docparse

DocParse - Document Parsing with Visual Citations

Parse documents and answer questions with spatially-grounded visual citations. Every cited fact is traced to an exact location on the source page.

Variables

• $ARGUMENTS - file path(s) or directory, optional question in quotes
• Examples:
  • /docparse datastore/corporate/presentations/deck.pdf
  • /docparse datastore/books/ "What are the key principles?"
  • /docparse report.pdf --report

Phase 0 - Setup & Context

1. Run prerequisite check:

   python "${CLAUDE_PLUGIN_ROOT}"/scripts/docparse.py setup --check
   

   If any check fails, print the install instructions and STOP.

2. Parse $ARGUMENTS:

   • Extract file path(s) or directory
   • Extract optional question (quoted string)
   • Check for --report flag or if user explicitly asks for "visual citations" / "visual report"
   • Check for --pages specification (e.g., --pages 1-10)

3. If a directory was given, discover supported files:

   # Use Glob to find PDFs, DOCX, PPTX, XLSX in the directory
   

4. For documents over 50 pages, ask the user which page range to parse. Never parse 100+ pages without an explicit --pages restriction.

Phase 1 - Parse

Run the parser:

python "${CLAUDE_PLUGIN_ROOT}"/scripts/docparse.py parse \
  --files "<file1>" "<file2>" \
  [--pages "1-10"] \
  [--dpi 150] \
  --output-json /tmp/docparse_parsed.json

Read the output JSON and present a summary:

• Number of files and pages parsed
• Cache hits vs fresh parses
• Total text items extracted

Phase 2 - Analysis

If the user asked a question:

1. Read /tmp/docparse_parsed.json focusing on each page's text field. Do NOT read raw text_items arrays - those are for the bounding box engine only.

2. Analyze the text to answer the question. Identify specific passages that support the answer.

3. For each cited passage, extract an exact verbatim quote from the parsed text:

   • Character-for-character exact (this is critical for bounding box lookup)
   • Prefer short quotes under 60 characters
   • Quote evidence (numbers, data values, key phrases), not just labels
   • Include 5-15 citations per answer