ghazaljohri/agentic-router-pro

Agentic Model Router – Pro

This Space runs a FastAPI app with multiple models routed automatically through a Docker backend.

🧩 Features

• —One endpoint (/query) that auto-routes to the best model (or force a task)
• —API Key authentication (header x-api-key)
• —Prometheus metrics at /metrics (request count, latency, task labels)
• —Optional Redis response cache
• —Minimal, pretty web UI
• —Tests + Dockerfile

⚙️ Environment

• —Python 3.11 recommended
• —macOS Apple Silicon (M1/M2/M3/M4): torch from PyPI supports CPU and Metal (MPS). Pip will choose the right wheel.

🚀 Run (Local)

python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
export API_KEY=dev-secret-123   # change this
./run_local.sh
# open http://localhost:8000

UI calls require the key; enter it in the top-right field (the UI includes it on requests).

🔐 API Usage

curl -X POST http://localhost:8000/query   -H 'Content-Type: application/json'   -H 'x-api-key: dev-secret-123'   -d '{"text":"Translate this to Hindi: I am learning AI.","task":"auto"}'

📈 Metrics

Visit http://localhost:8000/metrics for Prometheus exposition format. Add to Grafana via Prometheus data source.

🐳 Docker

docker build -t agentic-router-pro .
docker run -p 8000:8000 -e API_KEY=dev-secret-123 agentic-router-pro

☁️ Deploy

• —Render: Docker Web Service → set API_KEY, (optional) ENABLE_REDIS=true, REDIS_URL=....
• —Hugging Face Spaces: Space (Docker) → expose port 8000 → set API_KEY secret.

🧪 Tests

pytest -q

Notes

• —For Apple Silicon, PyTorch will use CPU by default. To try Metal (MPS), set PYTORCH_ENABLE_MPS_FALLBACK=1; pipelines will still run even if ops fall back to CPU.
• —You can swap the heuristic router with a learned classifier for extra credit.