Doc Multi-Agents Project

This project implements a Retrieval-Augmented Generation (RAG) agent using LlamaIndex and Meta's Llama 3.2 models.

Workflow Diagram

Goals:

• Train and index key aspects of your documentation as specialized bot agents (e.g.: one for conceptual content, another for tutorials and how-tos, and others for terms and glossaries, etc).
• When a user asks a question, the system analyzes the prompt to identify the type of query.
• The right agent (or multiple agents, if needed) is selected to answer specific parts of the prompt.
• The responses from each agent are sent to the LLM, which creates a cohesive, final response.

Setup

Windows OS:

1. Install Python 3.12.
2. Run pip install -r requirements.txt to install dependencies.
3. Create a new virtual environemnt: python -m venv multiagentsenv.
4. Activate the virtual environment: multiagentsenv/Scripts/activate.
5. Log into HuggingFace huggingface-cli login --token $env:HUGGINGFACE_TOKEN
6. Run the Flask agent: python flask_app.py...follow instructions in terminal to find the web endpoint for the Agent's Web UI. Start chatting with the Doc Bot Agent from there.

Features

• Web content ingestion using BeautifulSoup.
• Multiple RAG agents implementation using LlamaIndex.
• LLM using Meta Llama3.2 hosted through HuggingFace Hub
• Web UI (Chat Bot) powered by Flask

Remarks

• To help speed things up during execution, run process_data.py by itself first. This file takes a list of content URLs and scrape the content. The raw text will be placed under the folder data/raw
• Right now, load_data.py will try to load the vectorized indexes from the data/processed folder if they exists. Otherwise, it will read in the raw content files and index the content. This process can take awhile if you have a large set of files to process. So, you may also want to separate out the index processing into it's own file (similar to process_data.py) so that you can run that part of the process by itself. Once the raw files are indexed, reading them into the agent is relatively fast.
• For your convenient, I uploaded the raw and processed files to this project to help save you time. But, if you make any modifications to the raw content, you will need to re-run the indexing.
• LlamaIndex's ReAct Agent does what i wanted so the individual agents/ are no longer needed. In a ReAct Agent, they are turned into tools. But, I am keeping the agents/ for referencing. They are not used in the current setup.