document-parse

name

document-parse

description

Parse a raw file (PDF, DOCX, XLSX, PPTX, HTML, or scanned image) into a structured document representation (UDOM). Use this atom when a workflow needs to extract text, tables, and layout from an uploaded or crawled file BEFORE applying governance or analysis. Do NOT use for web page scraping — this handles file-based documents only. Do NOT use for reading-level estimation (call atom-reading-level on the extracted text).

document-parse

Parse raw file bytes into a structured document representation (UDOM tree). Used by document-intelligence to separate file parsing from governance, and reusable by standards-updater or any skill that needs to extract content from uploaded documents.

Input

{
  "file_path": "/path/to/uploaded.pdf",
  "file_type": "pdf",
  "parser_hints": {"ocr_engine": "pytesseract", "lang": "en"},
  "confidence_threshold": 0.8
}

Output

{
  "tool": "document-parse",
  "udom": {"type": "document", "children": [{"type": "heading", "text": "Chapter 1"}, {"type": "paragraph", "text": "..."}]},
  "confidence": 0.95,
  "recovery_note": "Full text extraction; no OCR gaps",
  "parser_used": "PyMuPDF",
  "unrecovered_regions": [],
  "human_review_required": true
}

Do NOT use this atom for

• Web page scraping (use source-crawl or feed-validate)
• Reading level estimation (pass extracted text to atom-reading-level)
• Document comparison or diff (this atom extracts, it does not compare)

Pipeline note

Follows references/method.md at the Ingestion step (raw file → UDOM). Output conforms to references/metadata-schema.md. Downstream steps (governance, analysis) are handled by the orchestrator. human_review_required: true — OCR/layout recovery is imperfect; teacher should verify critical content.