pdf-extract-skill（PDF 提取技能）

SKILL: OpenClaw PDF Supercharger with OpenDataLoader

0) Modular Map (.md)

To improve maintainability and allow targeted calls to specific .md files, this skill relies on helper documents:

• CLI quick start: docs/quickstart-cli.md
• Security before install: docs/security-before-install.md
• OpenClaw ready profiles: docs/profiles-openclaw.md
• Hybrid + OCR: docs/hybrid-mode-ocr.md
• RAG and bounding-box citations: docs/rag-citations.md
• Troubleshooting: docs/troubleshooting.md

Usage rules:

• If the task is setup/startup: load quickstart-cli.md
• Before any installation: load security-before-install.md
• If the task is command execution by scenario: load profiles-openclaw.md
• If the task involves scanned or complex table PDFs: load hybrid-mode-ocr.md
• If the task is RAG/citations: load rag-citations.md
• If there are errors: load troubleshooting.md

1) Goal

This skill maximizes PDF reading quality for OpenClaw in ClawHub using OpenDataLoader PDF.

Pillars:

• Local extraction (no cloud) for privacy.
• High-quality reading order and structure (columns, tables, layout).
• RAG and LLM-ready outputs (json + markdown).
• Simple end-user flow (CLI, no MCP).

2) When to Use This Skill

Use this skill when the user needs to:

• Extract clean text from PDFs.
• Improve table and multi-column parsing.
• Prepare data for RAG, embeddings, or coordinate-based citations.
• Process scanned PDFs with OCR.
• Describe images/charts to make them searchable.

Do not use this skill for:

• OCR of standalone image files outside PDF workflows.
• Cloud-only pipelines where local Java execution is not allowed.

3) Core Architecture Rule (No MCP)

Since the MCP does not exist yet, this skill must operate with CLI only:

• Client command: opendataloader-pdf
• Hybrid backend command: opendataloader-pdf-hybrid

Do not create complex wrappers or intermediate services unless strictly needed.

4) Robust Prerequisites

Always validate before conversion:

• Java 11+ in PATH.
• Python 3.10+.
• Package install policy:

- Do not use unpinned installs in production. - Use isolated environments (venv/container/VM). - Prefer pinned versions and verified sources. - See: docs/security-before-install.md

Quick checks:

• java -version
• pip index versions opendataloader-pdf
• pip show opendataloader-pdf
• opendataloader-pdf --help

If Java fails on Windows, reopen the terminal and verify PATH.

5) Standard OpenClaw Operating Flow

Step A: Classify user intent

1. General reading/summary -> markdown
2. RAG with metadata and citations -> json,markdown
3. Complex tables or scanned PDF -> hybrid docling-fast
4. Charts with image descriptions -> hybrid + hybrid-mode full + enrich-picture-description

Step B: Run in batches (required)

Always process multiple files in a single invocation to avoid JVM startup overhead per call.

Recommended example: opendataloader-pdf file1.pdf file2.pdf ./folder/ -o ./output -f json,markdown

Step C: Return a simple OpenClaw response format

Suggested response:

1. Status: ok or warning
2. Processed files
3. Output path
4. Generated formats
5. Suggested next action

Template: "Processing completed. N PDFs were converted to ./output with json,markdown format. If you want, I can now extract specific pages or enable OCR for scanned files."

6) Ready-to-Use CLI Profiles

Profile 1: Fast LLM reading

opendataloader-pdf ./pdfs/ -o ./output -f markdown

Profile 2: Recommended for RAG

opendataloader-pdf ./pdfs/ -o ./output -f json,markdown

Profile 3: Specific pages only

opendataloader-pdf report.pdf -o ./output -f json --pages "1,3,5-7"

Profile 4: Sensitive data sanitization

opendataloader-pdf report.pdf -o ./output -f markdown --sanitize

Profile 5: Preserve line breaks

opendataloader-pdf report.pdf -o ./output -f markdown --keep-line-breaks

Profile 6: Embedded or external images

opendataloader-pdf report.pdf -o ./output -f json --image-output external opendataloader-pdf report.pdf -o ./output -f json --image-output embedded

7) High-Precision Hybrid Mode

Use it when:

• Tables are complex or borderless.
• PDFs are scanned.
• Multi-language OCR is required.
• Image/chart descriptions are required.

7.1 Start backend

Standard: opendataloader-pdf-hybrid --port 5002

Forced OCR: opendataloader-pdf-hybrid --port 5002 --force-ocr

Multi-language OCR: opendataloader-pdf-hybrid --port 5002 --force-ocr --ocr-lang "es,en"

With image descriptions: opendataloader-pdf-hybrid --port 5002 --enrich-picture-description

7.2 Use backend from client

Hybrid auto mode: opendataloader-pdf --hybrid docling-fast file1.pdf file2.pdf ./folder/ -o ./output -f json,markdown

With timeout and fallback: opendataloader-pdf --hybrid docling-fast --hybrid-timeout 120000 --hybrid-fallback file1.pdf ./folder/ -o ./output -f json

Image descriptions enabled (full required): opendataloader-pdf --hybrid docling-fast --hybrid-mode full file1.pdf ./folder/ -o ./output -f json,markdown

Critical note: If the backend starts with --enrich-picture-description, the client must use --hybrid-mode full to include descriptions in output.

8) Key Robustness Parameters

• -f, --format: json, text, html, pdf, markdown, markdown-with-html, markdown-with-images
• --pages: page range (example: "1,3,5-7")
• --sanitize: anonymizes emails, phone numbers, IPs, cards, and URLs
• --reading-order xycut: keeps proper column reading order (recommended default)
• --use-struct-tree: improves extraction on tagged PDFs
• --table-method cluster: improves complex table detection
• --hybrid-url: backend endpoint (local by default)
• --hybrid-timeout: timeout in milliseconds (0 = no timeout)
• --hybrid-fallback: continue with Java mode if backend fails

9) Decision Matrix for OpenClaw

1. If the user wants speed and clean text: markdown.
2. If precise positional citations are needed: json (with bounding box) or json,markdown.
3. If output is empty/poor on scanned files: backend with --force-ocr.
4. If tables are very complex: enable --hybrid docling-fast.
5. If charts must be interpreted: backend with --enrich-picture-description and client with --hybrid-mode full.

10) Quick Troubleshooting

Problem: Java not found. Solution: install Java 11+ and verify with java -version.

Problem: Hybrid backend connection error. Solution: start opendataloader-pdf-hybrid in another terminal and verify port 5002.

Problem: Too slow. Solution: process in batches, increase hybrid timeout, and verify backend RAM.

Problem: Mixed columns. Solution: use default reading mode (xycut) and try --use-struct-tree for tagged PDFs.

Problem: Poor table quality. Solution: use json output + hybrid mode.

11) Best Practices for ClawHub

• Prioritize simple, predictable commands.
• Answer users with short, actionable steps.
• Recommend json,markdown as default for assistants and RAG.
• Keep safety filters enabled (do not disable content safety unless explicitly requested).
• Store output in run-specific folders for traceability.

12) Skill Quality Checklist

• CLI-only architecture defined (no MCP).
• Clear installation and prerequisites.
• Execution profiles ready to copy/paste.
• Full hybrid flow with OCR and image descriptions.
• Fallback and troubleshooting strategy.
• Simple communication format for OpenClaw.

13) Credits

This skill uses and credits the excellent OpenDataLoader project: https://opendataloader.org/

Official documentation used for this version: https://opendataloader.org/docs