paddleocr-doc-parsing

name

paddleocr-doc-parsing

description

PaddleOCR document parsing guidance for PDFs and images hosted on Baidu AI Studio. Use when the user wants to extract text or structured markdown from scans, screenshots, photos, tables, or PDFs and needs help choosing PaddleOCR options such as fileType, orientation correction, unwarping, layout detection, chart recognition, restructurePages, mergeTables, relevelTitles, prettifyMarkdown, and visualize. This skill is guidance, not a callable tool: read this file, then use the bundled Node scripts via exec to test the configured PaddleOCR connection or run document parsing against local files, URLs, data URLs, or base64 content.

PaddleOCR Doc Parsing

Use this skill when the user wants OCR or structured document parsing from images or PDFs and OpenClaw should actually run PaddleOCR instead of only describing the API.

The bundled scripts accept local files and remote PDF/image links. For remote links, they fetch the file first and send Base64 to PaddleOCR, which is more reliable than assuming the hosted API can dereference the URL itself.

Critical Rule

This skill is guidance, not a callable tool.

• Do not call paddleocr-doc-parsing as if it were a tool name.
• First use read to load this SKILL.md.
• Then use exec to run the bundled Node script with node.
• If the user expects OCR output, actually run the script. Do not stop at parameter advice.

Trigger Cues

Reach for this skill when the user asks for things like:

• OCR this image or screenshot
• extract text from this PDF
• parse this scan into markdown
• read a document from an image link or PDF URL
• pull text or structure from a receipt, table, report, or form

Script Directory

Determine this SKILL.md directory path as SKILL_DIR, then use these scripts:

Script	Purpose
scripts/test-connection.mjs	Verifies the saved PaddleOCR endpoint and token with a sample file
scripts/parse-document.mjs	Runs PaddleOCR doc parsing and returns JSON plus merged markdown

Both scripts are plain Node ESM and should be run with node.

Config Resolution

The scripts resolve credentials and defaults in this order:

1. CLI flags
2. PADDLEOCR_ENDPOINT / PADDLEOCR_TOKEN
3. ocr.providers.paddleocr in ~/.openclaw/openclaw.json
4. models.providers.baidu-aistudio.apiKey as the fallback token

If the user already configured PaddleOCR in ClawMaster or OpenClaw, prefer running the scripts without repeating endpoint and token on the command line.

Default Workflow

1. Diagnose the source type. Use fileType: 0 for PDFs and fileType: 1 for images.
2. Pick the smallest useful preset. clean-pdf for exported documents, mobile-scan for phone photos, layout-debug when layout ordering is wrong.
3. Verify the connection if the endpoint or token is new.
4. Run the parse script.
5. Summarize the parsed markdown and call out any layout or table limitations.

If the user asks for "markdown only", "recognized text only", or "no explanation", return only the parsed markdown from the script.

Commands

# Verify the saved OCR config against the default sample image
node ${SKILL_DIR}/scripts/test-connection.mjs

# Parse a local PDF with the clean PDF preset
node ${SKILL_DIR}/scripts/parse-document.mjs ./invoice.pdf --preset clean-pdf --markdown-out /tmp/invoice.md

# Parse an image URL with mobile scan settings
node ${SKILL_DIR}/scripts/parse-document.mjs https://example.com/photo.jpg --preset mobile-scan

Important Flags

Flag	Meaning
--file <value>	Local path, PDF/image URL, data URL, or raw base64. Positional file input also works.
`--file-type <image	pdf
`--preset <clean-pdf	mobile-scan
--output <path>	Writes the full JSON result to disk.
--markdown-out <path>	Writes merged markdown text to disk.
--visualize	Requests debug images from PaddleOCR.

Boolean option flags map directly to the PaddleOCR request body:

• --use-doc-orientation-classify
• --use-doc-unwarping
• --use-layout-detection
• --use-chart-recognition
• --restructure-pages
• --merge-tables
• --relevel-titles
• --prettify-markdown
• --visualize

Use --no-... to force any of those options off.

Preset Guidance

• clean-pdf Use for normalized PDFs that need readable markdown, cross-page table merge, and title cleanup.
• mobile-scan Use for camera captures, warped paper, or screenshots with mixed blocks.
• layout-debug Use when the user needs visualization output to inspect segmentation or ordering mistakes.

Response Expectations

When the script returns:

1. Report which preset or options were used and why.
2. Surface page count and whether layout images were generated.
3. Quote or summarize the important markdown output.
4. If parsing quality is weak, suggest the next option change instead of retrying blindly.