中文 | English

docdistill

Document → text via VLM. An independent Python MCP server with two tools: document_to_pdf (Office→PDF via real renderers) and analyze_doc (PDF→Markdown/LaTeX/JSON via MinerU's VLM). Local file paths only, no CDN, images localized.

Predecessor: seed-viz's analyze_doc (Node), now independent and pure-Python.

Repo: https://gitcode.com/Joe-zhouman/docdistill

Install

In your MCP client (.mcp.json):

{
  "mcpServers": {
    "docdistill": {
      "command": "uvx",
      "args": ["docdistill"],
      "env": { "BACKEND": "mineru-official", "AUTH": "<your-mineru-token>" }
    }
  }
}

Get a free MinerU token (1000/day): https://mineru.net/apiManage/token

PyPI: https://pypi.org/project/docdistill/

Backends

mineru-tasks-v3.2 is for users who self-host MinerU (e.g. an intranet-shared deployment: one install serves a whole campus LAN). No rate limits, no auth, data stays local.

Platform support (v1)

• PDF extraction (analyze_doc): all platforms (hosted backend, any networked machine).
• Office→PDF (document_to_pdf): Linux (LibreOffice) + Windows (Word/PowerPoint COM) only. macOS users get an honest error for Office files — export to PDF manually, then analyze_doc.

CLI

docdistill document-to-pdf paper.docx        # → paper.pdf, hint to use analyze_doc
docdistill analyze-doc paper.pdf              # → .docdistill/paper.md + .docdistill/images-<hash>/
docdistill analyze-doc thesis.pdf --output-format latex
docdistill analyze-doc doc.pdf --backend mineru-tasks-v3.2 --endpoint http://127.0.0.1:5580

Output lands in .docdistill/ next to the source. The images directory uses a short hash name (images-<hash>) rather than the source filename, so image references render correctly even when the source PDF has spaces/non-ASCII in its name. PDFs >180 pages auto-paginate concurrently; partial page failures don't abort the whole run. Files over 200MB get an honest error (split manually).

Known behavior: VLM splits multi-subfigure figures

The VLM extracts purely from visual content and does semantic segmentation, not pixel-chopping. On a large figure with several subfigures (e.g. a paper's "Fig. 4 (a)–(e)"), it tends to split it into one image block per visually bounded subfigure. Conversely, a region that looks like a table (clear grid lines) is extracted as a Markdown table rather than an image — which is usually what you want.

This is the VLM's own behavior, not a docdistill bug, and there is no API switch to disable it. Switching to the non-VLM pipeline model avoids the split but noticeably degrades text quality, so docdistill uses VLM. If you need the original single-image figure, refer to the source PDF.

Status

0.2.0 beta. Interface may change before 1.0.