RAG Project - Multi-Agent Q&A System

An intelligent RAG system using LangGraph with specialized agents for cars, countries, and math queries. Features tool-calling architecture where the cars agent dynamically chooses between SQL queries and vector search.

🚀 Key Features

• ReAct Agent with Tools: Cars agent intelligently selects between SQL database queries and vector search
• DuckDB Integration: SQL queries for counting, filtering, and aggregating car data
• Multi-Agent Workflow: Specialized routing for different query types
• Interactive UI: Chainlit web interface with expandable source citations
• Vector Database: ChromaDB for semantic search

Workflow Architecture

The system uses a classifier to route queries to the appropriate agent. The cars agent uses a ReAct pattern with two tools:

• query_cars_database: SQL queries for statistics and filtering
• search_cars_vector: Semantic search for detailed car information

🛠️ Setup

# Install uv package manager
curl -LsSf https://astral.sh/uv/install.sh | sh

# Install dependencies
uv sync

# Configure API key
cp .env.example .env
# Add your OPENAI_API_KEY to .env

# Initialize databases (ChromaDB + DuckDB)
uv run python -c "from main import initialize; initialize()"

🎯 Usage

Chainlit Web Interface

uv run chainlit run main.py

Visit http://localhost:8000 for interactive chat with source citations.

📁 Project Structure

├── main.py                    # Chainlit interface + initialization
├── src/
│   ├── agents.py              # Agent implementations with tools
│   ├── classifier.py          # Query classification
│   ├── workflow.py            # LangGraph workflow
│   ├── ingestion.py           # Database initialization
│   └── models.py              # LLM configuration
├── data/
│   ├── cars_dataset.csv       # Car data (2000 entries)
│   └── country_data.md        # Country information
├── chroma_db/                 # Vector database
└── duckdb_cars.db             # SQL database

🤖 Agents

Agent	Purpose	Examples
Cars Agent	Cars queries (ReAct with tools)	"How many Teslas?", "Tell me about BMW X5"
Countries Agent	Geographic information	"Capital of Norway?"
Math Agent	Calculations	"Square root of 144?"
General Agent	Fallback handler	General queries

Cars Agent Tools

The cars agent automatically selects the appropriate tool:

SQL Tool (query_cars_database):

• Counting: "How many cars in the database?"
• Filtering: "Cars with rating > 4.5"
• Aggregation: "Average rating by manufacturer"

Vector Search (search_cars_vector):

• Semantic queries: "Tell me about Tesla Model S"
• Features: "What are the safety features of BMW?"
• Descriptions: "Electric cars with good ratings"

🔧 Technologies

• LangChain + LangGraph: Agent orchestration
• OpenAI GPT-4o-mini: LLM
• ChromaDB: Vector database
• DuckDB: SQL database
• Chainlit: Web UI
• uv: Package management
• Langfuse: Observability and Tracing

🔍 Observability with Langfuse

This project is fully integrated with Langfuse for tracing and monitoring.

Setup

Add your Langfuse credentials to .env:

LANGFUSE_PUBLIC_KEY=pk-lf-...
LANGFUSE_SECRET_KEY=sk-lf-...
LANGFUSE_BASE_URL=https://cloud.langfuse.com

Features

• Full Traceability: Every query generates a unique trace ID
• Session Tracking: Group queries by user sessions
• Agent Monitoring: Track performance of individual agents (Cars, Math, etc.)
• Tool Usage: Monitor SQL vs Vector search tool selection
• Cost Tracking: Track token usage and costs per query

View your traces at cloud.langfuse.com.

Built with LangGraph 🦜🔗