parse-anything 0.1.0

One interface to parse any document: pick an engine (mineru, docling, unstructured, markitdown, lite) and get clean markdown out.

parse-anything

One interface to parse any document. Pick an engine — MinerU, Docling, unstructured, MarkItDown, or the PyMuPDF lite path — pass a file path, get clean markdown/text back. OCR runs on PaddleOCR (PP-OCRv5, default) or Tesseract.

from parse_anything import parse

result = parse("report.pdf", engine="miner-u")
print(result.content)          # markdown
print(result.duration_s)       # seconds

parse-anything report.pdf --engine miner-u -o report.md
parse-anything scan.png  --engine ocr --ocr paddle
parse-anything deck.pptx                     # auto-picks the best installed engine

Install

Core has zero dependencies; each engine is an optional extra:

pip install -e .[all]                        # everything
pip install -e .[mineru]                     # MinerU (scanned/complex PDFs)
pip install -e .[docling]                    # Docling (structure fidelity)
pip install -e .[unstructured]               # unstructured (broadest formats)
pip install -e .[markitdown]                 # MarkItDown (office/web -> md)
pip install -e .[lite]                       # PyMuPDF fast path
pip install -e .[paddle]                     # PaddleOCR (PP-OCRv5)
pip install -e .[tesseract]                  # pytesseract (+ install the tesseract binary)

Windows Tesseract binary: UB-Mannheim build or choco install tesseract.

Engines

--engine	Backed by	Reach for it when
mineru / miner-u	MinerU	scanned or complex PDFs — formulas, tables, multi-column; OCR built on PaddleOCR models
docling	Docling	you care about reading order, tables, office formats
unstructured	unstructured	odd formats: email, html, mixed batches
markitdown	MarkItDown	fast docx/pptx/xlsx/html → markdown (no OCR)
lite / liteparse	PyMuPDF4LLM	speed on PDFs that already have a text layer
ocr	PaddleOCR / Tesseract	images or fully scanned docs, raw text out
auto (default)	—	picks the best installed engine for the file type

OCR

--ocr paddle (default) uses PaddleOCR PP-OCRv5 — the current open-source benchmark leader. --ocr tesseract is the portable fallback. --ocr none skips OCR (text-layer extraction only).

Logging

Progress prints to stderr by default (auto-selected engine, OCR page-by-page, live MinerU subprocess output with a heartbeat on long runs). -v for debug detail, -q to silence it. As a library, nothing prints unless you call logging.basicConfig(...) yourself.

See SKILLS.md for the full skill contract (inputs, outputs, failure modes).

Layout

src/parse_anything/
├── __init__.py        # parse(path, engine=..., ocr=...) public API
├── factory.py         # engine registry, aliases, auto-selection
├── result.py          # ParseResult
├── cli.py             # parse-anything CLI
├── engines/           # one module per backend (lazy imports)
└── ocr/               # paddle + tesseract backends, PDF rasterization

Tests

pip install -e .[dev]
pytest