A Generative AI–powered Multi-Agent System that enables customer support teams to query structured databases and unstructured documents using natural language.

Built with LangChain, LangGraph, and Streamlit with flexible support for Google Gemini, OpenAI, Anthropic, and local LLMs (via LM Studio).

🎥 Demo Video

Watch the demo

---

🏗️ Architecture

Agent Roles

Agent	Purpose	Data Source
Router	Classifies queries and routes to specialist agents	LLM (Configurable)
SQL Agent	Queries structured customer data via natural language	SQLite database
RAG Agent	Searches uploaded policy documents	ChromaDB vector store

---

🔒 Security Architecture

The GenAI Customer Assistant implements a multi-layered security architecture designed to prevent prompt injection, data exfiltration, and unauthorized actions. This is achieved using LangChain Middleware to establish firm execution guardrails.

1. Middleware Intent Classification (@before_agent):
   • Before the sql_agent or rag_agent process any request, a custom @before_agent hook evaluates the user's intent.
   • If malicious intent is detected (e.g., prompt injection, requests to bypass rules, requests to drop tables), the middleware intercepts the execution, skips the LLM tool call entirely, and immediately returns a safe refusal template.
2. Tool-Level SQL Execution Guardrails:
   • Raw SQL execution via the LLM is wrapped in a custom validation function.
   • Checks are performed before database execution to block SELECT *, ensure a LIMIT clause is present, and entirely prohibit DML (Data Manipulation Language) statements like INSERT, UPDATE, DELETE, and DROP.
3. PII Middleware Sanitization (PIIMiddleware):
   • Even if a query legitimately retrieves data, the output is passed through LangChain's built-in PIIMiddleware before returning to the LLM or user.
   • Emails are redacted (e.g., e***@email.com).
   • Phone numbers are masked (e.g., +1-***-****).
4. Strict System Instructions:
   • All agents are prepended with a non-negotiable SYSTEM RULES (HIGHEST PRIORITY) block that dictates behavior boundaries that cannot be overridden by user input.
5. Red-Team Tested: Includes an automated test suite (test_security.py) to continually verify defenses against prompt overrides, mass data dumps, and SQL injections.

---

🚀 Quick Start

Prerequisites

• Python 3.10+
• LLM API Key (Google Gemini, OpenAI, or Anthropic)

1. Clone & Setup

cd genai-multi-agent-support
python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate
pip install -r requirements.txt

2. Configure Environment

Create a .env file in the project root to configure your LLM provider. The default provider is Google Gemini, but you can also use OpenAI or Anthropic.

# Set your preferred provider: gemini (default), openai, anthropic, or lmstudio
LLM_PROVIDER=gemini

# For Google Gemini (Default)
GEMINI_API_KEY=your-gemini-api-key-here

# For OpenAI
OPENAI_API_KEY=your-openai-api-key-here
# OPENAI_MODEL=gpt-4o

# For Anthropic
ANTHROPIC_API_KEY=your-anthropic-api-key-here
# ANTHROPIC_MODEL=claude-3-7-sonnet-20250219

Or copy from the example:

cp .env.example .env
# Edit .env and add your API key

3. Initialize Database

python setup_database.py

This creates customer_support.db with:

• 15 customers with realistic profiles
• 10 products across Software and Hardware categories
• 25+ support tickets with various statuses

4. Run the Application

streamlit run app.py

The app will open at http://localhost:8501

5. Load Sample Policies

In the Streamlit sidebar, click "📥 Load Sample Policies" to index the built-in Acme Corporation policy document. Or upload your own PDF files using the file uploader.

---

💬 Usage Examples

Structured Data Queries (SQL Agent)

• "Give me a quick overview of customer Ema's profile and past support ticket details."
• "Show me all open support tickets with high priority."
• "List all premium plan customers."
• "Which support tickets have been escalated and why?"

Policy Document Queries (RAG Agent)

• "What is the current refund policy?"
• "What is the warranty period for hardware products?"
• "What are the support tiers and response times?"
• "What is the data privacy policy for customer information?"

Combined Queries (Both Agents)

• "Does customer Ema qualify for a refund based on our policy?"
• "What is the SLA uptime guarantee for premium customers?"

---

🛠️ Technology Stack

Component	Technology
LLM	Configurable (Gemini, OpenAI, Anthropic, or LM Studio)
Embeddings	Configurable (Gemini, OpenAI, or local via LM Studio)
Framework	LangChain + LangGraph
SQL Database	SQLite
Vector Database	ChromaDB
MCP Server	FastMCP
UI	Streamlit

---

📁 Project Structure

genai-multi-agent-support/
├── README.md                  # This file
├── requirements.txt           # Python dependencies
├── .env.example               # Environment variable template
├── app.py                     # Streamlit UI entry point
├── mcp_server.py              # FastMCP server (MCP tools)
├── setup_database.py          # Database schema & synthetic data
├── create_sample_policy.py    # Sample policy document generator
├── agents/
│   ├── __init__.py
│   ├── router.py              # LangGraph router agent
│   ├── sql_agent.py           # SQL database agent
│   └── rag_agent.py           # RAG policy document agent
├── ui/
│   ├── chat.py                # Chat loop UI component
│   ├── sidebar.py             # Sidebar navigation and stats
│   └── styles.py              # CSS styling
├── tests/                     # Pytest suite
├── data/
│   ├── policies/              # Uploaded PDF documents
│   └── chroma_db/             # ChromaDB persistent storage
└── customer_support.db        # SQLite database (generated)

---

🔧 MCP Server Integration

The system exposes a Model Context Protocol (MCP) server for integration with MCP-compatible clients like Claude Desktop. This allows you to interact with the support agents as tools directly within other AI applications.

1. Available MCP Tools

Tool	Description
ask_support_assistant	Routes natural language questions to either the SQL or RAG agent automatically.
upload_policy_document	Uploads and indexes a PDF document into the knowledge base.
get_knowledge_base_stats	Returns statistics about the indexed knowledge base.

2. Connecting to Claude Desktop

To use these agents as tools directly in Claude Desktop:

1. Locate your Claude Desktop config file:

   • macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
   • Windows: %APPDATA%\Claude\claude_desktop_config.json

2. Add the server configuration: Update the mcpServers object with the absolute paths to your project. Use your full system path for command and args.

{
  "mcpServers": {
    "customer-support-agent": {
      "command": "/genai-multi-agent-support/venv/bin/python",
      "args": [
        "/genai-multi-agent-support/mcp_server.py"
      ],
      "env": {
        "LLM_PROVIDER": "gemini",
        "GEMINI_API_KEY": "your-gemini-api-key-here"
      }
    }
  }
}

3. Restart Claude Desktop: Once restarted, a 🔌 icon will appear. Claude can now call your local support assistant to query your database or policies.

3. Running Manually (for Debugging)

To run the server in stdio mode for testing or debugging:

source venv/bin/activate
python mcp_server.py

### Available MCP Tools

| Tool | Description |
|------|-------------|
| `ask_support_assistant` | Routes questions to the appropriate agent automatically |
| `upload_policy_document` | Uploads and indexes a PDF into the knowledge base |
| `get_knowledge_base_stats` | Returns knowledge base statistics |

---