Document to Markdown

Convert PDF, images, Office docs, and more to clean Markdown using the MinerU Open API CLI. No API key needed for basic use.

Language Rule

Reply to the user in the SAME language they use. This is non-negotiable.

Core Workflow

Extraction is often just the first step. The typical flow is:

1. Extract — Use mineru-open-api to convert the document to Markdown
2. Read & Process — Help the user with what they actually need

MinerU outputs raw Markdown — it doesn't interpret or restructure the content. If the user asks to "extract the tables", "summarize the paper", or "find the key findings", you need to read the output and do that work yourself. MinerU handles the OCR and layout; you handle the understanding.

Use -o to save to a file when the user wants persistent output (conversion, batch processing). Skip -o and read stdout directly when the content is consumed immediately (summarization, Q&A).

For example:

• "帮我把这个PDF转成markdown" → use -o to save to file, done
• "提取这篇论文里的表格" → use -o to save, then read the file and pull out the tables
• "这篇论文讲了什么" → stdout is fine, read the output directly and summarize
• "把PDF里的参考文献整理出来" → stdout or -o, then parse the references section

Two Extraction Modes

flash-extract — Fast, no auth

Best for quick reads. No API key, no setup.

mineru-open-api flash-extract report.pdf                               # to stdout (for immediate consumption)
mineru-open-api flash-extract report.pdf -o ./output/                  # save to file
mineru-open-api flash-extract report.pdf -o ./output/ --pages 1-10     # page range
mineru-open-api flash-extract report.pdf -o ./output/ --language en    # language hint
mineru-open-api flash-extract https://example.com/paper.pdf            # URL input

Supports: PDF, images (PNG, JPG, WebP...), DOCX, PPTX, Excel (XLS, XLSX) Limits: 10 MB / 20 pages per document Output: Markdown only — images, tables, and formulas may become placeholders

Use flash-extract as the default unless the user needs more.

extract — Precision, auth required

Use when the user needs full-fidelity output: preserved images, accurate tables, LaTeX formulas, or non-Markdown formats. Requires a token via mineru-open-api auth.

mineru-open-api extract report.pdf                              # to stdout
mineru-open-api extract report.pdf -o ./out/                    # save with all assets
mineru-open-api extract report.pdf -o ./out/ -f md,docx         # multiple output formats
mineru-open-api extract report.pdf -o ./out/ --ocr          # force OCR for scanned docs
mineru-open-api extract *.pdf -o ./results/                 # batch processing
mineru-open-api extract --list files.txt -o ./results/      # batch from file list

Supports: PDF, images, DOC, DOCX, PPT, PPTX, HTML Limits: 200 MB / 600 pages per document Output formats: md, json, html, latex, docx (comma-separated with -f) Features: formula recognition (on by default), table recognition (on by default), OCR toggle, batch mode, model selection (vlm, pipeline, html)

If the user hasn't authenticated yet, guide them to run mineru-open-api auth first.

When to Use Which

Language Support

Default is ch (Chinese + English). Use --language to specify others. Common codes:

80+ languages supported in total — use the PaddleOCR language code for any language not listed above.

Data Flow

Both commands send the document to MinerU's API (mineru.net) for processing. This is a stateless API call with no persistent storage. MinerU is open-source by OpenDataLab (Shanghai AI Lab): https://github.com/opendatalab/MinerU

Troubleshooting

• Debug API requests: Add -v flag to see HTTP request/response details (e.g., mineru-open-api flash-extract report.pdf -v)
• CLI not found: Install via one of:
  • npm i -g mineru-open-api (Node.js)
  • uv tool install mineru-open-api (Python/uv)
  • macOS/Linux: curl -fsSL https://cdn-mineru.openxlab.org.cn/open-api-cli/install.sh | sh
  • Windows: irm https://cdn-mineru.openxlab.org.cn/open-api-cli/install.ps1 | iex
• Auth error on extract: Run mineru-open-api auth to set up your token
• Timeout on large files: Increase with --timeout 600 (seconds)
• Wrong language output: Set --language explicitly (e.g., --language en for English docs)