What it does
DiffPaca compares two revisions of a PDF (construction drawings, blueprints, design plans; also spreadsheets, patents, question banks) and returns a verified change report: every changed region with visual evidence crops, OCR citations, a self-verification status (verified / needs_human), and a description in Japanese, English and Chinese.
A deterministic diff engine guarantees recall: no changed pixel escapes detection. The language-model layer only explains detected changes and never invents new ones. It never sees full pages, only cropped regions, so confidential documents stay compartmentalized.
Operations and pricing
| Operation | Price (USDC) | Mode | Returns |
|---|---|---|---|
POST /v1/compare | 0.50 | async, 2–4 min | task id; poll /v1/tasks/{id}, then /result and evidence files |
POST /v1/compare/preview | 0.005 | sync, 10–30 s | region count, bounding boxes, annotated pages (no interpretation) |
Max 50 MB per file. Payment is settled only after a successful response.
Privacy
- Input PDFs are deleted immediately after processing; task directories are purged after 24 hours.
- The model (Vertex AI, Tokyo region, no training on inputs) receives cropped evidence regions only, never a full page.
- Logs hold metadata only (page count, size, duration), never document content.
Full policy: Privacy.
Endpoint
https://diff-api-586747566167.asia-northeast1.run.app