MinerU

pdftotext gives you the words in roughly the wrong order. Classic OCR gives you the words in a scan, with the table turned into soup and the equation into punctuation. Both are fine until you need to use the document — feed it to a model, put it in notes, diff two revisions of a spec — and then the structure you lost was the point.

MinerU parses the document instead of reading characters off it: it recovers reading order, keeps tables as tables, formulas as LaTeX, and figures as images, and writes Markdown and JSON. Version 2.5 moved to a compact vision-language model — around 1.2B parameters — that scores at the top of the public document-parsing benchmarks while being small enough to run on a modest GPU, or on CPU through the pipeline backend.

Install and run link

  pip install uv
uv pip install -U "mineru[all]"

mineru -p report.pdf -o out/
mineru -p ./inbox -o out/ -b pipeline     # CPU-only machines
  

Output lands in out/ as Markdown, a JSON structure with the blocks and their coordinates, and the extracted images. The Markdown is what you read; the JSON is what you build on when you need positions or block types.

Python 3.10–3.13, 16 GB RAM as a floor, and about 20 GB of disk for a full install with models. There’s a Gradio web UI (mineru-gradio), a REST API (mineru-api), Docker images, and an MCP server if you want an agent to call it directly.

Choosing a backend link

The pipeline backend is the honest default for a first run: it’s fast, stable, and — the line from the project’s own docs worth repeating — it doesn’t hallucinate. A VLM reconstructing a smudged table can produce a plausible number, which is a different and more dangerous failure than an obviously mangled one.

What it’s genuinely good at link

• Academic and technical PDFs. Multi-column reading order, footnotes, and formulas as LaTeX rather than as mangled symbols.
• Tables. Structure preserved as HTML or Markdown, including tables that span pages.
• Scanned documents, with recognition across 109 languages.
• Office formats. DOCX, PPTX, and XLSX go through the same pipeline, so one command handles a mixed folder.

The obvious use is feeding a RAG index or an agent: chunking Markdown with real headings beats chunking a wall of extracted text, and it’s the difference between retrieval that finds the right section and retrieval that finds the right page.

Local processing is the other half of that. The documents worth parsing are often the ones you can’t upload to an API — contracts, patient records, internal specs — and this runs on your own machine.

Check the output before you trust it link

Every document parser is a probability model wearing a CLI. Two habits:

• Spot-check the numbers. Tables and figures are where errors hide, and a wrong digit in a Markdown table reads exactly like a right one. Verify anything you’ll make a decision on against the original page.
• Keep the source. Store the PDF alongside the Markdown with the page mapping from the JSON, so a surprising claim can be traced back rather than argued about.

Licence, and the alternatives link

MinerU ships under its own licence, based on Apache 2.0 with additional conditions. That’s more permissive than the AGPL it used to carry, and “additional conditions” is exactly the phrase to read carefully before building a commercial product on it rather than assuming.

Reach for the small tools first. pdftotext on a clean digital PDF is instant and exact; MinerU earns its install when the layout is the information.

Next link

The same job with a wider mouth and a pipeline attached → Docling