exactdoc

Turn a PDF into a Word document that looks like the original and that you can actually edit.

exactdoc writes a DOCX made of real headings, paragraphs, lists, tables, footnotes and page headers, and lays them out so each page matches the PDF. Its output is measured by opening it in Microsoft Word, LibreOffice and Google Docs and comparing every page with the original. It runs entirely on your machine: no telemetry, no account, and no network unless you ask for the optional Google Docs measurement (details).

_{Page 1 of one of the project's test whitepapers, converted with the default settings; Google Docs shows the DOCX made for it (still image, how it was made). Small differences remain: Google Docs adds a thin white strip above the dark band, and the footer's page number sits further left than in the PDF, set larger in Word and Google Docs.}

Beta (0.3.0b1), on PyPI. Not announced yet: see the results on documents it has never seen below.

Install

You need Python 3.9 or newer.

pip install --pre exactdoc
exactdoc --version

Optional, and recommended: LibreOffice (free). When it is installed, exactdoc opens its own DOCX in it, compares each page with the PDF and corrects page breaks and spacing. Without it, exactdoc converts in one pass and says so. docs/install.md covers the fonts LibreOffice needs on Linux, how long the check takes, and the optional extras.

Quick start

exactdoc report.pdf                                    # writes report.docx next to it
exactdoc report.pdf --output-profile gdocs --refine 0  # a DOCX made for Google Docs
exactdoc --input-dir pdfs --out-dir docx --recursive   # a whole folder

from exactdoc import convert
convert("report.pdf", "report.docx")                                  # with LibreOffice's check
convert("report.pdf", "report.docx", refine_rounds=0, oracle="none")  # without LibreOffice

exactdoc never replaces a file you did not name. The Google Docs option writes the same document in a form Google Docs' importer reads correctly; it is still made offline, and you upload it to Drive yourself. Every option, the Python API and the exit codes are in docs/usage.md.

What you get

The DOCX is built from the same pieces you would use to write the document in Word yourself, not from text boxes pinned to the page:

• Headings use Word's heading styles, so the navigation pane and Google Docs' outline work.
• Paragraphs flow and re-wrap when you edit them.
• Lists are real bulleted and numbered lists.
• Tables are real tables, with merged cells and rows that run over pages.
• Headers and footers repeat on every page, with live page numbers. Footnotes are real Word footnotes (plain text in the Google Docs output, because Docs moves real ones), and links still work.
• Charts, logos and artwork stay pictures, placed where the PDF has them.

What works, and what does not yet

What exactdoc promises depends on the program that made the PDF. Documents written as flowing text in Word, Google Docs, LibreOffice or LaTeX are promised. Designed layouts are not, yet. Numbers below are from the release measurement (record) and, where they say "never seen", from the test below (results).

Works: tables become real Word tables, merged cells included.	Not yet: in a dense arXiv paper (CC BY 4.0) the text converts, but equations come apart.

To see which program made a PDF, open File → Properties in a PDF viewer and read the Producer and Application fields, or run exactdoc --diagnose your.pdf, which prints them without any of the document's text. The full list, with numbers, is in docs/deep-dive/limitations.md; more side-by-side pictures are in docs/examples.md.

How we know

Measured, not eyeballed. Every change must pass 16 frozen test documents, gated on every push in CI inside the pinned measurement image, and each milestone converts 95 frozen PDFs (those 16 and 79 more, most of them public real-world documents), opens every DOCX in LibreOffice, Microsoft Word and Google Docs, and compares it with the PDF page by page. In the final measurement for the beta (record, scorecard):

• all 41 promised documents of up to 20 pages kept exactly their page count in all three;
• of the 21 promised documents over 20 pages, all 21 in LibreOffice and Word, and 18 in Google Docs, kept their page count within 2% (or one page) with at least 85% of their words on the right page;
• across everything, including the kinds marked "Not yet", Word opened all 90 DOCX files without asking to repair one, and in Google Docs 66 of the 90 compared came back with exactly the right number of pages. The other five are PDFs exactdoc refuses on purpose.

These are the documents exactdoc was developed against, so they show how well it does on PDFs it has seen. How the measurement works: docs/deep-dive/measured-state.md.

Tested on documents it has never seen

Before the release, 32 public PDFs never used in development were sealed and then converted once, on the release candidate. 23 were classed as promised before the run, including the web page now marked "Not yet". The verdicts were the same in LibreOffice and Word; Google Docs was not part of this test (results, data):

• No crashes, and Word opened all 32 cleanly: pass.
• Short promised documents: fail. 13 of 16 kept exactly their page count, but only 12 of 16 had at least 90% of their words on the right page, below the 80% bar. The four misses: a CMU sample résumé set (9 pages for 8), a federal-style résumé, a LaTeX résumé (88% of words) and a Wikipedia article printed from Chrome (8 pages for 7).
• No unusable result: fail. The federal-style résumé came back with 6 pages for 3 and about a third of its text on the right page.
• Long promised documents (reported): a median page-count error of 1.3%; 4 of 7 within 2% (or one page), and 3 of 7 with at least 85% of their words on the right page.

This did not meet the bar we set in advance for announcing it; 0.3.0b2 is the fix release.

Fonts

The DOCX names only fonts that a standard Windows 10 or 11 computer with Microsoft Office has, with one exception: Chinese, Japanese, Korean, Arabic, Hebrew, Persian and Thai text keeps the PDF's own font name, so a reader who has that font sees it and one who does not gets Word's default for the script. The fonts the test documents name this way are Noto Sans CJK JP, IPAPGothic, WenQuanYi Zen Hei, DFKai-SB, DejaVu Sans, Noto Naskh Arabic, David, Narkisim, BNazanin and Thonburi.

Contributing

PRs welcome. See CONTRIBUTING.md for the one-command quality check every PR runs. If a conversion goes wrong, report it with the output of exactdoc --diagnose your.pdf instead of the PDF itself.

Learn more

Licence

exactdoc is Apache-2.0. A default install pulls in no copyleft code. The optional mupdf extra (pip install "exactdoc[mupdf]") adds PyMuPDF, which is AGPL-3.0; it is only a reference for measurements and does not change the output. Details: docs/deep-dive/licensing.md and docs/license-audit.md.