-needing page (cheap — the rec models are ~10 MB).
2741 pub fn set_ocr_lang(&mut self, lang: Option<ocr::OcrLang>) {
2742 let lang = lang.unwrap_or_else(ocr::OcrLang::from_env);
2743 self.ocr_lang = lang;
2744 for worker in self.primary.iter_mut().chain(self.pool.iter_mut()) {
2745 if worker.ocr_lang != lang {
2746 worker.ocr_lang = lang;
2747 worker.ocr = OcrSlot::Unloaded;
2748 }
2749 }
2750 }
2751
2752 /// Which OCR engine recognizes text (#460, see [`OcrEngine`]): the
2753 /// built-in PP-OCRv3 recognizer by default, or the system `tesseract`
2754 /// binary (docling's `TesseractCliOcrOptions`) — same layout-region
2755 /// crops, same cells; Tesseract does its own line/word segmentation and
2756 /// reads orientation from its OSD. `None` keeps the process default
2757 /// (`DOCLING_RS_OCR_ENGINE`, else PP-OCR). Set before the first
2758 /// conversion; for a warm pipeline use
2759 /// [`set_ocr_engine`](Self::set_ocr_engine).
2760 pub fn ocr_engine(mut self, engine: Option<ocr::OcrEngine>) -> Self {
2761 self.set_ocr_engine(engine);
2762 self
2763 }
2764
2765 /// In-place variant of [`ocr_engine`](Self::ocr_engine): a model switch
2766 /// like [`set_ocr_lang`](Self::set_ocr_lang) — workers holding the other
2767 /// engine drop it, to be lazily reloaded.
2768 pub fn set_ocr_engine(&mut self, engine: Option<ocr::OcrEngine>) {
2769 let engine = engine.unwrap_or_else(ocr::OcrEngine::from_env);
2770 self.ocr_engine = engine;
2771 for worker in self.primary.iter_mut().chain(self.pool.iter_mut()) {
2772 if worker.ocr_engine != engine {
2773 worker.ocr_engine = engine;
2774 worker.ocr = OcrSlot::Unloaded;
2775 }
2776 }
2777 }
2778
2779 /// Tesseract's languages (#460): the `-l` argument, tessdata stems joined
2780 /// with `+` — build it from an `ocr_lang` value with
2781 /// [`tesseract_lang_arg`], which also maps BCP-47 tags and the PP-OCR
2782 /// codes. `None` runs Tesseract's default (`eng`). Ignored under the
2783 /// PP-OCR engine, whose language is [`ocr_lang`](Self::ocr_lang). For a
2784 /// warm pipeline use [`set_tesseract_lang`](Self::set_tesseract_lang).
2785 pub fn tesseract_lang(mut self, lang: Option<String>) -> Self {
2786 self.set_tesseract_lang(lang);
2787 self
2788 }
2789
2790 /// In-place variant of [`tesseract_lang`](Self::tesseract_lang): a model
2791 /// switch like [`set_ocr_lang`](Self::set_ocr_lang) — a worker whose
2792 /// Tesseract was probed for other languages drops it.
2793 pub fn set_tesseract_lang(&mut self, lang: Option<String>) {
2794 self.tesseract_lang = lang.clone();
2795 for worker in self.primary.iter_mut().chain(self.pool.iter_mut()) {
2796 if worker.tesseract_lang != lang {
2797 worker.tesseract_lang = lang.clone();
2798 if worker.ocr_engine == ocr::OcrEngine::Tesseract {
2799 worker.ocr = OcrSlot::Unloaded;
2800 }
2801 }
2802 }
2803 }
2804
2805 /// Resolve the configured 1-based window against a page count into the
2806 /// 0-based inclusive form the backend walks, validating it selects at
2807 /// least one existing page.
2808 fn resolve_range(&self, total: usize) -> Result<Option<(usize, usize)>, PdfError> {
2809 let Some((first, last)) = self.page_range else {
2810 return Ok(None);
2811 };
2812 if first == 0 || last < first {
2813 return Err(PdfError::Pdfium(format!(
2814 "invalid page range {first}-{last} (pages are 1-based, first <= last)"
2815 )));
2816 }
2817 if first > total {
2818 return Err(PdfError::Pdfium(format!(
2819 "page range {first}-{last} is outside the document ({total} page(s))"
2820 )));
2821 }
2822 Ok(Some((first - 1, last.min(total) - 1)))
2823 }
2824
2825 /// Enable the opt-in enrichment passes (docling's
2826 /// `do_picture_classification` / `do_code_enrichment` /
2827 /// `do_formula_enrichment`). Each enabled pass lazily loads its model on
2828 /// the first matching region; a missing model warns once and is skipped.
2829 /// Set before the first conversion (no effect on already-loaded workers).
2830 pub fn enrichments(mut self, opts: EnrichmentOptions) -> Self {
2831 self.enrich = opts;
2832 self
2833 }
2834
2835 /// Skip loading and running the TableFormer table-structure model. Table
2836 /// regions still get emitted, but reconstructed geometrically from cell
2837 /// positions instead of via the ONNX model's predicted structure — faster
2838 /// (no model load, no per-table inference) at the cost of table fidelity.
2839 /// No effect if a worker is already loaded; set this before the first
2840 /// conversion.
2841 pub fn no_table_former(mut self, disable: bool) -> Self {
2842 self.no_table_former = disable;
2843 self
2844 }
2845
2846 /// Keep every detected `picture` region as a picture. By default an
2847 /// *uncaptioned* picture that reads like a dense, uniform text panel (a
2848 /// terms-and-conditions box exported as an image) is demoted into
2849 /// paragraphs (#157); a chart the layout mislabels can still trip that
2850 /// heuristic on scanned pages, and image-extraction workflows may simply
2851 /// want every crop — this flag disables the demotion entirely (#173).
2852 /// No effect on already-loaded workers; set before the first conversion.
2853 pub fn no_text_panels(mut self, disable: bool) -> Self {
2854 self.no_text_panels = disable;
2855 self
2856 }
2857
2858 /// Skip layout detection, OCR, and TableFormer entirely — no model load, no
2859 /// inference of any kind. The PDF's embedded text cells are grouped by line
2860 /// and emitted as plain paragraphs in reading order: no headings, lists,
2861 /// tables, code blocks, or pictures, since that structure comes from the
2862 /// layout model. The fastest possible PDF path, but pages with no embedded
2863 /// text layer (scanned/image-only PDFs) yield no text at all — convert those
2864 /// without this flag. Implies `no_table_former`. A PDF conversion reads
2865 /// the flag per call (its pages never touch a worker); image and METS
2866 /// pages read the flag the workers were loaded with, so set it before
2867 /// the first conversion for those.
2868 pub fn no_ocr(mut self, disable: bool) -> Self {
2869 self.no_ocr = disable;
2870 self
2871 }
2872
2873 /// Decide per document between the two: a PDF whose embedded text layer
2874 /// carries its text converts as under [`no_ocr`](Self::no_ocr) — the
2875 /// text layer alone, flat paragraphs, no model loaded, no inference —
2876 /// and one without a usable layer (scanned / image-only, or a vestigial
2877 /// layer of a few typed-in form fields; [`text_layer_has_text`]) runs the
2878 /// full pipeline, which OCRs its pages. The verdict is the document's,
2879 /// not a page's: a digital file with a few scanned pages inserted keeps
2880 /// the text path, and those pages come back empty as under `no_ocr`.
2881 /// `no_ocr` wins when both are set; forcing (`force_full_page_ocr`,
2882 /// `ocr_mode`) is read only on the pipeline side of the verdict, as
2883 /// under `no_ocr`. A per-conversion setting: the workers and their
2884 /// models are untouched, so a warm pipeline serving digital and scanned
2885 /// documents in turn reloads nothing.
2886 pub fn text_layer_auto(mut self, enable: bool) -> Self {
2887 self.text_layer_auto = enable;
2888 self
2889 }
2890
2891 /// Set [`text_layer_auto`](Self::text_layer_auto) on a built pipeline.
2892 pub fn set_text_layer_auto(&mut self, enable: bool) {
2893 self.text_layer_auto = enable;
2894 }
2895
2896 /// Never run OCR, but keep layout detection and TableFormer — docling's
2897 /// independent `do_ocr=False` (#244), the counterpart of
2898 /// [`no_table_former`](Self::no_table_former). Unlike
2899 /// [`no_ocr`](Self::no_ocr) (which skips the whole ML stack), structured
2900 /// output — headings, tables, pictures, reading order — is preserved;
2901 /// only text that exists solely as pixels is lost: scanned pages come
2902 /// back with their regions empty, and the speculative OCR of large
2903 /// embedded images never runs. The OCR model is never loaded. Ignored
2904 /// when `no_ocr` is set (there is no OCR to skip);
2905 /// takes precedence over [`force_full_page_ocr`](Self::force_full_page_ocr),
2906 /// mirroring docling where forcing is a sub-option of `do_ocr`.
2907 pub fn skip_ocr(mut self, disable: bool) -> Self {
2908 self.skip_ocr = disable;
2909 self
2910 }
2911
2912 /// OCR every page from its rendered image even when the page carries an
2913 /// embedded text layer — docling's `force_full_page_ocr`. The escape hatch
2914 /// for text layers that exist but lie: broken encodings, subset fonts with
2915 /// garbage mappings, a scanned form with a few typed-in fields. Ignored
2916 /// when [`no_ocr`](Self::no_ocr) is set, mirroring docling (there
2917 /// `force_full_page_ocr` is a sub-option of `do_ocr`).
2918 pub fn force_full_page_ocr(mut self, force: bool) -> Self {
2919 self.force_full_page_ocr = force;
2920 self
2921 }
2922
2923 /// Which document regions feed the OCR — docling's `OcrMode` (#254). The
2924 /// default (`default` = `pdf_aware_layout_regions`) is the standard
2925 /// text-layer-aware behavior; `full_page`/`layout_regions` discard the
2926 /// text layer like [`force_full_page_ocr`](Self::force_full_page_ocr)
2927 /// (see [`ocr::OcrMode`] for why both map onto it). Whichever of the flag
2928 /// and the mode demands forcing wins, mirroring docling's
2929 /// `force_full_page_ocr` → `mode=full_page` bridge. `None` keeps the
2930 /// process default (`DOCLING_RS_OCR_MODE`, else `default`).
2931 pub fn ocr_mode(mut self, mode: Option<ocr::OcrMode>) -> Self {
2932 self.ocr_mode = mode.unwrap_or_else(ocr::OcrMode::from_env);
2933 self
2934 }
2935
2936 /// In-place variants of [`force_full_page_ocr`](Self::force_full_page_ocr),
2937 /// [`ocr_mode`](Self::ocr_mode) and [`ocr_scale`](Self::ocr_scale) for a
2938 /// long-lived pipeline (docling-serve's warm instance): all three are pure
2939 /// per-worker configuration — no model reloads — so they apply per request
2940 /// like [`set_pages`](Self::set_pages). Set them before every conversion so
2941 /// no request inherits a previous one's choice.
2942 pub fn set_force_full_page_ocr(&mut self, force: bool) {
2943 self.force_full_page_ocr = force;
2944 self.sync_ocr_config();
2945 }
2946
2947 /// See [`set_force_full_page_ocr`](Self::set_force_full_page_ocr).
2948 pub fn set_ocr_mode(&mut self, mode: Option<ocr::OcrMode>) {
2949 self.ocr_mode = mode.unwrap_or_else(ocr::OcrMode::from_env);
2950 self.sync_ocr_config();
2951 }
2952
2953 /// How the text layer splits words where the PDF draws no space glyph
2954 /// (#921): docling-parse ≥ 7's line-local [`WordGap::Adaptive`]
2955 /// threshold (the default), a fixed [`WordGap::Factor`] of the two
2956 /// flanking glyphs' mean advance, or [`WordGap::Legacy`] (docling-parse
2957 /// ≤ 6's running-mean rule). `None` keeps `DOCLING_RS_PDF_WORD_GAP`,
2958 /// else the default.
2959 pub fn set_word_gap(&mut self, gap: Option<WordGap>) {
2960 self.word_gap = gap;
2961 }
2962
2963 /// See [`set_force_full_page_ocr`](Self::set_force_full_page_ocr).
2964 pub fn set_ocr_scale(&mut self, scale: Option<f32>) {
2965 self.ocr_scale = scale
2966 .filter(|s| s.is_finite() && *s > 0.0)
2967 .or_else(ocr::scale_from_env);
2968 self.sync_ocr_config();
2969 }
2970
2971 /// Resolution of the picture-OCR enrichment (#645) in pixels per
2972 /// picture pixel, before RapidOCR's 2000 px longer-side cap (#666) —
2973 /// RapidOCR reads a picture at its native pixels (`1.0`) and only
2974 /// shrinks what exceeds 2000 px. `None` (default:
2975 /// `DOCLING_RS_PICTURE_OCR_SCALE`, else unset) keeps the image-input
2976 /// rule pictures shared with scanned images: [`ocr_scale`](Self::ocr_scale)
2977 /// when set, else 3 px/pt shrunk to the 2000 px cap — which upscales a
2978 /// 720 × 180 screenshot to 2000 × 500. Pages and image inputs are
2979 /// unaffected. Non-positive values are ignored.
2980 pub fn picture_ocr_scale(mut self, scale: Option<f32>) -> Self {
2981 self.set_picture_ocr_scale(scale);
2982 self
2983 }
2984
2985 /// In-place [`picture_ocr_scale`](Self::picture_ocr_scale) for a warm
2986 /// pipeline.
2987 pub fn set_picture_ocr_scale(&mut self, scale: Option<f32>) {
2988 self.picture_ocr_scale = scale
2989 .filter(|s| s.is_finite() && *s > 0.0)
2990 .or_else(ocr::picture_scale_from_env);
2991 }
2992
2993 /// RapidOCR's handling of text-line orientation (#666), off by default
2994 /// (`DOCLING_RS_OCR_TEXTLINE_ORIENTATION` sets the process default): each
2995 /// detected line is cropped along the detector's rotated rectangle (a
2996 /// perspective warp, `get_rotate_crop_image`) instead of its axis-aligned
2997 /// box, a crop at least 1.5× taller than wide is turned 90°
2998 /// counter-clockwise, and the PP-LCNet text-line orientation classifier
2999 /// (`.models/ocr_textline_ori.onnx`, `DOCLING_OCR_TEXTLINE_ORI_ONNX`)
3000 /// turns lines it reads as upside down (`180` above 0.9). Vertical axis
3001 /// titles, rotated headers, skewed and upside-down lines then read as
3002 /// text. Applies to every PP-OCR line read off the detector — pages and
3003 /// pictures alike — but not to table words or the projection fallback;
3004 /// the reported boxes stay axis-aligned. A missing classifier warns once
3005 /// and keeps the crop and the 90° turn. Tesseract orients on its own.
3006 pub fn ocr_textline_orientation(mut self, enable: bool) -> Self {
3007 self.set_ocr_textline_orientation(Some(enable));
3008 self
3009 }
3010
3011 /// In-place [`ocr_textline_orientation`](Self::ocr_textline_orientation)
3012 /// for a warm pipeline (serve, the bindings' `Pipeline`), applied to
3013 /// loaded workers too; `None` restores the process default
3014 /// (`DOCLING_RS_OCR_TEXTLINE_ORIENTATION`), like
3015 /// [`set_ocr_scale`](Self::set_ocr_scale).
3016 pub fn set_ocr_textline_orientation(&mut self, enable: Option<bool>) {
3017 self.textline_orientation = enable.unwrap_or_else(ocr::textline_orientation_from_env);
3018 self.sync_ocr_config();
3019 }
3020
3021 /// Whether page extraction should decode the text layer at all. Forced
3022 /// full-page OCR (the flag or `ocr_mode=full_page|layout_regions`) clears
3023 /// every extracted cell unread, so the decode is skipped outright —
3024 /// docling#4061's `skip_cell_extraction` (2.122). `no_ocr` wins over the
3025 /// forcing, as everywhere else: its fast path *is* the text layer.
3026 fn extract_text_layer(&self) -> bool {
3027 self.no_ocr || !(self.force_full_page_ocr || self.ocr_mode.forces_full_page())
3028 }
3029
3030 /// When the walk renders the 2x bitmaps of pages with a text layer
3031 /// (#897). Adaptively by default: while the document's recent pages read
3032 /// theirs, each starts on the helper thread as the page leaves the walk,
3033 /// overlapping layout (#682); while they do not (prose), only a page that
3034 /// reads it renders it, after layout. Page images read every bitmap, so
3035 /// they keep the #682 schedule. `DOCLING_RS_PAGE_BITMAP` picks one
3036 /// explicitly: `eager` (every bitmap, during layout), `adaptive`,
3037 /// `on-demand` (only the bitmaps read — the least CPU, with every render
3038 /// after layout).
3039 fn bitmap_render(&self) -> pdfium_backend::BitmapRender {
3040 use pdfium_backend::BitmapRender;
3041 match env::nonempty("DOCLING_RS_PAGE_BITMAP")
3042 .map(|v| v.to_ascii_lowercase())
3043 .as_deref()
3044 {
3045 Some("eager") => BitmapRender::Eager,
3046 Some("on-demand" | "ondemand" | "lazy") => BitmapRender::OnDemand,
3047 _ if self.images.page_images => BitmapRender::Eager,
3048 _ => BitmapRender::Adaptive,
3049 }
3050 }
3051
3052 /// Push the current OCR forcing/scale choice onto already-loaded workers
3053 /// (new workers read it at [`Worker::load`]).
3054 fn sync_ocr_config(&mut self) {
3055 let force = self.force_full_page_ocr || self.ocr_mode.forces_full_page();
3056 let scale = self.ocr_scale;
3057 let orient = self.textline_orientation;
3058 for worker in self.primary.iter_mut().chain(self.pool.iter_mut()) {
3059 worker.force_full_page_ocr = force;
3060 worker.ocr_scale = scale;
3061 worker.set_textline_orientation(orient);
3062 }
3063 }
3064
3065 /// Pixels per PDF point for picture crops and page images — docling's
3066 /// `images_scale` (#519/#520). `None` (the default) delivers crops at the
3067 /// pipeline's own page render, 2.0 px/pt (144 dpi), untouched. Another
3068 /// value resamples that render (CatmullRom, docling's PIL BICUBIC), so
3069 /// values above 2.0 upsample rather than re-render. Each picture's
3070 /// `dpi` (docling-core's `ImageRef.dpi`) is 72·scale, so consumers mapping
3071 /// pixels back to points stay exact. Layout, OCR and TableFormer inputs
3072 /// are unaffected. Non-positive values are ignored.
3073 pub fn images_scale(mut self, scale: Option<f32>) -> Self {
3074 self.set_images(ImageOutput {
3075 scale,
3076 ..self.images
3077 });
3078 self
3079 }
3080
3081 /// Keep each page's render as the document's page image — docling's
3082 /// `generate_page_images` (#520): the JSON export writes it as the page's
3083 /// `image` (docling-core's `PageItem.image`, at [`Self::images_scale`],
3084 /// `dpi` = 72·scale), which is what docling-core's
3085 /// `TableItem.get_image` / `FormulaItem.get_image` crop from. Off by
3086 /// default: a page image is a full-page PNG per page held in memory.
3087 /// Pages converted with `no_ocr` (text layer only) are never rendered and
3088 /// get none; streaming conversions carry no page map and drop them.
3089 pub fn generate_page_images(mut self, enabled: bool) -> Self {
3090 self.set_images(ImageOutput {
3091 page_images: enabled,
3092 ..self.images
3093 });
3094 self
3095 }
3096
3097 /// Both image outputs at once, also on already-loaded workers — the
3098 /// per-request form a warm pipeline (serve, the Node `Pipeline`) uses,
3099 /// like [`set_ocr_scale`](Self::set_ocr_scale). A non-finite or
3100 /// non-positive scale reads as unset.
3101 pub fn set_images(&mut self, images: ImageOutput) {
3102 self.images = ImageOutput {
3103 scale: images.scale.filter(|s| s.is_finite() && *s > 0.0),
3104 ..images
3105 };
3106 let images = self.images;
3107 for worker in self.primary.iter_mut().chain(self.pool.iter_mut()) {
3108 worker.images = images;
3109 }
3110 }
3111
3112 /// OCR render scale in pixels per PDF point — docling's `OcrOptions.scale`
3113 /// (#254, upstream docling#3877; their default 3 = 216 dpi). `None`
3114 /// (default: `DOCLING_RS_OCR_SCALE`, else unset) feeds the recognizer the
3115 /// pipeline's own page render (2.0 px/pt = 144 dpi) for a rendered page,
3116 /// and for a standalone image input docling's own resolution (#570): 3
3117 /// px/pt shrunk so the longer side stays within RapidOCR's 2000 px — what
3118 /// its models see after RapidOCR's `max_side_len` pass (a 754 × 1000 scan
3119 /// reads at 2.0). A set value resamples the render / image for the OCR
3120 /// input only — layout and TableFormer keep their pinned-resolution
3121 /// pixels, so the conformance baseline never moves. Lower it when the
3122 /// source raster is already high-resolution and upscaling degrades
3123 /// recognition; raise it toward docling's 216 dpi for parity experiments.
3124 /// Non-positive values are ignored.
3125 pub fn ocr_scale(mut self, scale: Option<f32>) -> Self {
3126 self.ocr_scale = scale
3127 .filter(|s| s.is_finite() && *s > 0.0)
3128 .or_else(ocr::scale_from_env);
3129 self
3130 }
3131
3132 /// Keep typographic punctuation as drawn — `„ “ ‚ ‘`, dashes, `…` —
3133 /// instead of folding it to ASCII as docling-parse's sanitizer does
3134 /// (`pdf_keep_typography`, #921). `None` keeps
3135 /// `DOCLING_RS_PDF_KEEP_TYPOGRAPHY`, else folding.
3136 pub fn set_keep_typography(&mut self, keep: Option<bool>) {
3137 self.keep_typography = keep;
3138 }
3139
3140 /// See [`set_keep_typography`](Self::set_keep_typography).
3141 pub fn keep_typography(mut self, keep: Option<bool>) -> Self {
3142 self.keep_typography = keep;
3143 self
3144 }
3145
3146 /// The text-layer settings the page walk applies.
3147 fn text_options(&self) -> pdfium_backend::TextOptions {
3148 pdfium_backend::TextOptions {
3149 word_gap: self.word_gap,
3150 keep_typography: self
3151 .keep_typography
3152 .unwrap_or_else(assemble::keep_typography_from_env),
3153 }
3154 }
3155
3156 /// See [`set_word_gap`](Self::set_word_gap).
3157 pub fn word_gap(mut self, gap: Option<WordGap>) -> Self {
3158 self.word_gap = gap;
3159 self
3160 }
3161
3162 /// The shared TableFormer slot handed to each worker, or `None` when the
3163 /// pipeline options skip TableFormer entirely.
3164 fn tables_slot(&self) -> Option<SharedTables> {
3165 if self.no_table_former || self.no_ocr {
3166 None
3167 } else {
3168 Some(Arc::clone(&self.tables))
3169 }
3170 }
3171
3172 /// The shared enrichment slots for a worker (`None` per model unless its
3173 /// flag is on; `no_ocr` skips layout, so there are no regions to enrich).
3174 fn enrich_slots(&self) -> (Option<SharedClassifier>, Option<SharedCodeFormula>) {
3175 if self.no_ocr || !self.enrich.any() {
3176 return (None, None);
3177 }
3178 (
3179 self.enrich
3180 .picture_classification
3181 .then(|| Arc::clone(&self.classifier)),
3182 (self.enrich.code || self.enrich.formula).then(|| Arc::clone(&self.code_formula)),
3183 )
3184 }
3185
3186 /// Eagerly load the models (the full-intra serial worker: layout + OCR, and
3187 /// the shared TableFormer unless disabled) so the first conversion doesn't pay
3188 /// the load cost. Idempotent; respects `no_ocr` / `no_table_former` (with
3189 /// `no_ocr` there is nothing to load). The docling.rs analogue of docling's
3190 /// `DocumentConverter.initialize_pipeline`.
3191 pub fn warm_up(&mut self) -> Result<(), PdfError> {
3192 self.primary()?;
3193 // The shared TableFormer loads on the first table otherwise — a
3194 // document with one would still pay for it (#548).
3195 if let Some(tables) = self.tables_slot() {
3196 tables.lock().unwrap_or_else(|p| p.into_inner()).load();
3197 }
3198 Ok(())
3199 }
3200
3201 /// [`warm_up`](Self::warm_up) plus the parallel page-worker pool, so the
3202 /// first *multi-page* conversion is no slower than later ones either: a
3203 /// document of `DOCLING_RS_PDF_PARALLEL_MIN` pages (6) or more fans out across the pool,
3204 /// whose workers each load their own layout/OCR sessions on first use
3205 /// (#548 — docling-serve's `--warmup` left a 9-page PDF paying that load).
3206 /// Costs the pool's memory up front (~0.4 GB a worker); with a pool of one
3207 /// worker it is exactly `warm_up`.
3208 pub fn warm_up_all(&mut self) -> Result<(), PdfError> {
3209 self.warm_up()?;
3210 if self.target_workers >= 2 {
3211 self.ensure_pool()?;
3212 }
3213 Ok(())
3214 }
3215
3216 /// The full-intra serial worker, loaded on first use.
3217 fn primary(&mut self) -> Result<&mut Worker, PdfError> {
3218 if self.primary.is_none() {
3219 self.primary = Some(Worker::load(
3220 intra_threads(),
3221 self.tables_slot(),
3222 self.enrich_slots(),
3223 self.enrich,
3224 self.no_ocr,
3225 self.skip_ocr,
3226 // The mode-shaped spelling (#254) and the flag are one engine
3227 // truth: whichever demands forcing wins, mirroring docling's
3228 // `force_full_page_ocr` → `mode=full_page` bridge.
3229 self.force_full_page_ocr || self.ocr_mode.forces_full_page(),
3230 self.no_text_panels,
3231 self.ocr_lang,
3232 self.ocr_engine,
3233 self.tesseract_lang.clone(),
3234 self.ocr_scale,
3235 self.textline_orientation,
3236 self.images,
3237 )?);
3238 }
3239 Ok(self.primary.as_mut().unwrap())
3240 }
3241
3242 /// Convert a PDF (bytes) to a [`DoclingDocument`]. A document with fewer than
3243 /// `parallel_min` pages (or a pool size of 1) streams through the full-intra
3244 /// primary; a larger one renders on this thread (pdfium is not thread-safe) and
3245 /// fans the pages out across the worker pool, reassembled in page order so the
3246 /// output is byte-identical to the serial path.
3247 pub fn convert(
3248 &mut self,
3249 bytes: &[u8],
3250 password: Option<&str>,
3251 name: &str,
3252 ) -> Result<DoclingDocument, PdfError> {
3253 self.convert_outcome(bytes, password, name)
3254 .map(|c| c.document)
3255 }
3256
3257 /// [`convert`](Self::convert) that also says how the conversion ended —
3258 /// whether the [`document_timeout`](Self::document_timeout) cut it short
3259 /// and how many pages made it (`Completion`).
3260 pub fn convert_outcome(
3261 &mut self,
3262 bytes: &[u8],
3263 password: Option<&str>,
3264 name: &str,
3265 ) -> Result<Converted, PdfError> {
3266 let deadline = self.deadline();
3267 let pages = pdfium_backend::page_count(bytes, password)?;
3268 let range = self.resolve_range(pages)?;
3269 // Serial vs parallel is decided by the pages actually converted: a
3270 // 3-page window over a 500-page PDF should not pay the pool load.
3271 let selected = range.map_or(pages, |(a, b)| b - a + 1);
3272 let (document, done, timed_out) = if self.text_only_for(bytes, password)? {
3273 self.convert_text_only(bytes, password, name, range, selected, deadline)?
3274 } else if self.target_workers >= 2 && selected >= self.parallel_min {
3275 self.convert_parallel(bytes, password, name, range, selected, deadline)?
3276 } else {
3277 self.convert_serial(bytes, password, name, range, selected, deadline)?
3278 };
3279 timing::report();
3280 Ok(Converted {
3281 document,
3282 completion: self.completion(timed_out, done, selected),
3283 })
3284 }
3285
3286 /// Whether this document converts from its text layer alone: always
3287 /// under [`no_ocr`](Self::no_ocr); under
3288 /// [`text_layer_auto`](Self::text_layer_auto), when the layer carries
3289 /// the text ([`text_layer_has_text`]). Decided before any worker — and
3290 /// so any model — is touched.
3291 fn text_only_for(&self, bytes: &[u8], password: Option<&str>) -> Result<bool, PdfError> {
3292 if self.no_ocr {
3293 return Ok(true);
3294 }
3295 if !self.text_layer_auto {
3296 return Ok(false);
3297 }
3298 let has_text = text_layer_has_text(bytes, password)?;
3299 docling_core::debug_log!(
3300 "text_layer_only=auto: {}",
3301 if has_text {
3302 "the text layer carries the document, converting it alone"
3303 } else {
3304 "no usable text layer, running the full pipeline"
3305 }
3306 );
3307 Ok(has_text)
3308 }
3309
3310 /// The text layer alone, page by page — no worker, no model, no page
3311 /// bitmap: [`text_only_page`] on each extracted page, in order. The path
3312 /// of [`no_ocr`](Self::no_ocr) and of an [`auto`](Self::text_layer_auto)
3313 /// document with text; byte-identical to the serial walk, since the
3314 /// worker's `no_ocr` branch is the same function.
3315 fn convert_text_only(
3316 &mut self,
3317 bytes: &[u8],
3318 password: Option<&str>,
3319 name: &str,
3320 range: Option<(usize, usize)>,
3321 selected: usize,
3322 deadline: Option<std::time::Instant>,
3323 ) -> Result<(DoclingDocument, usize, bool), PdfError> {
3324 let mut doc = DoclingDocument::new(name);
3325 let mut confs = std::collections::BTreeMap::new();
3326 let images = self.images;
3327 let progress = self.progress.clone();
3328 let mut done = 0usize;
3329 let timed_out = Self::render_ahead(
3330 bytes,
3331 password,
3332 false,
3333 pdfium_backend::BitmapRender::Eager,
3334 true,
3335 range,
3336 deadline,
3337 self.text_options(),
3338 |n, page| {
3339 let (mut nodes, links, conf) = text_only_page(&page, images.scale);
3340 assemble::stamp_page_no(&mut nodes, n + 1);
3341 if let Some(img) = page_image_under(images, n, &page) {
3342 doc.page_images.insert(n + 1, img);
3343 }
3344 doc.nodes.extend(nodes);
3345 doc.links.extend(links);
3346 confs.insert(n + 1, conf);
3347 done += 1;
3348 if let Some(cb) = &progress {
3349 cb(PageProgress {
3350 page_no: n + 1,
3351 pages_done: done,
3352 pages_selected: selected,
3353 });
3354 }
3355 Ok(())
3356 },
3357 )?;
3358 assemble::merge_continuations(&mut doc.nodes);
3359 self.apply_heading_hierarchy(&mut doc.nodes, Some(bytes));
3360 doc.confidence = Some(docling_core::ConfidenceReport::from_pages(confs));
3361 Ok((doc, done, timed_out))
3362 }
3363
3364 /// Stream pages one at a time through the primary worker — render → process →
3365 /// drop — so the document holds ~one page bitmap (~5 MB) at a time.
3366 fn convert_serial(
3367 &mut self,
3368 bytes: &[u8],
3369 password: Option<&str>,
3370 name: &str,
3371 range: Option<(usize, usize)>,
3372 selected: usize,
3373 deadline: Option<std::time::Instant>,
3374 ) -> Result<(DoclingDocument, usize, bool), PdfError> {
3375 let mut doc = DoclingDocument::new(name);
3376 let mut confs = std::collections::BTreeMap::new();
3377 let mut ocr_cells_by_page = std::collections::BTreeMap::new();
3378 let render_image = !self.no_ocr;
3379 let bitmap = self.bitmap_render();
3380 let extract_text = self.extract_text_layer();
3381 let progress = self.progress.clone();
3382 let mut done = 0usize;
3383 let text = self.text_options();
3384 let worker = self.primary()?;
3385 let timed_out = Self::render_ahead(
3386 bytes,
3387 password,
3388 render_image,
3389 bitmap,
3390 extract_text,
3391 range,
3392 deadline,
3393 text,
3394 |n, mut page| {
3395 let (mut nodes, links, conf, page_image, ocr_cells) =
3396 worker.process(n, &mut page)?;
3397 assemble::stamp_page_no(&mut nodes, n + 1);
3398 if let Some(img) = page_image {
3399 doc.page_images.insert(n + 1, img);
3400 }
3401 doc.nodes.extend(nodes);
3402 doc.links.extend(links);
3403 confs.insert(n + 1, conf);
3404 ocr_cells_by_page.insert(n + 1, ocr_cells);
3405 done += 1;
3406 if let Some(cb) = &progress {
3407 cb(PageProgress {
3408 page_no: n + 1,
3409 pages_done: done,
3410 pages_selected: selected,
3411 });
3412 }
3413 Ok(())
3414 },
3415 )?;
3416 assemble::merge_continuations(&mut doc.nodes);
3417 self.apply_heading_hierarchy(&mut doc.nodes, Some(bytes));
3418 doc.confidence = Some(confidence_report(confs, ocr_cells_by_page));
3419 Ok((doc, done, timed_out))
3420 }
3421
3422 /// The serial path's schedule (#682): pages render on a scoped thread
3423 /// and queue through a [`RENDER_AHEAD`]-page channel while the primary
3424 /// worker processes the previous one on the calling thread, so a page's
3425 /// two rasters (about a third of a digital page's cost) hide behind
3426 /// layout and TableFormer instead of preceding them — the render-ahead
3427 /// the worker pool always had, for documents too small for the pool.
3428 /// One consumer, pages in order: the output is byte-identical to the
3429 /// sequential walk. `consume` runs on the calling thread (a streaming
3430 /// `emit` need not be `Send`). A spent budget (#497) ends the walk at
3431 /// the renderer and the consumer alike; a consumer error closes the
3432 /// channel, which stops the renderer, and is the one reported.
3433 #[allow(clippy::too_many_arguments)]
3434 fn render_ahead<F>(
3435 bytes: &[u8],
3436 password: Option<&str>,
3437 render_image: bool,
3438 bitmap: pdfium_backend::BitmapRender,
3439 extract_text: bool,
3440 range: Option<(usize, usize)>,
3441 deadline: Option<std::time::Instant>,
3442 text: pdfium_backend::TextOptions,
3443 mut consume: F,
3444 ) -> Result<bool, PdfError>
3445 where
3446 F: FnMut(usize, PdfPage) -> Result<(), PdfError>,
3447 {
3448 let (tx, rx) = sync_channel::<(usize, PdfPage)>(RENDER_AHEAD);
3449 std::thread::scope(|s| {
3450 let render = s.spawn(move || {
3451 pdfium_backend::walk_pages(
3452 bytes,
3453 password,
3454 render_image,
3455 extract_text,
3456 range,
3457 bitmap,
3458 text,
3459 |n, _total, page| {
3460 if expired(deadline) {
3461 return Err(timeout_sentinel());
3462 }
3463 tx.send((n, page))
3464 .map_err(|_| PdfError::Pdfium("page channel closed".into()))
3465 },
3466 )
3467 });
3468 let mut timed_out = false;
3469 let mut failed = None;
3470 for (n, page) in rx.iter() {
3471 if expired(deadline) {
3472 timed_out = true;
3473 break;
3474 }
3475 if let Err(e) = consume(n, page) {
3476 failed = Some(e);
3477 break;
3478 }
3479 }
3480 // Closing the channel ends a walk still rendering.
3481 drop(rx);
3482 let walk = render
3483 .join()
3484 .map_err(|_| PdfError::Pdfium("page render thread panicked".into()))?;
3485 if let Some(e) = failed {
3486 return Err(e);
3487 }
3488 match walk {
3489 Ok(()) => Ok(timed_out),
3490 Err(PdfError::Timeout(_)) => Ok(true),
3491 Err(_) if timed_out => Ok(true),
3492 Err(e) => Err(e),
3493 }
3494 })
3495 }
3496
3497 /// Render pages serially on this thread (pdfium) and process them in parallel
3498 /// across the worker pool. A bounded channel applies backpressure so only a
3499 /// handful of page bitmaps are resident at once; results carry their page
3500 /// index and are reassembled in order, so the output is byte-identical to the
3501 /// serial path.
3502 fn convert_parallel(
3503 &mut self,
3504 bytes: &[u8],
3505 password: Option<&str>,
3506 name: &str,
3507 range: Option<(usize, usize)>,
3508 selected: usize,
3509 deadline: Option<std::time::Instant>,
3510 ) -> Result<(DoclingDocument, usize, bool), PdfError> {
3511 let text = self.text_options();
3512 self.ensure_pool_for(selected)?;
3513 let progress = self.progress.clone();
3514 let pages_done = std::sync::atomic::AtomicUsize::new(0);
3515 let timed_out = std::sync::atomic::AtomicBool::new(false);
3516 let n_workers = self.pool.len();
3517 let render_image = !self.no_ocr;
3518 let bitmap = self.bitmap_render();
3519 let extract_text = self.extract_text_layer();
3520 let layout_batch = pdf_layout_batch();
3521 // Bound sized so every worker can accumulate a full layout batch while
3522 // rendering stays ahead (and never below the pre-#73 render-ahead of
3523 // two pages per worker); still a hard cap on resident page bitmaps.
3524 let (work_tx, work_rx) = sync_channel::<(usize, PdfPage)>(n_workers * layout_batch.max(2));
3525 let work_rx: Arc<Mutex<Receiver<(usize, PdfPage)>>> = Arc::new(Mutex::new(work_rx));
3526 let results: Arc<Mutex<Vec<(usize, PageOut)>>> = Arc::new(Mutex::new(Vec::new()));
3527 let first_err: Arc<Mutex<Option<PdfError>>> = Arc::new(Mutex::new(None));
3528
3529 // Move the pool into the scope so each worker gets an exclusive `&mut`.
3530 let mut workers = std::mem::take(&mut self.pool);
3531 std::thread::scope(|s| {
3532 for worker in workers.iter_mut() {
3533 let work_rx = Arc::clone(&work_rx);
3534 let results = Arc::clone(&results);
3535 let first_err = Arc::clone(&first_err);
3536 let progress = progress.clone();
3537 let pages_done = &pages_done;
3538 let timed_out = &timed_out;
3539 s.spawn(move || {
3540 worker.run_pool(&work_rx, layout_batch, deadline, |idx, out| {
3541 match out {
3542 Ok(out) => {
3543 results.lock().unwrap().push((idx, out));
3544 let d = pages_done
3545 .fetch_add(1, std::sync::atomic::Ordering::Relaxed)
3546 + 1;
3547 if let Some(cb) = &progress {
3548 cb(PageProgress {
3549 page_no: idx + 1,
3550 pages_done: d,
3551 pages_selected: selected,
3552 });
3553 }
3554 }
3555 // The budget ran out before this page's turn: the
3556 // page is left out, the document is partial.
3557 Err(PdfError::Timeout(_)) => {
3558 timed_out.store(true, std::sync::atomic::Ordering::Relaxed);
3559 }
3560 Err(e) => {
3561 let mut slot = first_err.lock().unwrap();
3562 if slot.is_none() {
3563 *slot = Some(e);
3564 }
3565 }
3566 }
3567 true
3568 });
3569 });
3570 }
3571 // Render on this thread and feed the workers; backpressure blocks here
3572 // when the channel is full. Dropping `work_tx` afterwards signals the
3573 // workers (recv → Err) to finish. A spent budget ends the walk before
3574 // the next page is rendered.
3575 let render = pdfium_backend::walk_pages(
3576 bytes,
3577 password,
3578 render_image,
3579 extract_text,
3580 range,
3581 bitmap,
3582 text,
3583 |i, _total, page| {
3584 if expired(deadline) {
3585 return Err(timeout_sentinel());
3586 }
3587 work_tx
3588 .send((i, page))
3589 .map_err(|_| PdfError::Pdfium("page-worker channel closed".into()))
3590 },
3591 );
3592 drop(work_tx);
3593 match render {
3594 Ok(()) => {}
3595 Err(PdfError::Timeout(_)) => {
3596 timed_out.store(true, std::sync::atomic::Ordering::Relaxed);
3597 }
3598 Err(e) => {
3599 let mut slot = first_err.lock().unwrap();
3600 if slot.is_none() {
3601 *slot = Some(e);
3602 }
3603 }
3604 }
3605 });
3606 // Threads have joined; restore the pool for the next conversion.
3607 self.pool = workers;
3608
3609 if let Some(e) = first_err.lock().unwrap().take() {
3610 return Err(e);
3611 }
3612 let mut results = Arc::try_unwrap(results)
3613 .unwrap_or_else(|arc| Mutex::new(arc.lock().unwrap().clone()))
3614 .into_inner()
3615 .unwrap();
3616 results.sort_by_key(|(idx, _)| *idx);
3617 let done = results.len();
3618 let mut doc = DoclingDocument::new(name);
3619 let mut confs = std::collections::BTreeMap::new();
3620 let mut ocr_cells_by_page = std::collections::BTreeMap::new();
3621 for (idx, (mut nodes, links, conf, page_image, ocr_cells)) in results {
3622 assemble::stamp_page_no(&mut nodes, idx + 1);
3623 if let Some(img) = page_image {
3624 doc.page_images.insert(idx + 1, img);
3625 }
3626 doc.nodes.extend(nodes);
3627 doc.links.extend(links);
3628 confs.insert(idx + 1, conf);
3629 ocr_cells_by_page.insert(idx + 1, ocr_cells);
3630 }
3631 assemble::merge_continuations(&mut doc.nodes);
3632 self.apply_heading_hierarchy(&mut doc.nodes, Some(bytes));
3633 doc.confidence = Some(confidence_report(confs, ocr_cells_by_page));
3634 let timed_out = timed_out.load(std::sync::atomic::Ordering::Relaxed);
3635 Ok((doc, done, timed_out))
3636 }
3637
3638 /// Convert a PDF in **streaming** mode: `emit` is called with each finalized,
3639 /// in-document-order batch of nodes (and that span's recovered links) as pages
3640 /// complete, so a caller can serialize Markdown page by page instead of waiting
3641 /// for the whole document. The batches are exactly the buffered [`convert`]'s
3642 /// nodes, split at safe block boundaries by [`assemble::StreamAssembler`] — the
3643 /// parallel path reorders pages back into document order before emitting, so
3644 /// the output is identical regardless of worker scheduling.
3645 ///
3646 /// `emit` runs on the calling thread (never a worker), so it needn't be `Send`
3647 /// and its backpressure throttles the whole pipeline. Returning `Err` from
3648 /// `emit` aborts the conversion with that error.
3649 pub fn convert_streaming<F>(
3650 &mut self,
3651 bytes: &[u8],
3652 password: Option<&str>,
3653 name: &str,
3654 emit: F,
3655 ) -> Result<(), PdfError>
3656 where
3657 F: FnMut(Vec<Node>, Vec<(String, String)>) -> Result<(), PdfError>,
3658 {
3659 self.convert_streaming_outcome(bytes, password, name, emit)
3660 .map(|_| ())
3661 }
3662
3663 /// [`convert_streaming`](Self::convert_streaming) that also says how the
3664 /// conversion ended ([`Completion`]): a spent
3665 /// [`document_timeout`](Self::document_timeout) stops the walk, the pages
3666 /// finished so far are emitted (the tail included) and the outcome is
3667 /// `TimedOut` — never an error.
3668 pub fn convert_streaming_outcome<F>(
3669 &mut self,
3670 bytes: &[u8],
3671 password: Option<&str>,
3672 name: &str,
3673 emit: F,
3674 ) -> Result<Completion, PdfError>
3675 where
3676 F: FnMut(Vec<Node>, Vec<(String, String)>) -> Result<(), PdfError>,
3677 {
3678 let _ = name; // page nodes carry no name; the caller owns the document name.
3679 let deadline = self.deadline();
3680 let pages = pdfium_backend::page_count(bytes, password)?;
3681 let range = self.resolve_range(pages)?;
3682 let selected = range.map_or(pages, |(a, b)| b - a + 1);
3683 let r = if self.text_only_for(bytes, password)? {
3684 self.convert_streaming_text_only(bytes, password, range, deadline, selected, emit)
3685 } else if self.target_workers >= 2 && selected >= self.parallel_min {
3686 self.convert_streaming_parallel(bytes, password, range, deadline, selected, emit)
3687 } else {
3688 self.convert_streaming_serial(bytes, password, range, deadline, selected, emit)
3689 };
3690 timing::report();
3691 let (done, timed_out) = r?;
3692 Ok(self.completion(timed_out, done, selected))
3693 }
3694
3695 /// Streaming from the text layer alone ([`convert_text_only`](Self::convert_text_only)'s
3696 /// emitting twin): extract → [`text_only_page`] → emit, one page at a
3697 /// time, holding back only the tail that might still merge into the next
3698 /// page. No worker, no model, no bitmap.
3699 fn convert_streaming_text_only<F>(
3700 &mut self,
3701 bytes: &[u8],
3702 password: Option<&str>,
3703 range: Option<(usize, usize)>,
3704 deadline: Option<std::time::Instant>,
3705 selected: usize,
3706 mut emit: F,
3707 ) -> Result<(usize, bool), PdfError>
3708 where
3709 F: FnMut(Vec<Node>, Vec<(String, String)>) -> Result<(), PdfError>,
3710 {
3711 let mut asm = assemble::StreamAssembler::new();
3712 let images = self.images;
3713 let progress = self.progress.clone();
3714 let mut done = 0usize;
3715 let timed_out = Self::render_ahead(
3716 bytes,
3717 password,
3718 false,
3719 pdfium_backend::BitmapRender::Eager,
3720 true,
3721 range,
3722 deadline,
3723 self.text_options(),
3724 |n, page| {
3725 let (nodes, links, _conf) = text_only_page(&page, images.scale);
3726 done += 1;
3727 if let Some(cb) = &progress {
3728 cb(PageProgress {
3729 page_no: n + 1,
3730 pages_done: done,
3731 pages_selected: selected,
3732 });
3733 }
3734 emit(asm.push(nodes), links)
3735 },
3736 )?;
3737 emit(asm.finish(), Vec::new())?;
3738 Ok((done, timed_out))
3739 }
3740
3741 /// Serial streaming: render → process → emit, one page at a time, holding back
3742 /// only the tail that might still merge into the next page.
3743 fn convert_streaming_serial<F>(
3744 &mut self,
3745 bytes: &[u8],
3746 password: Option<&str>,
3747 range: Option<(usize, usize)>,
3748 deadline: Option<std::time::Instant>,
3749 selected: usize,
3750 mut emit: F,
3751 ) -> Result<(usize, bool), PdfError>
3752 where
3753 F: FnMut(Vec<Node>, Vec<(String, String)>) -> Result<(), PdfError>,
3754 {
3755 let mut asm = assemble::StreamAssembler::new();
3756 let render_image = !self.no_ocr;
3757 let bitmap = self.bitmap_render();
3758 let extract_text = self.extract_text_layer();
3759 let progress = self.progress.clone();
3760 let text = self.text_options();
3761 let worker = self.primary()?;
3762 let mut done = 0usize;
3763 let timed_out = Self::render_ahead(
3764 bytes,
3765 password,
3766 render_image,
3767 bitmap,
3768 extract_text,
3769 range,
3770 deadline,
3771 text,
3772 |n, mut page| {
3773 // Confidence is dropped on the streaming path: the report is
3774 // only complete once every page has run, which defeats
3775 // page-by-page emission — buffered `convert` carries it.
3776 let (nodes, links, _conf, _page_image, _cells) = worker.process(n, &mut page)?;
3777 done += 1;
3778 if let Some(cb) = &progress {
3779 cb(PageProgress {
3780 page_no: n + 1,
3781 pages_done: done,
3782 pages_selected: selected,
3783 });
3784 }
3785 emit(asm.push(nodes), links)
3786 },
3787 )?;
3788 emit(asm.finish(), Vec::new())?;
3789 Ok((done, timed_out))
3790 }
3791
3792 /// Parallel streaming: pages render serially on a dedicated thread (pdfium is
3793 /// not thread-safe) and process across the worker pool; results carry their
3794 /// page index and are reordered on the calling thread into a
3795 /// [`assemble::StreamAssembler`], which emits each page in document order as
3796 /// soon as its predecessors have arrived. Bounded channels keep only a handful
3797 /// of pages resident and let `emit`'s backpressure reach the renderer.
3798 fn convert_streaming_parallel<F>(
3799 &mut self,
3800 bytes: &[u8],
3801 password: Option<&str>,
3802 range: Option<(usize, usize)>,
3803 deadline: Option<std::time::Instant>,
3804 selected: usize,
3805 mut emit: F,
3806 ) -> Result<(usize, bool), PdfError>
3807 where
3808 F: FnMut(Vec<Node>, Vec<(String, String)>) -> Result<(), PdfError>,
3809 {
3810 let text = self.text_options();
3811 self.ensure_pool_for(selected)?;
3812 let n_workers = self.pool.len();
3813 let render_image = !self.no_ocr;
3814 let bitmap = self.bitmap_render();
3815 let extract_text = self.extract_text_layer();
3816 let progress = self.progress.clone();
3817 let layout_batch = pdf_layout_batch();
3818 // Bound sized so every worker can accumulate a full layout batch while
3819 // rendering stays ahead (and never below the pre-#73 render-ahead of
3820 // two pages per worker); still a hard cap on resident page bitmaps.
3821 let (work_tx, work_rx) = sync_channel::<(usize, PdfPage)>(n_workers * layout_batch.max(2));
3822 let work_rx: Arc<Mutex<Receiver<(usize, PdfPage)>>> = Arc::new(Mutex::new(work_rx));
3823 // Workers and the renderer report here; the calling thread drains it in
3824 // page order. Bounded so workers block (bounding resident bitmaps) when the
3825 // consumer falls behind.
3826 let (res_tx, res_rx) = sync_channel::<Result<(usize, PageOut), PdfError>>(n_workers * 2);
3827
3828 let mut workers = std::mem::take(&mut self.pool);
3829 let mut asm = assemble::StreamAssembler::new();
3830 let mut first_err: Option<PdfError> = None;
3831 let mut timed_out = false;
3832 let mut done = 0usize;
3833
3834 std::thread::scope(|s| {
3835 // Workers: pull a batch of pages (whatever is already rendered, up
3836 // to the layout batch size), process it, report (index-tagged)
3837 // results.
3838 for worker in workers.iter_mut() {
3839 let work_rx = Arc::clone(&work_rx);
3840 let res_tx = res_tx.clone();
3841 s.spawn(move || {
3842 worker.run_pool(&work_rx, layout_batch, deadline, |idx, out| {
3843 // `false` once the consumer is gone.
3844 res_tx.send(out.map(|o| (idx, o))).is_ok()
3845 });
3846 });
3847 }
3848 // Renderer: feed pages to the pool on its own thread (pdfium stays on a
3849 // single thread); report a render error through the same channel. A
3850 // spent budget ends the walk before the next page is rendered.
3851 {
3852 let res_tx = res_tx.clone();
3853 s.spawn(move || {
3854 let render = pdfium_backend::walk_pages(
3855 bytes,
3856 password,
3857 render_image,
3858 extract_text,
3859 range,
3860 bitmap,
3861 text,
3862 |i, _total, page| {
3863 if expired(deadline) {
3864 return Err(timeout_sentinel());
3865 }
3866 work_tx
3867 .send((i, page))
3868 .map_err(|_| PdfError::Pdfium("page-worker channel closed".into()))
3869 },
3870 );
3871 drop(work_tx); // signal workers to finish
3872 if let Err(e) = render {
3873 let _ = res_tx.send(Err(e));
3874 }
3875 });
3876 }
3877 // Drop our own sender so the channel closes once the threads finish.
3878 drop(res_tx);
3879
3880 // Collector (this thread): reorder into document order and emit.
3881 // With a page window, indices start at the window's first page.
3882 // A timed-out page leaves a hole in the sequence: everything
3883 // before it emits in order, the pages after it stay buffered —
3884 // they are the pages the budget did not cover, and the document
3885 // ends at the cut like the buffered path's.
3886 let mut buffer: BTreeMap<usize, PageOut> = BTreeMap::new();
3887 let mut next = range.map_or(0, |(first, _)| first);
3888 for msg in res_rx.iter() {
3889 match msg {
3890 Err(PdfError::Timeout(_)) => timed_out = true,
3891 Err(e) => {
3892 if first_err.is_none() {
3893 first_err = Some(e);
3894 }
3895 }
3896 Ok((idx, out)) => {
3897 buffer.insert(idx, out);
3898 if first_err.is_some() {
3899 continue; // keep draining so the threads can exit
3900 }
3901 // A stream carries Markdown, not the page map: page
3902 // images (#520) have nowhere to go and are dropped.
3903 while let Some((nodes, links, _conf, _page_image, _cells)) =
3904 buffer.remove(&next)
3905 {
3906 if let Err(e) = emit(asm.push(nodes), links) {
3907 first_err = Some(e);
3908 break;
3909 }
3910 next += 1;
3911 done += 1;
3912 // Reported in document order, as emitted: a page
3913 // that finished early counts once it is emitted.
3914 if let Some(cb) = &progress {
3915 cb(PageProgress {
3916 page_no: next,
3917 pages_done: done,
3918 pages_selected: selected,
3919 });
3920 }
3921 }
3922 }
3923 }
3924 }
3925 // Pages that finished after a hole (a page the budget skipped
3926 // before them) cannot follow it in document order; count them
3927 // out of the document like the skipped one.
3928 if timed_out && !buffer.is_empty() {
3929 buffer.clear();
3930 }
3931 });
3932 // Threads have joined; restore the pool for the next conversion.
3933 self.pool = workers;
3934
3935 if let Some(e) = first_err {
3936 return Err(e);
3937 }
3938 emit(asm.finish(), Vec::new())?;
3939 Ok((done, timed_out))
3940 }
3941
3942 /// Lazily grow the pool to `target_workers`, loading the new workers
3943 /// concurrently (model load is mostly I/O + mmap, so N loads overlap to roughly
3944 /// one load's wall-time). Cached for reuse across documents.
3945 fn ensure_pool(&mut self) -> Result<(), PdfError> {
3946 self.ensure_pool_for(self.target_workers)
3947 }
3948
3949 /// [`ensure_pool`](Self::ensure_pool) for a document of `pages` pages:
3950 /// no more workers than pages (#682 — a two-page document on a 16-core
3951 /// host loads two workers, not eight), never fewer than two, and never
3952 /// more than `target_workers`; a later, longer document grows the pool.
3953 fn ensure_pool_for(&mut self, pages: usize) -> Result<(), PdfError> {
3954 let want = pages.max(2).min(self.target_workers);
3955 let need = want.saturating_sub(self.pool.len());
3956 if need == 0 {
3957 return Ok(());
3958 }
3959 let intra = pdf_intra();
3960 let no_ocr = self.no_ocr;
3961 let skip_ocr = self.skip_ocr;
3962 let force = self.force_full_page_ocr || self.ocr_mode.forces_full_page();
3963 let ntp = self.no_text_panels;
3964 let ocr_lang = self.ocr_lang;
3965 let ocr_engine = self.ocr_engine;
3966 let tesseract_lang = self.tesseract_lang.clone();
3967 let ocr_scale = self.ocr_scale;
3968 let textline_orientation = self.textline_orientation;
3969 let images = self.images;
3970 let enrich = self.enrich;
3971 let tables = self.tables_slot();
3972 let enrich_slots = self.enrich_slots();
3973 let loaded: Vec<Result<Worker, PdfError>> = std::thread::scope(|s| {
3974 let handles: Vec<_> = (0..need)
3975 .map(|_| {
3976 let tables = tables.clone();
3977 let enrich_slots = enrich_slots.clone();
3978 let tesseract_lang = tesseract_lang.clone();
3979 s.spawn(move || {
3980 Worker::load(
3981 intra,
3982 tables,
3983 enrich_slots,
3984 enrich,
3985 no_ocr,
3986 skip_ocr,
3987 force,
3988 ntp,
3989 ocr_lang,
3990 ocr_engine,
3991 tesseract_lang,
3992 ocr_scale,
3993 textline_orientation,
3994 images,
3995 )
3996 })
3997 })
3998 .collect();
3999 handles.into_iter().map(|h| h.join().unwrap()).collect()
4000 });
4001 for w in loaded {
4002 self.pool.push(w?);
4003 }
4004 Ok(())
4005 }
4006
4007 /// Convert a standalone image (PNG/JPEG/TIFF/WebP/…) as a single page —
4008 /// docling routes images through the same layout+OCR pipeline as a PDF page.
4009 pub fn convert_image(&mut self, bytes: &[u8], name: &str) -> Result<DoclingDocument, PdfError> {
4010 let image = decode_image_limited(bytes)?;
4011 let (w, h) = image.dimensions();
4012 // The image is its own page rendered at 1 px per "point" (scale 1.0); a
4013 // standalone image has no text layer, so OCR supplies the cells.
4014 let page = PdfPage {
4015 width: w as f32,
4016 height: h as f32,
4017 scale: 1.0,
4018 cells: Vec::new(),
4019 code_cells: Vec::new(),
4020 checkboxes: Vec::new(),
4021 word_cells: Vec::new(),
4022 // A standalone image *is* its own scale-1.0 page image, so the
4023 // layout model sees it through the docling-exact PIL kernel.
4024 image_layout: Some(image.clone()),
4025 image,
4026 links: Vec::new(),
4027 rotation: 0,
4028 keep_typography: false,
4029 pending_image: None,
4030 };
4031 self.process_pages(vec![page], name)
4032 }
4033
4034 /// Run layout (+ OCR for cell-less pages) and assemble each already-rendered
4035 /// page (image / METS inputs, which are small and already materialised).
4036 /// Public so [`mets::convert_mets_gbs_with_pipeline`] can drive a
4037 /// caller-configured pipeline (#244).
4038 pub fn process_pages(
4039 &mut self,
4040 mut pages: Vec<PdfPage>,
4041 name: &str,
4042 ) -> Result<DoclingDocument, PdfError> {
4043 let mut doc = DoclingDocument::new(name);
4044 let mut confs = std::collections::BTreeMap::new();
4045 let mut ocr_cells_by_page = std::collections::BTreeMap::new();
4046 let progress = self.progress.clone();
4047 let total = pages.len();
4048 let worker = self.primary()?;
4049 for (n, page) in pages.iter_mut().enumerate() {
4050 let (mut nodes, links, conf, page_image, ocr_cells) = worker.process(n, page)?;
4051 assemble::stamp_page_no(&mut nodes, n + 1);
4052 if let Some(img) = page_image {
4053 doc.page_images.insert(n + 1, img);
4054 }
4055 doc.nodes.extend(nodes);
4056 doc.links.extend(links);
4057 confs.insert(n + 1, conf);
4058 ocr_cells_by_page.insert(n + 1, ocr_cells);
4059 if let Some(cb) = &progress {
4060 cb(PageProgress {
4061 page_no: n + 1,
4062 pages_done: n + 1,
4063 pages_selected: total,
4064 });
4065 }
4066 }
4067 assemble::merge_continuations(&mut doc.nodes);
4068 // No PDF behind these pages (images, METS): the heading-hierarchy
4069 // stage degrades to the numbering signal — exactly docling without
4070 // an outline or parsed pages.
4071 self.apply_heading_hierarchy(&mut doc.nodes, None);
4072 doc.confidence = Some(confidence_report(confs, ocr_cells_by_page));
4073 Ok(doc)
4074 }
4075}
4076
4077/// Number of pages in a PDF, without converting anything — what the CLI batch
4078/// mode prints in its per-document start line.
4079#[cfg(feature = "ml")]
4080pub fn page_count(bytes: &[u8], password: Option<&str>) -> Result<usize, PdfError> {
4081 pdfium_backend::page_count(bytes, password)
4082}
4083
4084#[cfg(feature = "ml")]
4085/// Convenience one-shot conversion (loads the pipeline per call). Errors are
4086/// detailed and surfaced (never silently skipped).
4087pub fn convert(
4088 bytes: &[u8],
4089 password: Option<&str>,
4090 name: &str,
4091) -> Result<DoclingDocument, PdfError> {
4092 convert_with_options(
4093 bytes,
4094 password,
4095 name,
4096 false,
4097 false,
4098 false,
4099 false,
4100 EnrichmentOptions::default(),
4101 None,
4102 None,
4103 )
4104}
4105
4106#[cfg(feature = "ml")]
4107/// Like [`convert`], but optionally skips loading/running TableFormer (see
4108/// [`Pipeline::no_table_former`]) and/or layout+OCR+TableFormer entirely (see
4109/// [`Pipeline::no_ocr`]), and/or enables the enrichment passes (see
4110/// [`Pipeline::enrichments`]).
4111// One positional per pipeline switch mirrors the Pipeline builder; growing
4112// past clippy's arity cap is the price of keeping this one-shot signature
4113// stable-ish instead of churning callers into an options struct mid-series.
4114#[allow(clippy::too_many_arguments)]
4115pub fn convert_with_options(
4116 bytes: &[u8],
4117 password: Option<&str>,
4118 name: &str,
4119 no_table_former: bool,
4120 no_ocr: bool,
4121 force_full_page_ocr: bool,
4122 no_text_panels: bool,
4123 enrich: EnrichmentOptions,
4124 pages: Option<(usize, usize)>,
4125 ocr_lang: Option<OcrLang>,
4126) -> Result<DoclingDocument, PdfError> {
4127 Pipeline::new()?
4128 .no_table_former(no_table_former)
4129 .no_ocr(no_ocr)
4130 .force_full_page_ocr(force_full_page_ocr)
4131 .no_text_panels(no_text_panels)
4132 .enrichments(enrich)
4133 .pages(pages)
4134 .ocr_lang(ocr_lang)
4135 .convert(bytes, password, name)
4136}
4137
4138#[cfg(feature = "ml")]
4139/// Convenience one-shot image conversion (loads the pipeline per call).
4140pub fn convert_image(bytes: &[u8], name: &str) -> Result<DoclingDocument, PdfError> {
4141 convert_image_with_options(
4142 bytes,
4143 name,
4144 false,
4145 false,
4146 false,
4147 EnrichmentOptions::default(),
4148 None,
4149 )
4150}
4151
4152#[cfg(feature = "ml")]
4153/// Like [`convert_image`], but optionally skips loading/running TableFormer (see
4154/// [`Pipeline::no_table_former`]) and/or layout+OCR+TableFormer entirely (see
4155/// [`Pipeline::no_ocr`]), and/or enables the enrichment passes.
4156pub fn convert_image_with_options(
4157 bytes: &[u8],
4158 name: &str,
4159 no_table_former: bool,
4160 no_ocr: bool,
4161 no_text_panels: bool,
4162 enrich: EnrichmentOptions,
4163 ocr_lang: Option<OcrLang>,
4164) -> Result<DoclingDocument, PdfError> {
4165 Pipeline::new()?
4166 .no_table_former(no_table_former)
4167 .no_ocr(no_ocr)
4168 .no_text_panels(no_text_panels)
4169 .enrichments(enrich)
4170 .ocr_lang(ocr_lang)
4171 .convert_image(bytes, name)
4172}
4173
4174#[cfg(feature = "ml")]
4175/// Convert pre-segmented pages (image + already-known text cells, e.g. METS/hOCR
4176/// scans) through the shared layout + assembly pipeline.
4177pub fn convert_pages(pages: Vec<PdfPage>, name: &str) -> Result<DoclingDocument, PdfError> {
4178 convert_pages_with_options(
4179 pages,
4180 name,
4181 false,
4182 false,
4183 false,
4184 EnrichmentOptions::default(),
4185 )
4186}
4187
4188#[cfg(feature = "ml")]
4189/// Like [`convert_pages`], but optionally skips loading/running TableFormer (see
4190/// [`Pipeline::no_table_former`]) and/or layout+OCR+TableFormer entirely (see
4191/// [`Pipeline::no_ocr`]), and/or enables the enrichment passes.
4192pub fn convert_pages_with_options(
4193 pages: Vec<PdfPage>,
4194 name: &str,
4195 no_table_former: bool,
4196 no_ocr: bool,
4197 no_text_panels: bool,
4198 enrich: EnrichmentOptions,
4199) -> Result<DoclingDocument, PdfError> {
4200 Pipeline::new()?
4201 .no_table_former(no_table_former)
4202 .no_text_panels(no_text_panels)
4203 .no_ocr(no_ocr)
4204 .enrichments(enrich)
4205 .process_pages(pages, name)
4206}
4207
4208#[cfg(feature = "ml")]
4209#[cfg(all(test, feature = "ml"))]
4210mod image_limit_tests {
4211 use super::decode_image_with_max_side;
4212
4213 /// A small valid PNG encoded via the `image` crate (robust vs. a hand-rolled
4214 /// byte literal).
4215 fn png_bytes(w: u32, h: u32) -> Vec<u8> {
4216 use std::io::Cursor;
4217 let img = image::RgbImage::new(w, h);
4218 let mut out = Vec::new();
4219 img.write_to(&mut Cursor::new(&mut out), image::ImageFormat::Png)
4220 .unwrap();
4221 out
4222 }
4223
4224 #[test]
4225 fn normal_image_decodes_under_the_cap() {
4226 let img = decode_image_with_max_side(&png_bytes(8, 8), 30_000).expect("8x8 decodes");
4227 assert_eq!(img.dimensions(), (8, 8));
4228 }
4229
4230 /// docling#4247 (2.128): the EXIF orientation is applied when a frame is
4231 /// loaded — a 4×2 JPEG tagged "rotate 90° CW" (orientation 6) decodes as
4232 /// 2×4, with the pixels turned.
4233 #[test]
4234 fn exif_orientation_is_applied() {
4235 use std::io::Cursor;
4236 let mut img = image::RgbImage::new(4, 2);
4237 img.put_pixel(0, 0, image::Rgb([255, 0, 0]));
4238 let mut jpeg = Vec::new();
4239 img.write_to(&mut Cursor::new(&mut jpeg), image::ImageFormat::Jpeg)
4240 .unwrap();
4241 // Splice an APP1 Exif segment (little-endian TIFF, one IFD entry:
4242 // Orientation = 6) right after the SOI marker.
4243 let mut tiff = b"II*\0".to_vec();
4244 tiff.extend(8u32.to_le_bytes());
4245 tiff.extend(1u16.to_le_bytes());
4246 tiff.extend(0x0112u16.to_le_bytes());
4247 tiff.extend(3u16.to_le_bytes());
4248 tiff.extend(1u32.to_le_bytes());
4249 tiff.extend(6u16.to_le_bytes());
4250 tiff.extend([0, 0]);
4251 tiff.extend(0u32.to_le_bytes());
4252 let mut app1 = b"Exif\0\0".to_vec();
4253 app1.extend(tiff);
4254 let mut out = jpeg[..2].to_vec();
4255 out.extend([0xff, 0xe1]);
4256 out.extend((app1.len() as u16 + 2).to_be_bytes());
4257 out.extend(app1);
4258 out.extend(&jpeg[2..]);
4259 let plain = decode_image_with_max_side(&jpeg, 30_000).unwrap();
4260 assert_eq!(plain.dimensions(), (4, 2));
4261 let turned = decode_image_with_max_side(&out, 30_000).unwrap();
4262 assert_eq!(turned.dimensions(), (2, 4), "quarter turn swaps the sides");
4263 // Rotating 90° CW moves the top-left pixel to the top-right corner.
4264 assert!(turned.get_pixel(1, 0).0[0] > 128);
4265 assert!(turned.get_pixel(0, 0).0[0] <= 128);
4266 }
4267
4268 #[test]
4269 fn dimensions_over_the_cap_are_rejected_not_aborted() {
4270 // A per-side cap below the image's declared size must yield a
4271 // recoverable Err, never an allocation-abort — the mechanism that stops
4272 // a crafted image declaring 60000×60000 from OOM-killing the process.
4273 let r = decode_image_with_max_side(&png_bytes(8, 8), 4);
4274 assert!(
4275 r.is_err(),
4276 "decode must fail under the pixel cap, not abort"
4277 );
4278 }
4279}
4280
4281#[cfg(all(test, feature = "ml"))]
4282mod ocr_scale_tests {
4283 use super::*;
4284
4285 /// #570: an explicit scale wins everywhere; a rendered page keeps its
4286 /// render; a scale-1.0 image page reads at docling's 3 px/pt, shrunk to
4287 /// RapidOCR's 2000 px longer side (a FUNSD scan at 2.0, a 3000 px photo
4288 /// downsampled to 0.67), and an image that lands at 1.0 is not resampled.
4289 #[test]
4290 fn image_inputs_follow_rapidocrs_resolution() {
4291 assert_eq!(page_ocr_scale(Some(1.5), 754.0, 1000.0, 1.0), Some(1.5));
4292 assert_eq!(page_ocr_scale(None, 612.0, 792.0, 2.0), None);
4293 assert_eq!(page_ocr_scale(None, 754.0, 1000.0, 1.0), Some(2.0));
4294 assert_eq!(page_ocr_scale(None, 400.0, 600.0, 1.0), Some(3.0));
4295 let s = page_ocr_scale(None, 3000.0, 2000.0, 1.0).unwrap();
4296 assert!((s - 2.0 / 3.0).abs() < 1e-6, "{s}");
4297 assert_eq!(page_ocr_scale(None, 2000.0, 1500.0, 1.0), None);
4298 }
4299}
4300
4301#[cfg(test)]
4302mod median_tests {
4303 #[test]
4304 fn median_of_empty_is_zero_not_a_panic() {
4305 // A crafted table can leave a row/column with zero matched cells; the
4306 // even-count branch would index values[0 - 1] and panic (→ remote crash
4307 // via docling-serve) without the empty guard.
4308 assert_eq!(super::tf_match::median_for_test(&mut []), 0.0);
4309 assert_eq!(super::tf_match::median_for_test(&mut [4.0, 2.0]), 3.0);
4310 assert_eq!(super::tf_match::median_for_test(&mut [5.0, 1.0, 3.0]), 3.0);
4311 }
4312}
4313
4314#[cfg(test)]
4315mod unreadable_tests {
4316 /// The text-layer path (the wasm build's only one) tells a file nothing
4317 /// can open apart from a PDF that merely has no text layer: garbage is
4318 /// an error about the file, a real PDF converts.
4319 #[test]
4320 fn text_layer_path_reports_an_unreadable_file() {
4321 for bytes in [b"garbage".as_slice(), b"", b"%PDF-1.7\n%%EOF\n"] {
4322 let err = super::convert_text_layer(bytes, "x.pdf").unwrap_err();
4323 assert!(err.to_string().contains("not a readable PDF"), "{err}");
4324 }
4325 let real = std::fs::read(
4326 std::path::Path::new(env!("CARGO_MANIFEST_DIR"))
4327 .join("../../tests/data/pdf/sources/multi_page.pdf"),
4328 )
4329 .unwrap();
4330 assert!(!super::convert_text_layer(&real, "multi_page.pdf")
4331 .unwrap()
4332 .nodes
4333 .is_empty());
4334 }
4335}
4336
4337#[cfg(test)]
4338mod send_check {
4339 /// The Node bindings (`docling-node`) run a shared [`super::Pipeline`] on
4340 /// libuv worker threads (`Arc<Mutex<Pipeline>>`), which is only sound while
4341 /// `Pipeline: Send` holds — this fails to compile if a non-`Send` field
4342 /// (e.g. an `Rc` or a raw pdfium handle) ever lands in the pipeline.
4343 fn assert_send<T: Send>() {}
4344
4345 #[test]
4346 fn pipeline_is_send() {
4347 assert_send::<super::Pipeline>();
4348 }
4349}
4350
4351#[cfg(all(test, feature = "ml"))]
4352mod ocr_input_tests {
4353 /// #254: without an `ocr_scale` (or with one equal to the render scale)
4354 /// the OCR reads the page render untouched and the cache stays cold; a
4355 /// different scale builds one resampled view, reuses it across calls, and
4356 /// reports the requested px/pt so cell geometry divides back to points.
4357 #[test]
4358 fn ocr_input_resamples_only_on_a_real_scale_change() {
4359 let img = image::RgbImage::new(200, 100);
4360 let mut cache = None;
4361 let (v, s) = super::ocr_input(&mut cache, &img, 2.0, None);
4362 assert!(std::ptr::eq(v, &img) && s == 2.0 && cache.is_none());
4363 let (v, s) = super::ocr_input(&mut cache, &img, 2.0, Some(2.0));
4364 assert!(std::ptr::eq(v, &img) && s == 2.0 && cache.is_none());
4365
4366 let (v, s) = super::ocr_input(&mut cache, &img, 2.0, Some(3.0));
4367 assert_eq!((v.width(), v.height(), s), (300, 150, 3.0));
4368 let first = cache.as_ref().map(|c| c as *const image::RgbImage);
4369 let (v, _) = super::ocr_input(&mut cache, &img, 2.0, Some(3.0));
4370 assert_eq!(
4371 Some(v as *const image::RgbImage),
4372 first,
4373 "cached, not rebuilt"
4374 );
4375
4376 let mut down = None;
4377 let (v, s) = super::ocr_input(&mut down, &img, 2.0, Some(1.0));
4378 assert_eq!((v.width(), v.height(), s), (100, 50, 1.0));
4379 }
4380}
4381
4382#[cfg(all(test, feature = "ml"))]
4383mod text_layer_auto_tests {
4384 use std::path::Path;
4385
4386 fn fixture(rel: &str) -> Vec<u8> {
4387 std::fs::read(
4388 Path::new(env!("CARGO_MANIFEST_DIR"))
4389 .join("../..")
4390 .join(rel),
4391 )
4392 .unwrap()
4393 }
4394
4395 /// The verdict behind `text_layer_only=auto`: a digital PDF's layer
4396 /// carries it, a scan's does not, and garbage is the file's error.
4397 #[test]
4398 fn text_layer_verdict_tells_digital_from_scanned() {
4399 let digital = fixture("tests/data/pdf/sources/multi_page.pdf");
4400 assert!(super::text_layer_has_text(&digital, None).unwrap());
4401 let scanned = fixture("tests/data/scanned/sources/ocr_test_raster.pdf");
4402 assert!(!super::text_layer_has_text(&scanned, None).unwrap());
4403 let err = super::text_layer_has_text(b"garbage", None).unwrap_err();
4404 assert!(err.to_string().contains("not a readable PDF"), "{err}");
4405 }
4406
4407 /// On a digital PDF `auto` is the `no_ocr` path byte for byte — the
4408 /// buffered and the streaming conversion alike — and touches no model:
4409 /// the pipeline below has none to load (`Pipeline::new` loads lazily),
4410 /// so a conversion that reached a worker would fail.
4411 #[test]
4412 fn auto_on_a_digital_pdf_is_the_text_layer_path() {
4413 let bytes = fixture("tests/data/pdf/sources/multi_page.pdf");
4414 let mut only = super::Pipeline::new().unwrap().no_ocr(true);
4415 let mut auto = super::Pipeline::new().unwrap().text_layer_auto(true);
4416 let a = only.convert(&bytes, None, "multi_page.pdf").unwrap();
4417 let b = auto.convert(&bytes, None, "multi_page.pdf").unwrap();
4418 assert!(!a.nodes.is_empty());
4419 assert_eq!(a.export_to_markdown(), b.export_to_markdown());
4420 assert_eq!(a.confidence, b.confidence);
4421 let streamed = |p: &mut super::Pipeline| {
4422 let mut nodes = Vec::new();
4423 p.convert_streaming(&bytes, None, "multi_page.pdf", |page, _| {
4424 nodes.extend(page);
4425 Ok(())
4426 })
4427 .unwrap();
4428 nodes
4429 };
4430 assert_eq!(streamed(&mut only), streamed(&mut auto));
4431 }
4432}