Overview

This archive is a self-contained quickstart for text2sql ReAct agents on Databricks. It provides Text2SQL question-answering functionality on financial datasets found in the data_setup directory

Files

• agent_config.yaml: Contains key parameters for the ReAct Agent, including:
  • Model inference params: model endpoint name, temperature, max_tokens, etc.
  • SQL warehouse params: warehouse id
  • Databricks resource params: hostname, catalog, schema
• 01_sql_react_agent contains code for the text2sql ReAct agent, implemented using LangChain
• 02_evaluate: Tests/runs the code in 01_sql_react_agent. It contains code that:
  • Logs the chain to MLflow
  • Implements mlflow.evaluate() to evaluate agent against a benchmark dataset
  • Registers the chain to Unity Catalog
  • Deploys the chain to a serving endpoint and starts a review UI
• data_setup: Contains data (csv) files for tables and notebook to create Delta tables for testing

Setup

1. Update agent_config.yaml with the Databricks Resources (warehouse, hostname, catalog, schema), model resources (model endpoints, temperature, max tokens, etc.), and base prompt
2. Run data_setup to create data tables for testing
3. Review and customize 01_sql_react_agent as needed
4. Test the agent code using 02_evaluate. Once the code is stabilized, register the model, run evaluation, register and deploy.

Requirements

• Permissions to create schemas and tables in Databricks
• Permissions to deploy model serving endpoints
• Enablement of AI-assisted features on your Databricks workspace