rapiddoc-cli

A command-line interface for RapidAI/RapidDoc — parse PDF / image / Office documents into Markdown / JSON / Word / HTML with a single command.

RapidDoc is a high-performance document parsing pipeline (MinerU fork, VLM removed, ONNX-based) that runs on CPU — no GPU required.

Install

# Recommended: isolated environment (doesn't pollute your system Python)
uv tool install rapiddoc-cli

# Or with pip
pip install rapiddoc-cli

Both expose the rapiddoc command. The first run downloads ONNX models (~50–100 MB) to a cache directory; subsequent runs reuse them.

Usage

rapiddoc paper.pdf                      # → paper.out.md (default: Markdown)
rapiddoc report.pdf -f json             # → report.out.json
rapiddoc notes.docx -f word             # → notes.out.docx (Office native parsing)
rapiddoc scan.png -f html               # → scan.out.html
rapiddoc book.pdf -o book.md --lang en  # English OCR + custom output path
rapiddoc scan.png --no-formula          # skip formula recognition (faster)

Options

rapiddoc <input> [options]

  -o, --output <path>               Output file path
  -f, --format <md|json|word|html>  Output format (default: md)
      --lang <ch|en|...>            OCR language (default: ch)
      --no-formula                  Disable formula recognition
      --no-table                    Disable table recognition
  -h, --help                        Show help

Supported inputs

Platform support

Why not Intel Mac?

This is not a limitation of rapiddoc-cli itself — it's an upstream constraint: the required dependencies onnxruntime and openvino (2025.1+) no longer publish wheels for macOS x86_64. As a result pip install rapiddoc-cli will fail on Intel Macs because the underlying model-inference libraries can't be installed.

Intel Mac users can run rapiddoc-cli inside a Linux container/VM (e.g. Docker with a Linux image, or a Linux cloud VM), where x86_64 wheels are fully supported.

How it works

rapiddoc-cli is a thin wrapper around the rapid-doc Python package. It:

1. Initializes a RapidDoc engine (OCR, layout, formula, table models loaded on demand)
2. Parses the input into an intermediate middle_json representation
3. Renders the requested output format:
   • Markdown / JSON: produced directly by the engine
   • Word / HTML: Markdown is generated first, then converted via markdown_to_docx / markdown_to_html (pypandoc / markdown-it-py)

License

Apache-2.0 (same as upstream rapid-doc).