Local Corrective RAG Agent

Project Description

Jupyter notebook tutorial for a local Corrective RAG Agent. The goal is to provide accurate, relevant, and hallucination-free answers by combining the strengths of retrieval-based systems with generative AI and web search capabilities.

Tech Stack

• Llama3
• Firecrawl 🔥
• Langchain 🦜️🔗

Key Features

• Document retrieval and relevance grading
• Web search integration for enhanced information gathering
• Hallucination detection and correction mechanism
• Iterative answer improvement process

Workflow

1. Question Input: User-provides question.

2. Retrieve Documents: Relevant documents retrieved based on the input question.

3. Grade Documents: Retrieved documents are graded for relevance and quality.

4. Relevance Check: The system checks if any retrieved document is relevant to the question.

   • If Yes: Proceeds to generate an answer.
   • If No: Initiates a web search for additional information.

5. Generate Answer: Using the relevant documents or web search results, the system generates an answer.

6. Hallucination Check: The generated answer is checked for potential hallucinations or inaccuracies.

   • If No hallucinations detected: The answer is provided to the user.
   • If Yes (hallucinations detected): The system asks if the answer addresses the question.
     • If Yes: The final answer is provided.
     • If No: The system initiates a web search for more accurate information.

7. Web Search: When needed, the system performs a web search to gather additional, up-to-date information to supplement or correct the answer.

Setup and Usage

Installation

To set up the Corrective RAG Agent, you need to:

1. Install Ollama using the official download link or using the homebrew command:

brew install ollama

2. Install the required packages. Run the following command:

pip install -U langchain-nomic langchain_community tiktoken langchainhub chromadb langchain langgraph tavily-python gpt4all firecrawl-py