A production-grade multi-agent AI platform that handles customer support queries autonomously and escalates to humans when needed.
- Classifies incoming queries by topic and confidence score
- Resolves high-confidence queries autonomously using a ReAct agent with tools
- Scores response quality with a critic agent before sending
- Escalates low-confidence or low-quality responses to a human queue with full context
- Tracks resolution rate, escalation rate, and latency on a live dashboard
| Layer | Technology |
|---|---|
| LLM | Groq llama-3.3-70b-versatile |
| Embeddings | sentence-transformers (local) |
| Vector DB | ChromaDB (local) |
| Session store | Redis |
| Agent framework | LangGraph |
| Backend | FastAPI + WebSocket |
| Frontend | React + TypeScript |
# 1. Clone and enter the project
git clone <repo-url>
cd support-agent
# 2. Copy env file and add your Groq key
copy .env.example .env
# 3. Install uv (if not already installed)
powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"
# 4. Install dependencies
uv sync
# 5. Start Redis (must be running before the backend)
redis-server
# 6. Start the backend
uv run uvicorn backend.main:app --reloadAPI docs: http://localhost:8000/docs
| Phase | Name | Status |
|---|---|---|
| 0 | Foundation + Knowledge Base | Not started |
| 1 | Intent Classifier + Router | Pending |
| 2 | Autonomous ReAct Agent + Tools | Pending |
| 3 | Confidence Checker + Escalation | Pending |
| 4 | Eval Dashboard + Frontend | Pending |
See ARCHITECTURE.md for the full system diagram and data flow.