Convert Document To Markdown

Use this skill when a user wants a supported local file converted into Markdown for later processing.

What this skill does

• Converts supported local files into Markdown:

.pdf, .docx, .pptx, .xlsx, .jpg, .jpeg, .png, .gif, .bmp, .txt, .json, .xml, .md

• Image handling modes are file-type dependent:

ocr / vl / none for .docx, .pptx, .xlsx, and image files;

ocr / vl / vl-page / none for .pdf

• Only runs through Docker. Do not use local Python execution as an operational path.
• Uses a prebuilt Aliyun CR image with fixed version 0.0.1:

convert-document-to-markdown-arm64:0.0.1 on ARM64 hosts,

convert-document-to-markdown-x64:0.0.1 on x64 hosts

• Returns structured JSON by default so later tool calls can consume markdown, logs, and meta.
• Reads one-time VL configuration from OpenClaw skill config or the repository .env file, then forwards it into the container automatically.
• Only exposes the file command. URL, health, and version commands are intentionally removed to keep startup lean.
• Do not use latest, do not build a fallback image at runtime, and do not treat .doc, .ppt, .xls, audio files, or unlisted image formats as supported inputs.

Required workflow

1. By default the scripts use crpi-4auaoyyj6r36p6lb.cn-hangzhou.personal.cr.aliyuncs.com/huozige_lab.
2. Let the wrapper script resolve the host architecture and choose convert-document-to-markdown-arm64:0.0.1 or convert-document-to-markdown-x64:0.0.1.
3. If needed, override with IMAGE_REGISTRY or IMAGE_NAME.
4. For a local file, run:

scripts/run_docker_cli.sh file --format json

1. Parse the JSON result.
2. If success is false, surface error.message and relevant logs.
3. If success is true, use markdown as the canonical output for downstream work.

One-time VL configuration

This skill is designed so the user does not need to re-enter Vision API settings on each run.

Preferred OpenClaw configuration in ~/.openclaw/openclaw.json:

{
  "skills": {
    "entries": {
      "convert_document_to_markdown": {
        "enabled": true,
        "apiKey": "sk-xxx",
        "env": {
          "VL_BASE_URL": "https://api.openai.com/v1",
          "VL_MODEL": "gpt-4.1-mini"
        }
      }
    }
  }
}

This works because:

• skillKey is convert_document_to_markdown
• primaryEnv is VL_API_KEY, so apiKey maps to VL_API_KEY
• env can hold VL_BASE_URL and VL_MODEL

Repository-local runtime configuration:

• copy .env.example to .env
• fill VL_BASE_URL, VL_API_KEY, and VL_MODEL
• by default the scripts use crpi-4auaoyyj6r36p6lb.cn-hangzhou.personal.cr.aliyuncs.com/huozige_lab
• optionally override with IMAGE_REGISTRY or IMAGE_NAME
• use scripts/run_docker_cli.sh, which loads .env, forwards any host VL_* variables into docker run, and pulls the correct fixed-version image if missing

Command patterns

Local file:

scripts/run_docker_cli.sh file ./notes.pdf --image-process-model ocr --format json

Parameters

• --image-process-model ocr

Default mode. Use Tesseract OCR for images.

• --image-process-model vl

Use a Vision API. Only choose this when the environment provides VL_API_KEY and related variables.

• --image-process-model none

Skip image recognition for speed.

• --image-process-model vl-page

PDF only. Do not use this mode for Office documents or image files.

• --format json|markdown

Use json unless the user explicitly wants raw Markdown on stdout.

• --output

Save the Markdown to a file. Prefer this only when you invoke docker run directly with a writable host mount.

• --log-file

Save detailed logs to a file. Prefer this only when you invoke docker run directly with a writable host mount.

Operational notes

• For very large local files, stay with the Docker CLI path; do not wrap the file content into base64 or a temporary HTTP service.
• The skill is Docker-only. Do not instruct users to run uv, python, or any other local runtime path for production use.
• The wrapper scripts choose the image by host architecture. Override with IMAGE_ARCH only when you have a concrete reason.
• Prefer IMAGE_REGISTRY plus the fixed version 0.0.1; only use IMAGE_NAME when you need to pass the full image reference explicitly.
• When the user asks for VL or VL-page, first check whether VL_BASE_URL, VL_API_KEY, and VL_MODEL are already configured via OpenClaw skill config or .env.
• If the user only needs extracted Markdown and not the raw JSON wrapper, read the JSON and return the markdown field.
• If the user provides an unsupported extension such as .doc, .ppt, .xls, .wav, .mp3, .m4a, or .mp4, say the current skill does not reliably support it.

Safety notes

• Treat file paths as untrusted input. Quote shell arguments correctly.
• Do not claim success unless the command returns success: true.