paperextract

Turn a folder of scientific PDFs into a searchable, citable, machine-readable library, locally, without losing the equations, tables or figures.

paperextract is built for technical literature, where a misread exponent or a shifted table cell matters. It runs state-of-the-art document models on your own machine or cluster, keeps every result traceable to the page it came from, and tells you what it could not verify instead of guessing.

paper.pdf  ──►  Travers_2019_HighEnergyPulse/
                ├── paper.md          Markdown with LaTeX equations, tables, figures
                ├── document.json     every block with its page, position and source
                ├── figures/ tables/  crops, captions, tables as HTML, JSON and CSV
                ├── citation.bib      identity validated against Crossref or DataCite
                ├── validation.json   what could not be verified
                └── original/         your PDF, byte for byte

Features

• Equations, tables and figures as data. Display and inline math as LaTeX, table cells exactly as printed, figures with their panels and captions, supplements linked to the main text.
• Nothing invented. Values are never repaired or filled in; original files are preserved with hashes; every doubt is recorded as a finding.
• Real bibliographic identity. DOIs, titles and authors checked against Crossref, DataCite and arXiv, with BibTeX, clear names, and document versions (preprint, accepted manuscript, published) kept apart.
• Duplicates caught before extraction. Identical files, copies with the same text and shared DOIs are found before any model runs.
• A library you can search and share. Full-text search, lookup by DOI or file, library.bib, portable folders with relative links, and an agent skill so AI assistants can find and cite your papers.
• Figure descriptions. Optional, clearly labelled machine descriptions of every figure from a local vision-language model or the Claude API.
• From laptop to cluster. MinerU, Docling or Marker in isolated environments on a Mac or Linux; GPU batches on Slurm clusters, resumable and published in one step. 162 papers took under an hour on four GPUs.

Install

Requires uv and macOS on Apple silicon or Linux on x86_64.

git clone https://github.com/jtravs/paperextract.git && cd paperextract
uv sync --locked
uv sync --directory workers/mineru --locked       # add --extra cuda on an NVIDIA GPU
uv run paperextract models fetch mineru           # or mineru-cuda; pinned and verified

Use

uv run paperextract paper.pdf --library ~/papers          # one paper
uv run paperextract batch incoming/ --library ~/papers    # a folder, duplicates handled
uv run paperextract search "soliton self-compression" --library ~/papers
uv run paperextract describe --all --library ~/papers     # optional figure descriptions

Documentation

The manual is at paperextract.readthedocs.io:

• Getting started: installation, a first extraction and what it produces
• Worked example: a real paper followed from PDF to library, with the frozen input and output in examples/
• Command line: every command and setting
• Running on a Slurm cluster: GPU batches
• Figure descriptions
• Known limitations: read before relying on a result
• Model licences: the terms of each backend's models

To build it locally, run uv run poe docs and open docs/_build/html/index.html.

Status

paperextract is alpha software. The pipeline is complete and has been run on about 170 papers on macOS and on NVIDIA A40 GPUs; accuracy has not yet been measured against a reviewed reference set. See the changelog.

Contributing

uv sync --locked
uv run poe check      # lint, types, docstrings, spelling, tests, manual

The default test suite is offline and needs no models. See AGENTS.md for conventions and tooling for the checks.

Licence

Apache License 2.0; see LICENSE and NOTICE. paperextract distributes no model weights: each backend's models keep their own licences (summarized in model licences), and papers keep their publishers' copyright.