sunrisever/glm-ocr

GLM-OCR Skill

When to Use

Use this skill when the task involves:

• OCR for PDF / PPT / PPTX / images
• Converting textbooks, lecture slides, or screenshots into Markdown
• Verifying whether OCR output is actually complete
• Investigating legacy mixed output, failed placeholders, or silent OCR corruption

Project Layout

• input/ — source files waiting for OCR
• output/ — Markdown output, extracted images, and any _failed_segments/*.failed.json
• _cache/ppt_pdf/ — cached PDFs converted from PPT/PPTX
• ocr.py — main OCR pipeline
• verify_ocr.py — acceptance check
• audit_ocr_integrity.py — deep integrity audit
• reference_book_metadata.py — textbook directory page, page offset, and QR resource generation
• backfill_reference_book_directory_pages.py — batch refresh textbook metadata
• rerun_pdf_segments.py — rerun only failed or suspicious page ranges
• duplicate_image_reviewer.py — local UI for duplicate/similar image review
• clean_junk_images.py — duplicate audit, similarity search, purge, and legacy size-clean fallback
• markdown_cleanup.py / repair_math_delimiters.py — OCR-side Markdown and LaTeX delimiter cleanup
• KNOWLEDGE_PIPELINE.md — source-library -> OCR intermediate -> Obsidian note workflow

Required Workflow

1. Put source files into input/.
2. Run python ocr.py.
3. Run both python verify_ocr.py and python audit_ocr_integrity.py.
4. For textbooks/reference books, generate 目录页.md, page offsets, and QR metadata via reference_book_metadata.py or backfill_reference_book_directory_pages.py.
5. Only treat the batch as complete when both checks are clean and the corresponding output directory has no _failed_segments/*.failed.json.
6. Before downstream note generation, use the OCR intermediate output rather than importing raw OCR folders directly into the final knowledge vault.

Failure Semantics

• The fallback chain is segment PDF upload -> per-page image OCR -> native PDF text fallback.
• If a segment still fails after all fallbacks, the pipeline writes _failed_segments/*.failed.json.
• Failed placeholders and failed segment reports mean the OCR result is incomplete. Do not silently pass or move such output downstream.

Common Problems and Recommended Handling

• 1301 contentFilter: split the segment first, rerun the blocked page separately, then use a secondary OCR / vision path only for the blocked page if needed.
• Legacy segment_*.md: do not delete until ranged .md coverage and content have been compared.
• Garbled file names from ZIP/RAR extraction: fix source names first, then keep input/, output/, and downstream library names in sync.
• Empty native PDF text fallback: common on scanned books; be ready to switch to image OCR or another vision path.
• Formula delimiters like $ 2x+1 $: clean them at OCR output time, not later in the note-writing stage.
• Header/footer strips, logos, watermarks, and glyph fragments: suppress them before they become images/.
• Ordinary orphan images should be audited first; do not blindly delete every unreferenced image.
• If a page is filled through a non-GLM route, mark it as AI visual supplementation (non-GLM-OCR output) or equivalent instead of pretending it came from the main OCR pipeline.
• For the exact marker format and audit rules, follow OCR_AUDIT_POLICY.md.

Notes

• PPT/PPTX temporary PDFs belong in _cache/ppt_pdf/, not in output/.
• verify_ocr.py is the minimum acceptance gate.
• audit_ocr_integrity.py should be used whenever you need confidence that there is no legacy mixed output or silent corruption.
• For the full downstream note workflow, also read KNOWLEDGE_PIPELINE.md.