Multi-Agent System

Overview

Build a multi-agent customer service system where a Supervisor Agent intelligently routes questions between a Genie Space (for SQL-based data queries) and a Knowledge Assistant (for document-based policy answers). Deploy the system behind a Gradio chat interface on Databricks Apps.

Skills used: databricks-genie, databricks-agent-bricks, databricks-model-serving, databricks-app-python MCP tools used: create_or_update_genie, manage_ka, manage_mas, query_serving_endpoint, create_or_update_app

Prerequisites

• A Databricks workspace with Unity Catalog and Model Serving enabled
• Tables for the Genie Space (e.g. main.sales.transactions, main.sales.customers)
• PDF or text documents for the Knowledge Assistant (e.g. company policies in a Volume)
• A SQL warehouse for Genie Space queries

Steps

1. Create a Genie Space for data queries

   Set up the SQL-based agent that handles product, order, and revenue questions.

   Create a Genie Space called "Sales Explorer" that connects to
   main.sales.transactions and main.sales.customers with sample questions:
   "What was total revenue last month?", "Who are our top 10 customers?",
   "Show monthly revenue trends".

2. Create a Knowledge Assistant for documents

   Set up the RAG-based agent that handles policy and FAQ questions.

   Create a Knowledge Assistant that answers questions about our company's
   policies using documents stored in /Volumes/main/support/policy_docs/.
   Add instructions to always cite the specific document and section in
   its answers.

3. Build a Supervisor Agent

   Create the orchestration layer that routes questions to the right specialist.

   Create a Supervisor Agent that routes questions to:
   1. A "Sales Analyst" Genie Space for revenue, order, and customer questions
   2. A "Policy Assistant" Knowledge Assistant for company policy, HR, and
      FAQ questions
   Give the supervisor instructions to identify the question type before routing,
   and to combine information from both agents when a question spans both domains.

4. Verify the agent endpoint

   Test the Supervisor Agent is online and routing correctly.

   Check the status of my Supervisor Agent endpoint. Once it's online, test it
   with these questions:
   - "What was our top-selling product last quarter?" (should route to Genie)
   - "What is our return policy?" (should route to KA)
   - "How do refunds affect our revenue numbers?" (should use both)

5. Deploy as a Gradio chat app

   Build a user-friendly chat interface for the multi-agent system.

   Build a Gradio app on Databricks that provides a chat interface to my
   Supervisor Agent endpoint. Include:
   - A chat window with message history
   - An indicator showing which agent is answering each question
   - A sidebar with suggested questions for each agent type

What You Get

• Genie Space that converts natural language to SQL against your sales tables
• Knowledge Assistant that answers questions from your policy documents with citations
• Supervisor Agent that intelligently routes questions to the right specialist
• Gradio app with a chat UI deployed on Databricks Apps

Next Steps

• Add MLflow Evaluation to measure routing accuracy and answer quality
• Expand with additional agents: a Model Serving endpoint for product recommendations
• Store conversation history in Lakebase for analytics
• Use Vector Search for a custom embedding index with more control over retrieval