pdf-ocr

PDF OCR via GLM-OCR + Ollama

Convert PDF files to markdown using local OCR. Each page is rendered to an image, processed through GLM-OCR running on Ollama, and the recognized text is assembled into a single markdown file.

Workflow

Phase 1: Dependency Check

Run these checks and report a status table. If any required dependency is missing, use AskUserQuestion to offer installation.

Dependency	Check	Install
Ollama running	curl -sf http://127.0.0.1:11434/api/tags > /dev/null	Cannot auto-start. Tell user to start Ollama and stop.
glm-ocr model	Check that Ollama tags response contains glm-ocr	ollama pull glm-ocr
pdftoppm	which pdftoppm	brew install poppler
pdfinfo	which pdfinfo	Comes with poppler
sips	which sips	Built-in on macOS. Warn if missing.

Check all five in a single Bash call:

echo "=== Ollama ===" && curl -sf http://127.0.0.1:11434/api/tags && echo "" && echo "=== pdftoppm ===" && which pdftoppm && echo "=== pdfinfo ===" && which pdfinfo && echo "=== sips ===" && which sips

Parse the output:

• If curl fails: Ollama is not running. Stop and tell user to start Ollama.
• If glm-ocr is not in the tags list: ask user if they want to pull it (ollama pull glm-ocr).
• If pdftoppm or pdfinfo missing: ask user if they want to install poppler (brew install poppler).
• If sips missing: warn user (built-in on macOS, no auto-install).

If all checks pass, proceed to Phase 2.

Phase 2: Execute OCR

1. Validate the argument is a path to an existing .pdf file. If no argument was provided, print usage and stop:

   Usage: /pdf-ocr <path-to-pdf> Converts a PDF to markdown using local GLM-OCR via Ollama.

2. Determine the output path: same directory as the PDF, same base name with .md extension.

3. Get the page count to decide foreground vs background execution:

pdfinfo "<pdf_path>" | grep "^Pages:"

4. Run the OCR script:

python3 ${CLAUDE_PLUGIN_ROOT}/skills/pdf-ocr/scripts/pdf_ocr.py "<input.pdf>" "<output.md>"

• 10 pages or fewer: Run in foreground (timeout 600000ms / 10 min).
• More than 10 pages: Run in background so the user can continue working. Use timeout: 600000.

Phase 3: Report Results

After the script completes:

1. Read the first 50 lines of the output .md file to show a preview.
2. Report:
   • Page count processed
   • Output file path
   • Total character count (from the script's stdout)
   • Approximate time taken (from the script's stdout)

Key Technical Details

• Images are rendered at 72 DPI then resized to max 1024px (the sweet spot for GLM-OCR accuracy)
• The prompt must be exactly "Text Recognition:" — other prompts degrade quality
• Each page gets a <!-- Page N --> marker with --- separators between pages
• The script uses only Python 3 stdlib (no pip dependencies)
• Temp files are created in /tmp/claude/ocr_pages and cleaned up per-page