Imported from VectorSpaceLab/AREX-Skill (skills/repositories/repo-skills/docling/sub-skills/extraction/SKILL.md). Install upstream with npx skills add VectorSpaceLab/AREX-Skill --skill extraction. Copyright stays with the author (MIT).

Docling Structured Extraction

Use this sub-skill when the user wants structured field extraction from PDF or image documents with DocumentExtractor, extraction templates, and page-level extraction results.

Route elsewhere when the task is primarily:

• General document conversion, parsing, or DoclingDocument creation: use the conversion sub-skill.
• Output serialization to Markdown, JSON, YAML, HTML, text, doctags, VTT, or Docling JSON: use the document-outputs sub-skill.
• VLM backend selection, accelerator configuration, remote service wiring, or model/download strategy: use the advanced-pipelines sub-skill.

Quick Use

• Install an extraction-capable Docling package variant before runtime extraction; VLM extraction needs model/backend dependencies beyond a minimal import-only install.
• Start with DocumentExtractor(allowed_formats=[InputFormat.IMAGE, InputFormat.PDF]) because default extraction backends are configured for images and PDFs.
• Pass a template as a JSON-like string, dict, Pydantic model instance, or Pydantic model class.
• Validate result.pages[*].extracted_data with your own Pydantic model before trusting fields downstream.
• Keep fixture runs explicit and local; do not assume model artifacts are already downloaded unless the environment owner has prepared them.

References

• references/api-reference.md covers DocumentExtractor, extraction options, template types, and result models.
• references/workflows.md provides recipes for small field extraction, Pydantic validation, and safe fixture runs.
• references/troubleshooting.md lists common failures including unsupported formats, missing model extras, template mismatches, and remote opt-in.
• scripts/extract_fixture.py is a safe helper for local image/PDF fixture extraction; it prints help and never downloads by default.