Agentic-RAG-with-Search

This repository contains an implementation of an adaptive Retrieval-Augmented Generation (RAG) system using LangChain and Cohere APIs. The system intelligently routes user questions to either a vector store containing pre-indexed documents or performs a web search based on the nature of the question. It leverages a series of prompts and models to ensure the generation is accurate and relevant.

Features

Environment Setup:

• Set user agent and API keys for Cohere and Tavily services.

Indexing:

• Load documents from specified URLs.
• Split the documents into manageable chunks.
• Store the chunks in a Chroma vector store with Cohere embeddings.

Question Routing:

• Route user questions to either the vector store or web search or None (model will directly answer)
• Use few-shot prompting to train the model on routing decisions.

Retrieval and Generation:

• Retrieve relevant documents from the vector store.
• Generate answers using the retrieved documents or fallback to the LLM if necessary.
• Grade the relevance of retrieved documents and the generated answers.

Grading and Quality Control:

• Grade the relevance of documents and generated answers.
• Ensure the generated answers are grounded in the retrieved documents to prevent hallucinations.
• Re-write and optimize questions for better retrieval if initial results are unsatisfactory.

Graph Workflow:

• Define a workflow using a state graph to manage the entire process, from question routing to answer generation and grading.

Usage

The system can be run with different input questions to demonstrate its capabilities. The workflow adapts based on the input question, performing retrieval, generation, grading, and re-writing as needed.

inputs = {
    "question": "What are the types of different graph neural nets?"
}
for output in app.stream(inputs):
    for key, value in output.items():
        print(f"Node '{key}':")
    print("\n---\n")
    time.sleep(6)
    print("\n---\n")

print(value["generation"])