foliodoc

Fast document conversion for Python (PDF / scans / images / DOCX → Markdown, JSON, text) on Windows, Linux and macOS. It solves the same problem as docling with no layout neural network and no GPU: exact text from the PDF text layer, OCR only where a page has none, and geometric reconstruction of reading order, headings, lists and tables. See Benchmarks for where that wins and where it doesn't.

from foliodoc import convert

doc = convert("report.pdf")        # or .png/.jpg/.tiff scan, or .docx
print(doc.to_markdown())
doc.tables                         # [[["Item", "Qty"], ["Apple", "3"]], ...]
doc.to_json()                      # blocks with type, text, bbox, page, heading level, cells

foliodoc report.pdf scan.png -o out/ --to md --timings

Install

pip install foliodoc              # digital PDFs and Word files
pip install "foliodoc[ocr]"       # + OCR for scans/images: Apple Vision on macOS, RapidOCR on every OS

Python 3.10+ on Windows, Linux and macOS. Individual engines: foliodoc[apple], foliodoc[rapidocr], foliodoc[tesseract] (the last also needs the tesseract binary). Documentation: https://meet2147.github.io/foliodoc/

How it gets its accuracy

Benchmarks

Two benchmarks, both reproducible from bench/. Every number below comes from the final code. Both systems ran one after the other on an Apple M2, warm (model loading excluded), with default settings unless a row says otherwise. Docling is version 2.63.

Summary

foliodoc is not more accurate than docling across the board. It is much faster, it reads text as well or better on digital PDFs, forms and receipts, and it loses clearly on layout and table structure in real documents.

Real-world: 7,822 public documents

Default configuration (on macOS both systems use Apple Vision for OCR; docling through ocrmac):

Cross-platform configuration. Both systems use RapidOCR (PaddleOCR models on ONNX Runtime, CPU), the OCR they use on Windows and Linux. This ran on a subset: all FUNSD forms, the SROIE test split, and the English and mixed-language OmniDocBench pages that both finished (493). DocLayNet needs no OCR, so its results above apply on every OS.

Notes on the real-world numbers:

• 102 DocLayNet pages are excluded from the text metric for both systems because DocLayNet's own ground-truth text is garbled on them (PDFs with broken font encodings, e.g. Ó Ç ä ä Ê).
• Text inside figures is excluded on both sides. Docling's layout model was trained on DocLayNet, so DocLayNet is home ground for it.
• Chinese pages: both systems ran with default (Latin-script) OCR settings and score poorly on them. foliodoc accepts languages=["zh-Hans", "en-US"] for Apple Vision; that setting was not benchmarked.
• A failed document counts as empty output. Every document that crashed or hung was recorded as a failure; none were silently skipped.
• The cross-platform run was cut short to save time: OmniDocBench covers 493 of the planned 871 pages. Both systems are scored on exactly the same pages.
• Bugs found in foliodoc while running this benchmark were fixed before the final numbers: effective font size, word spacing in PDFs without space glyphs, letter-spaced and rotated text, page-border frames, content outside the visible page, a pdfium threading crash, a Vision stall, and an XY-cut infinite loop on mirrored glyphs.

Reproduce:

python bench/real_prep.py funsd sroie doclaynet omnidocbench    # after downloading the datasets to bench/real/
python bench/real_run.py foliodoc doclaynet                        # resumable; one JSON line per document
bench/.venv-docling/bin/python bench/real_run.py docling doclaynet
python bench/real_run.py folio-rapidocr funsd                   # cross-platform OCR configuration
python bench/real_score.py > bench/real/results.json

Synthetic: exact ground truth

bench/corpus.py generates documents where every character, heading and table cell is known: reports, two-column papers, invoices and dense pages, each as a digital PDF, a clean 300 dpi scan, and a noisy 200 dpi scan (skew, blur, noise, JPEG). The table shows the held-out set (seed 1000, generated after tuning; 36 files, 78 pages).

On these synthetic documents docling's character error is high for two reasons confirmed by diffing: it reorders blocks in two-column layouts, and on justified text it cuts off the last word of some lines ($94,118.11 → $94,11). The synthetic set is narrower than real documents, which is why the real-world results above are the ones to trust.

Layout

foliodoc/pdf.py       pdfium text layer (chars → segments with words), vector rulings, images, scan extraction
foliodoc/ocr.py       OCR engines behind one interface (apple / rapidocr / tesseract)
foliodoc/raster.py    deskew, rule detection, weak-line re-read, ink-gap splitting, stroke-based styling
foliodoc/layout.py    rows, ruled/unruled tables, XY-cut reading order, paragraphs, headings, lists, furniture, stitching
foliodoc/docx.py      DOCX straight from XML
foliodoc/convert.py   orchestration, streaming OCR pool
bench/             corpus generator, scorer, runner, per-document inspector