Vanna Multi-Agent Query Router

A minimal FastAPI app that demonstrates how to run multi-agent, multi-database workflows using Vanna.

📹 Demo Video

Watch the video walkthrough here:

✨ What it does

• Takes a natural language query and a start_agent ("agent_1" or "agent_2")
• Calls the first agent and enforces a consistent output using a standardized background prompt
• Extracts a shared identifier (email) from the first agent's output
• Uses those emails to query the second agent
• Returns both steps in the response

🤖 Why it's useful

LLMs are inconsistent without structure. This app shows how to:

• Standardize agent outputs using a background prompt (email column)
• Chain LLM-based agents across different data sources
• Ensure reproducible, structured multi-agent workflows

🧠 Key idea: background prompt

Both agents get a custom instruction + a shared standardization prompt like:

Always include a column named 'email' that uniquely identifies a user.

This ensures that:

• Output from one agent is reliably usable by the other
• The interface between agents is stable and predictable

🚀 Deploy to Cloud Run

gcloud builds submit --tag gcr.io/YOUR_PROJECT_ID/vanna-multi-agent
gcloud run deploy vanna-multi-agent \
  --image gcr.io/YOUR_PROJECT_ID/vanna-multi-agent \
  --platform managed \
  --region us-central1 \
  --allow-unauthenticated \
  --set-env-vars VANNA_API_KEY=your-api-key USER_EMAIL=your-email@example.com

🧪 Example Request

POST /query
{
  "start_agent": "agent_1",
  "query": "show me the top users by engagement"
}

🗂 Files

• main.py: Main FastAPI app
• requirements.txt: Dependencies
• Dockerfile: For containerizing
• .dockerignore: For cleaner builds
• .env.example: Example environment config
• openapi.yaml: OpenAPI spec for GPT tool use

---

Let us know if you'd like to contribute or extend this to other Vanna agent workflows!