DOCS/task_log.md

Upgrade: Keyword-Match → LLM Intent Detection

• Backend: LLM Intent Detector

  • Add detect_intent(query) → calls LLM → returns {intent, entities}
  • Define intent types: trace_order, show_orders, billing_analysis, explain, fraud_check, etc.

• Backend: Intent-Based SQL

  • Replace keyword generate_sql() with generate_sql_from_intent(intent, entities)
  • Deterministic SQL per intent, no more if "trace" in q

• Backend: LLM Summary

  • Add generate_llm_summary(query, result) for clean natural language answers
  • Fallback to rule-based if LLM fails

• Backend: Conditional Highlights

  • Only return highlights for trace/graph intents
  • Empty highlights for show_orders, billing_analysis, etc.

• Frontend: Smart Graph Control

  • Only highlight when highlights array is non-empty
  • Clear previous highlights for non-graph queries

• Verification

  • curl test all intents
  • Browser demo

• Feature: Visual Flow Trace in Chat

  • Add renderFlowTrace component to ChatPanel.jsx
  • Render document chain with arrows (Order → Delivery → Billing → Journal)
  • Support clicking on IDs in the flow to highlight them in the graph
  • Sync main.py and main_OPENROUTER_FINAL.py for correct flow data output

• Feature: Resizable Chat Panel

  • Add chatWidth state to App.jsx
  • Implement draggable resize handle between Graph and Chat panels
  • Update ChatPanel.jsx to accept and use dynamic width
  • Add min/max width constraints for better UX

• Feature: Monorepo Refactor for Deployment

  • Create backend/ and frontend/ directories
  • Move backend logic and DB to backend/
  • Create backend/requirements.txt
  • Update frontend to use import.meta.env.VITE_API_URL
  • Create frontend/.env for local development
  • Verify local run in browser
  • Commit all changes locally (Monorepo + Updates)
  • Final Push to GitHub (Success 🚀)
  • Fix Vercel 404: Add frontend/vercel.json and push
  • Fix API 422: Refactor /query from GET to POST