OCR

Run GLM-OCR on this Mac through Ollama. Page images never leave the machine. The script writes the result to a Markdown file and prints only a summary, so the text stays out of the conversation.

Instructions

1. Take the file path from the arguments. If none was given, ask for one.
2. Pick the mode:
   • text (default) for ordinary pages.
   • table when the image or page is entirely a table: a screenshot or crop of one, a statement page or a spreadsheet. It returns an HTML <table>. On a mixed page of text plus a table, it reads the text and skips the table (tested). Use text there, or crop the table first.
   • formula for maths.
3. Run the script with run_in_background: true. The Bash tool's 2-minute default timeout is shorter than a 3-page run. Add -p only for a page range, and -o only when the user names an output folder:

bash ~/.claude/skills/ocr/scripts/ocr.sh -m <mode> [-p 1-12] [-o <folder>] "<file>"

• Default output: ~/Documents/obsidian/00_inbox/<name>_ocr.md.
• Private sources (under ~/Sync/ or a personal_records folder) never go into the vault, because it auto-pushes to GitHub.
  • Under ~/Sync/, the output goes next to the source.
  • Anywhere else, the script stops and asks for -o. Ask the user where to write. Never suggest the vault or the Google Drive mirror.
• It never overwrites an existing file.
• It starts its own Ollama server on 127.0.0.1:11435 with a 16k context and stops it on exit. Ollama's default 4k context cuts every page off after about a paragraph.
• Expect about 20 seconds to load the model, then roughly 40 seconds per dense page on an M3 with 16 GB (11 pages took about 7 minutes).

4. When the background task finishes, report the summary lines as printed.

Privacy Rules

• Do not Read, cat or grep the output file unless the user asks. Reading it sends the text to Anthropic; say so first when the source is private.
• Don't run pdfinfo on a private file without | grep '^Pages:', because its Title and Author fields can name people.
• Never run ollama pull, or ollama run with a model that isn't installed (it downloads automatically).
• Don't start Ollama as a login service (brew services start ollama).

Troubleshooting

• "model 'glm-ocr' is not installed": the user downloads it (2.7 GB) by running ollama serve in a separate Terminal window, then ollama pull glm-ocr in another, then stopping the server with Ctrl-C.
• "something is already serving on 127.0.0.1:11435": another /ocr run is in progress, or an earlier run was killed and left its server behind. Check with lsof -i :11435. Wait for the other run, or ask the user before stopping that process.
• "could not render the PDF": the PDF is password-protected or damaged, or the -p range starts past the last page. Check the page count with pdfinfo "<file>" | grep '^Pages:'.
• Gaps in the output:
  • Pages marked TRUNCATED: usually the model got stuck repeating itself. Re-run those pages with -p.
  • Tables missing from mixed pages: Ollama runs the model without GLM-OCR's separate layout step. Crop the table (pdftoppm -x -y -W -H) and run -m table on the crop.
  • "did not attach the image": Ollama didn't recognise the page path. Check ollama --version and report it.