Agentic RAG System Using CrewAI & LangChain

This project implements an advanced Retrieval-Augmented Generation (RAG) system using multiple AI agents to improve information retrieval and response quality. The system uses CrewAI for agent orchestration and LangChain for RAG functionality.

Features

• Multi-agent architecture for sophisticated RAG operations
• Automated routing between vector store and web search
• Multiple validation layers to ensure response quality
• Hallucination detection and prevention
• PDF and web search capabilities

Prerequisites

pip install crewai==0.28.8 crewai_tools==0.1.6 langchain_community==0.0.29 sentence-transformers langchain-groq

You'll also need the following API keys:

• Groq API key
• Tavily API key

Agent Architecture

The system implements five specialized agents:

1. Router Agent

• Role: Routes questions to appropriate search methods
• Function: Determines whether to use vector store or web search
• Goal: Optimize information retrieval path

2. Retriever Agent

• Role: Executes information retrieval
• Function: Uses either vector store or web search based on router's decision
• Goal: Extract relevant information for the given question

3. Grader Agent

• Role: Validates retrieval relevance
• Function: Evaluates if retrieved content matches the question
• Goal: Filter out irrelevant retrievals

4. Hallucination Grader Agent

• Role: Prevents hallucinations
• Function: Verifies if answers are factually grounded
• Goal: Ensure response accuracy

5. Answer Grader Agent

• Role: Final quality control
• Function: Reviews and refines answers
• Goal: Ensure responses are useful and accurate

Task Pipeline

1. Routing Task: Analyzes question keywords and determines search method
2. Retrieval Task: Executes search and extracts information
3. Grading Task: Evaluates retrieval relevance
4. Hallucination Check: Verifies factual accuracy
5. Answer Quality: Ensures response usefulness

Usage

# Initialize the crew with agents and tasks
rag_crew = Crew(
    agents=[Router_Agent, Retriever_Agent, Grader_agent, 
            hallucination_grader, answer_grader],
    tasks=[router_task, retriever_task, grader_task, 
           hallucination_task, answer_task],
    verbose=True,
)

# Execute the pipeline
inputs = {"question": "Your question here"}
result = rag_crew.kickoff(inputs=inputs)

Agent Flow

Question Input
    │
    ▼
Router Agent
    │
    ├─► Vector Store Search
    │         or
    ├─► Web Search
    │
    ▼
Retriever Agent
    │
    ▼
Grader Agent
    │
    ▼
Hallucination Grader
    │
    ▼
Answer Grader
    │
    ▼
Final Response

Key Benefits

1. Improved Accuracy: Multiple validation layers reduce errors
2. Flexible Retrieval: Smart routing between different information sources
3. Quality Control: Systematic checks for relevance and factuality
4. Hallucination Prevention: Dedicated agent for fact verification
5. Adaptive Responses: Falls back to web search if vector store results are inadequate

Notes

• The system uses Groq's LLM for agent operations
• PDF search capability is implemented using LangChain
• Web search is handled through the Tavily API
• All agents operate with verbose logging for debugging

Customization

You can modify agent behaviors by adjusting:

• Agent backstories
• Task descriptions
• Expected outputs
• Tool configurations
• LLM parameters

Limitations

• Requires multiple API keys
• Sequential processing may impact response time
• Dependent on external service availability
• Memory usage scales with PDF size