Converting PDF to Markdown in Python

Four libraries, what each is good at, and the licence question people usually find out about too late.

The quickest working version

pymupdf4llm is a thin wrapper over PyMuPDF that outputs Markdown directly, and for most text PDFs it is all you need:

pip install pymupdf4llm

import pymupdf4llm
md = pymupdf4llm.to_markdown("report.pdf")
pathlib.Path("report.md").write_text(md)

It detects headings from font size, reconstructs tables, and handles multi-column layouts reasonably well. On a three-page document expect around 0.1 seconds.

The licence catch

PyMuPDF is AGPL-licensed. For internal scripts and personal work that is fine. If you ship it inside a product you distribute, or run it as a network service, the AGPL's terms reach further than most people expect — check before you build on it commercially. A paid commercial licence is available from Artifex.

When to reach for something else

Scanned PDFs

None of these read a scan, because a scan has no text layer to extract. You need OCR first — Tesseract via pytesseract, or PyMuPDF's built-in OCR support, which shells out to Tesseract too. Detect the case rather than returning an empty file:

doc = pymupdf.open(path)
chars = sum(len(p.get_text().strip()) for p in doc)
if chars < doc.page_count * 40:
    ...  # it's a scan — OCR required

Things worth doing to the output

Raw extraction is rarely what you want to hand to a model or a wiki:

• Strip running headers and footers. A line appearing at the top or bottom of most pages is boilerplate. Keep the first occurrence so you never lose the document's only copy of its title.
• Strip page numbers — including combined forms like Page 4 of 32, which a naive digits-only rule misses.
• Repair mojibake. Broken font encodings turn en-dashes into replacement characters; a small substitution table fixes most of it.

Questions people ask