Multi-Agent Retreival Augemented Generation (MARAG) System

Overview

This project implements a Multi-Agent Retrieval-Augmented Generation (RAG) system. Unlike traditional RAG systems that use a single context retrieval and generation mechanism, this multi-agent system leverages multiple contexts and specialized agents to provide a comprehensive and accurate final answer. The system decomposes the input question into multiple contexts, retrieves relevant information for each context, and synthesizes the results from multiple agents to produce a final answer. The document used to answer questions is the following blog - https://lilianweng.github.io/posts/2023-06-23-agent/

Libraries Used

1. bs4

• Description: BeautifulSoup4 is a library for parsing HTML and XML documents.
• Usage: It is used to parse and extract content from web pages, specifically targeting the parts of the document that contain relevant information.

2. langchain

• Description: LangChain is a library for building language model applications.
• Usage: Provides the core framework and tools to define and run language model-based agents and pipelines.

3. langchain_community.document_loaders

• Description: A module within LangChain for loading documents from various sources.
• Usage: Used to load and preprocess content from web pages.

4. langchain_chroma

• Description: Chroma is a library for vectorizing and indexing documents.
• Usage: Used for creating vector embeddings of the document chunks and facilitating efficient retrieval of relevant chunks.

5. langchain_core.output_parsers

• Description: Provides utilities for parsing the output of language models.
• Usage: Used to structure and format the output generated by the language models.

6. langchain_core.runnables

• Description: Core utilities for building runnable tasks in LangChain.
• Usage: Provides the functionality to create and manage the runnable tasks that form the building blocks of the multi-agent system.

7. langchain_openai

• Description: LangChain integration for OpenAI language models.
• Usage: Facilitates the use of OpenAI’s language models for generating responses based on retrieved contexts.

8. langchain_text_splitters

• Description: A module for splitting large documents into manageable chunks.
• Usage: Used to split the documents into smaller, overlapping chunks to optimize the retrieval process.

System Functioning

Initialization

1. Loading and Parsing Documents: The system loads and parses documents from specified web paths using WebBaseLoader and BeautifulSoup.
2. Chunking and Vectorizing: The parsed documents are split into chunks using RecursiveCharacterTextSplitter and vectorized using Chroma with OpenAIEmbeddings.

Multi-Agent Mechanism

1. Context Identification: The input question is analyzed to identify five distinct contexts.
2. Agent Creation: For each context, a specialized agent is created. Each agent retrieves relevant document chunks specific to its context.
3. LLM Invocation: Each agent calls the LLM with its retrieved context to generate a response.
4. Aggregation: The responses from all agents are aggregated to form the final answer.

Advantages Over Single RAG System

1. Context Diversity: By decomposing the question into multiple contexts, the system captures a broader range of relevant information.
2. Specialized Retrieval: Each agent focuses on a specific aspect of the question, leading to more precise retrieval and response generation.
3. Comprehensive Answers: The final aggregation of multiple agent responses ensures a more thorough and nuanced answer, reducing the likelihood of missing critical information.

This multi-agent approach enhances the traditional RAG system by leveraging the strengths of multiple specialized agents, resulting in more accurate and comprehensive answers.