Tracker/Code/TODO.md

TODO

Phase 0: Infrastructure (DONE)

• Azure VMs created and configured (jt-ub, jt-win)
• Ubuntu VM provisioned (Docker, Go, Python, rclone, zsh, dotfiles)
• Directory structures on both VMs (William's layout)
• Sample document generator (7 generators, Azure Claude content pools, CJK fonts)
• Project restructure: Tracker/Setup.bat + Tracker/Code/{Sync,Tools,Tracker}
• Windows rclone sync setup (Setup.bat, Win-Setup.ps1, sync.ps1, config.ini)
  • rclone auto-download + self-contained install in _LLM/Code/Sync/rclone/
  • Password-based SFTP auth (demo), config.ini for server details
  • .sync_ignore generated on setup (excludes_LLM/, dotfiles, temp/lock files)
  • 5-minute scheduled task (JingTian-Sync)
  • Pull-first-then-push: bidirectional files pulled with --update, excluded from push
  • Bidirectional file list in config.ini (e.g. Admin/Tracker.xlsx)
  • Daily log rotation in _LLM/Code/Sync/logs/
• Ubuntu VM SSH password auth enabled (Azure cloud-init was blocking it)
• Both sync directions tested and working
• Old JingTian-Rclone repo archived on Gitea

Phase 1: Services (Services-First Approach)

Rationale: De-risk unknowns (Docling OCR quality, Ollama CPU inference speed) before building orchestrator.

1a. Docker Compose + Docling (DONE)

• Docker Compose base (Docling service)
• Test Docling API with sample PDFs (curl)
• Validate OCR quality on real PDFs + generated samples
• Test OCR on images (PNG) — works with force_ocr=true

Docling API Spec:

• Image: ds4sd/docling-serve:latest
• Port: 5001
• Endpoint: POST /v1/convert/file
• Form params: files=@<path>;type=application/pdf, to_formats=md, do_ocr=true
• Response: { "document": { "md_content": "..." }, "status": "success" }
• OCR engine: easyocr (default), supports CJK

OCR Options:

• do_ocr=true — enabled by default, processes bitmap content
• force_ocr=true — use for images/scanned PDFs (replaces existing text with OCR)
• ocr_engine — easyocr (default), tesseract, rapidocr, tesserocr, ocrmac
• ocr_lang — language codes (engine-specific), e.g. ch_sim for Simplified Chinese

Tested Formats:

• PDF (Filing Receipt) — extracted text, tables, CJK content ✅
• PNG (Email screenshot) — extracted subject, dates, recipient ✅ (CJK trademark garbled, may need ocr_lang)

1b. Ollama (DONE)

• Add Ollama to Docker Compose (Qwen3-1.7B, CPU mode)
• Test structured extraction prompt (curl)
• Measure inference time per document (~15-30s on 4 vCPU)
• Validate extraction accuracy (deadline, doc type, etc.)

Ollama API Spec:

• Image: ollama/ollama:0.9.3
• Port: 11434
• Model: qwen3:1.7b (~1.4 GB, auto-pulled on startup)
• Endpoint: POST /api/generate
• Body: {"model":"qwen3:1.7b","prompt":"...","stream":false,"options":{"temperature":0}}
• Response: { "response": "<json>" } (strip <think>...</think> tags in post-processing)
• Note: Qwen3 includes reasoning by default; /no_think leaves empty tags, so strip instead

1c. Python Tools Service

• FastAPI + openpyxl service
• Endpoints: read tracker, write/update rows, get column schema
• Test with sample Tracker.xlsx

1d. Manual Integration Test

• Chain test: PDF → Docling → Ollama → Tools → Excel updated
• Document any issues / adjustments needed

Phase 2: Go Orchestrator

• Init Go module in Tracker/
• Config loading (YAML)
• SQLite schema + migrations (documents, extractions, processing_log)
• Docling HTTP client
• Ollama HTTP client
• Tools service HTTP client
• File watcher (inotify on Ubuntu, 2-min stability check via SHA256 hash)
• State tracker integration (new/changed/skip logic)
• Pipeline orchestrator (read latest Tracker before writing, merge not overwrite)

Phase 3: Integration + Testing

• End-to-end test (drop file on Windows → syncs to Ubuntu → pipeline → Tracker updates → syncs back)
• Smoke test with real PDFs + generated samples
• README documentation

Future

• Key-based SFTP auth (replace password for production)
• Windows Go service (replace PowerShell scheduled task)
• File hash (SHA256) change detection on Windows side
• WhatsApp deadline notifications
• Processing retry logic
• Multi-client support
• Smarter bidirectional merge (cell-level, not file-level)