Databricks-Multi-agent-workflow: Prada Localized Shop Analytics (PLSA)

Example of multi agent workflow in databricks

1. System Overview

An enterprise multi-agent system designed for Shop Managers to access localized inventory, sales, and visual merchandising data using native Databricks tools. The system utilizes a hub-and-spoke orchestration model where a Supervisor routes queries to specialized agents.

2. Project Structure

Databricks-Multi-agent-workflow/
├── README.md               # Project documentation
├── databricks.yml          # Databricks Asset Bundle (DAB) configuration
├── config.yaml             # Centralized configuration for agents (Genie IDs, VS indexes, UC functions)
├── Makefile                # Automation scripts for deployment and cleanup
├── requirements.txt        # Python dependencies
├── .env                    # Local environment variables (not committed)
├── src/                    # Application source code
│   ├── app.py              # Streamlit Chatbot UI with thinking trace visualization
│   ├── orchestrator.py     # LangGraph workflow and Supervisor routing logic
│   ├── agents/             # Specialist agent implementations
│   │   ├── inventory_analyst.py    # Databricks Genie integration
│   │   ├── visual_merchandiser.py  # Mosaic AI Vector Search integration
│   │   └── creative_analyst.py     # Unity Catalog Functions as tools
│   └── utils/              # Shared utilities
│       └── config_loader.py        # Environment-aware configuration helper
├── setup/                  # Infrastructure and setup scripts
│   ├── DDL.sql             # SQL for creating demo tables and functions
│   └── deploy_instructions.md
└── .gitignore

3. Agent Orchestration (The Orchestrator)

The system is powered by LangGraph, implementing a Supervisor/Specialist pattern found in src/orchestrator.py.

Supervisor Node

The supervisor_node acts as the intelligent router. It uses ChatDatabricks to analyze the user's intent and select the most appropriate specialist:

• Inventory: Routed to the Genie Space analyst.
• Merchandiser: Routed to the Vector Search specialist.
• Creative: Routed to the UC Functions analyst.
• General: Handled directly for greetings or simple queries.

Workflow State

The system maintains an AgentState containing the query, message history, routing decision (next_step), and the final_response.

---

4. Sub-Agent Details

🛍️ Inventory Analyst (src/agents/inventory_analyst.py)

• Core Technology: Databricks Genie.
• Function: Provides a conversational interface over structured SQL data.
• Implementation: Uses GenieAgent to invoke a specific genie_space_id defined in config.yaml.

🖼️ Visual Merchandiser (src/agents/visual_merchandiser.py)

• Core Technology: Mosaic AI Vector Search.
• Function: Retrieves visual display guidelines and brand placement rules.
• Implementation: Performs similarity searches on a Databricks Vector Search index using DatabricksVectorSearch.

💡 Creative Analyst (src/agents/creative_analyst.py)

• Core Technology: Unity Catalog Functions & Tool Calling.
• Function: Complex sales analysis and multi-table interrogation.
• Implementation: Uses UCFunctionToolkit to wrap SQL/Python functions as LangChain tools. It operates as a ReAct agent to autonomously find product details before running analysis.

---

5. Deployment & Execution

The project is built to run as a Databricks App managed via DABs.

Prerequisites

• Databricks CLI configured for your workspace.
• Unity Catalog access with permissions for Vector Search and UC Functions.

Deployment Commands

make deploy

6. Observability

• MLflow Tracing: Integrated throughout the workflow. Each node in the LangGraph and every tool call is traced automatically.
• Trace Visualization: Traces can be viewed in the Databricks MLflow UI or captured directly within the Streamlit "Thinking" steps.