Agentic-Corrective-RAG-System

This project will cover a full hands-on workflow and demonstration of how to build an Agentic Corrective RAG (CRAG) System with LangGraph, it's adapted from Analytics Vidhya free courses.

Overview

The Agentic CRAG system enhances traditional RAG by incorporating checks for document relevance and performing web searches when retrieved documents are insufficient. The system is structured as a graph where each node represents a specific functionality, orchestrated using LangGraph.

Key steps include:

• Retrieval from Vector Database: Fetch relevant documents based on the user query.

• Document Relevance Check: Evaluate whether retrieved documents are relevant to the query.

• Query Rewriting and Web Search: If documents are not relevant, rewrite the query and perform a web search to gather additional context.

• Response Generation: Generate accurate answers using the retrieved or augmented context.

The system is designed to mitigate issues like poor retrieval and out-of-context or hallucinated answers by dynamically adjusting its workflow.

Features

• Agentic Workflow: Utilizes LangGraph to create a flexible, node-based workflow.

• Relevance Checking: Assesses the relevance of retrieved documents to ensure high-quality responses.

• Web Search Integration: Augments context with web searches when vector database results are inadequate.

• LLM Integration: Leverages local Ollama LLM models (gemma3n:e4b and llama3.2) and the nomic-embed-text embedding model for query processing and response generation.

Installation

To set up the project, follow these steps:

1- Clone the Repository:

git clone https://github.com/bassemalyyy/agentic-crag-system.git

2- Install Dependencies: Ensure you have Python 3.10 or higher installed. Then, install the required packages:

pip install langchain langchain-ollama langchain-community langgraph

The corrective-rag-system.ipynb notebook includes the full dependency installation commands, which are already satisfied in the provided environment:

• langchain
• langchain-groq
• langchain-community
• langgraph

3- Set Up Environment:

• Install and configure Ollama to run the local LLM models (gemma3n:e4b and llama3.2) and the embedding model (nomic-embed-text).

ollama pull gemma3n:e4b
ollama pull llama3.2
ollama pull nomic-embed-text

• Ensure Ollama is running locally:

ollama run

• Ensure you have access to a vector database and configure it as needed.

• Get you free Tavily API key, in the third cell of the code enter your API Key:

from getpass import getpass

TAVILY_API_KEY = getpass('Enter Tavily Search API Key: ')

• Run the Notebook: Open the corrective-rag-system.ipynb notebook in Jupyter or a compatible environment (e.g., VS Code, Colab) and execute the code.

Usage

The notebook demonstrates the Agentic CRAG system with example queries. To test the system:

1- Open the Notebook: Load corrective-rag-system.ipynb in a Jupyter environment.

2- Run the Cells:

• Execute the dependency installation cell to ensure all packages are available.

• Run the cell that displays the LangGraph workflow diagram to visualize the system.

• Test the system with provided queries like:

  • what is an agent?
  • what is CrewAI agent?
  • what is orchestration?

3- Example Queries: The notebook includes example queries and their responses:

• Query: "what is an agent?"

  • Response: Describes an agent as a system that independently accomplishes tasks using an LLM, with capabilities for workflow management and tool access. -> THIS RESPONSE IS RETRIEVED FROM THE DOCUMENTS AS THE QUERY IS RELEVANT TO THE DOCUMENTS

• Query: "what is CrewAI agent?"

  • Response: Explains a CrewAI agent as an autonomous unit within the CrewAI framework, capable of task execution, decision-making, and collaboration. -> THIS RESPONSE IS RETRIEVED FROM THE WEB SEARCH RESULTS AS THE QUERY IS NOT RELEVANT TO THE DOCUMENTS

• Query: "what is orchestration?"

  • Response: Defines orchestration as managing workflow execution in single or multi-agent systems, including patterns like the manager pattern. -> THIS RESPONSE IS RETRIEVED FROM THE DOCUMENTS AS THE QUERY IS RELEVANT TO THE DOCUMENTS

4- Custom Queries: Modify the query variable in the notebook to test custom questions. For example:

query = "your custom question here"
response = agentic_rag.invoke({"question": query})
display(Markdown(response['generation']))

Project Structure

• corrective-rag-system.ipynb: The main Jupyter notebook containing the implementation and test cases.

• README.md: This file, providing an overview and instructions for the project.

Workflow Diagram

The system’s workflow is visualized using a LangGraph-generated diagram, as shown below:

This diagram illustrates the nodes and decision points in the CRAG system, including retrieval, relevance checking, query rewriting, web search, and response generation.

Notes

• Environment: The notebook was tested in a Python 3.10 environment with the specified dependencies.

• Local Models: The system uses local Ollama models (gemma3:e4b, llama3.2, and nomic-embed-text).

• Vector Database: The system assumes access to a vector database for document retrieval. Configure your database connection as needed.

• Extensibility: The system can be extended by adding new nodes to the LangGraph workflow or integrating additional tools for web search or document processing.