Tracker/Code/README.md

JingTian-Tracker

Document processing pipeline for JingTian & Gongcheng IP/trademark workflow.

Pipeline

Windows VM                              Ubuntu VM
                                        (Docker Compose)
BenjaminTeam/                           
  user drops files ──► rclone sync ──►  inotifywait detects new files
                                            │
                                        State check (SQLite)
                                        - new file → process
                                        - changed file → re-process  
                                        - unchanged → skip
                                        - tracker.xlsx → skip
                                            │
                                        Docling (text extraction + OCR)
                                            │
                                        Ollama / Qwen 3-1.7B (structured extraction)
                                        → deadline, doc type, assigned person, etc.
                                            │
                                        Tools service (openpyxl)
                                        - read latest tracker.xlsx
                                        - append new rows / update re-uploaded
                                        - preserve manual edits
                                            │
                                        rclone sync tracker.xlsx back
                                            │
  tracker.xlsx updated ◄── rclone ◄──  done

Stack

Service	Role	Language
server	Orchestrator, watcher, state	Go
docling	Text extraction, OCR	Python
ollama	LLM structured extraction	-
tools	Excel read/write, extensible	Python
sqlite	State tracking (file, no container)	-

Key Rules

• LLM only writes to Excel on first processing or source doc re-upload
• Mr. Choi can edit any cell freely — edits are preserved
• _LLM/ directory is excluded from sync and file watching
• Windows watcher uses debounce (15s) and .last_sync to avoid feedback loops

TODO

See TODO.md.