LangGraph Agent with Human-in-the-Loop

A flexible LangGraph agent built with LangChain components, featuring:

• Multi-provider LLM support (Databricks, OpenAI, Azure OpenAI, Anthropic, Ollama)
• Human-in-the-loop middleware for tool approval and answer verification
• MLflow integration for deployment (optional)
• Modular architecture for easy maintenance and extension
• Environment-agnostic - runs in Databricks or standalone mode

Setup

This project uses Poetry for dependency management.

Prerequisites

• Python 3.10 or 3.11
• Poetry (install via: curl -sSL https://install.python-poetry.org | python3 -)

Installation

# Install dependencies
poetry install

# Activate the virtual environment
poetry shell

# Or run commands directly
poetry run python agent.py

Dependencies

Core:

• langchain - LangChain framework (v1.0.0+)
• langchain-core - LangChain core (v1.0.0+)
• langgraph - LangGraph for agent workflows
• python-dotenv - Environment variable management

Optional (provider-specific):

• databricks-langchain - For Databricks provider
• langchain-openai - For OpenAI/Azure OpenAI providers
• langchain-anthropic - For Anthropic provider
• langchain-community - For Ollama provider
• mlflow-skinny[databricks] - For MLflow integration (optional)

Quick Start

Configuration

The agent supports configuration via environment variables or a .env file:

# .env file
LLM_PROVIDER=azure_openai
LLM_ENDPOINT_NAME=gpt-4-deployment
AZURE_OPENAI_ENDPOINT=https://your-resource.openai.azure.com/
AZURE_OPENAI_API_KEY=your-api-key
AGENT_EXECUTION_MODE=standalone

Basic Usage

from agent import agent

config = {"configurable": {"thread_id": "1"}}
result = agent.invoke({
    "messages": [{"role": "user", "content": "List all schemas"}]
}, config)

Human-in-the-Loop

The agent includes built-in human-in-the-loop middleware:

• SQL queries require approval before execution (configurable)
• Final answers can be verified before sending to users

See USAGE.md for detailed usage instructions and AZURE_OPENAI_SETUP.md for Azure OpenAI configuration.

Architecture Decisions

Architecture Decision Records (ADRs) are documented in docs/adr/:

• ADR-0001: Use LangGraph for Agent Orchestration

Term Expansions Configuration

The search_tables_by_description tool uses a configurable term expansion system to map natural language descriptions to database column search terms.

Managing Term Expansions

Term expansions are stored in term_expansions.json. To add or modify expansions:

1. Edit the JSON file directly:

   {
     "your_term": ["term1", "term2", "term3", "synonym1", "synonym2"]
   }

2. The file is automatically loaded and cached - no code changes needed.

3. For large-scale management, you can optionally load from a Databricks table:

   • Uncomment the table loading code in _load_term_expansions()
   • Create a table with columns: term (string), expansions (JSON string)
   • Update the SQL query to point to your table

Example

To add support for "financial" terms:

{
  "financial": ["financial", "finance", "fiscal", "money", "currency", "dollar", "euro", "payment", "transaction"],
  "payment": ["payment", "pay", "transaction", "financial", "invoice", "billing"]
}

The system will automatically:

• Cache expansions for performance
• Expand search terms when users query
• Search across all schemas in the catalog