CaD-Track

Git-style change tracking for CAD drawings. Upload two versions (v1/v2) of the same drawing — PDF, PNG, or JPG — and get back the changes as added / removed / moved / modified regions, an annotated visualization, statistics, and a plain-English summary.

Unlike pixel-based diff tools, CaD-Track contains no pixel-comparison algorithm: it extracts the structured content of each drawing (vector geometry and text via PyMuPDF for native CAD exports, OCR text via Tesseract for scans) and diffs those entities directly. A global page shift, rescale, or DPI difference therefore cannot make the whole sheet read as "changed."

See docs/architecture.md for the full design and docs/requirements.md for the original requirements.

Features

• Accepts PDF (vector or scanned), PNG, JPG — auto-detects vector vs raster.
• Vector PDFs: exact geometry + text diff (primitives grouped into entities, matched by position and shape).
• Raster inputs: text/annotation diff via OCR plus approximate geometry diff — the drawing's ink is traced into connected structures on both sides and those entities are matched, so added/removed drawing elements are detected even in scans. Explicit warnings when a scan is too low-resolution to compare reliably.
• Change overlay, side-by-side view, per-region stats table, % area changed.
• Rule-based natural-language summary (no LLM, fully offline).

Setup

Requirements: Python 3.11+, Tesseract OCR (Windows: winget install UB-Mannheim.TesseractOCR; the backend finds the default install path automatically).

cd backend
pip install -r requirements.txt
python -m uvicorn app.main:app --reload

Open http://127.0.0.1:8000 — the web UI is served at the root; interactive API docs at /docs.

Usage

1. Drop the old version into "Drawing A" and the new version into "Drawing B".
2. Click Compare.
3. Review the change overlay (green = added, red = removed, amber = moved, blue = modified), the statistics table, and the generated summary.

Sample inputs live in samples/real_pair/.

API

POST /api/upload                 multipart file_a, file_b -> { job_id }
POST /api/compare/{job_id}       run the comparison
GET  /api/jobs/{job_id}          status + diff + stats + summary
GET  /api/jobs/{job_id}/visualization?mode=bbox|side_by_side|original_a|original_b
GET  /api/jobs/{job_id}/summary
GET  /api/jobs/{job_id}/stats
DELETE /api/jobs/{job_id}

Tests

python -m pytest tests

Unit tests cover the matcher, stats math, and location/severity bucketing; integration tests run the full pipeline on synthetic vector-PDF and raster pairs with known expected diffs, plus a smoke test on the real sample pair.

Limitations (v1)

• Raster geometry comparison is approximate (connected ink structures): a change fused into a larger connected structure is only detected when it noticeably alters that structure's ink share. There is still no pixel diffing — traced entities are matched structurally.
• Low-resolution scans fuse fine detail into blobs; results carry a warning when the two sides' entity yields are badly mismatched.
• Single page: page 1 of each document is compared.
• Job store is in-memory; restart clears results.