Rag using ReAct from scratch

This project implements a RAG model for PDF files using ReAct. We want to recreate an agent from scratch, without using Langchain.

Authors

• Théo Ripoll
• Nicolas Fidel
• Adrien Giget
• Quentin Fisch

Current state of the project

In a first time, we implemented a RAG agent using ReAct with Langchain, to see the potential of the ReAct agent. We obtain very positive results, which you can see in the notebook langchain_react_rag.ipynb (more explanations on how to use it in the next section).

The main ReAct framework tool is locate in the src folder, where you will find the foundations of the tool. You will find the final report in the pdf folder, which details the overall project.

Tasks

• Use libraries (Langchain) to see potential results (Quentin)
• “Play” with different LLMs and try to see which seems to be best to fit to the task (Adrien, Quentin)
  • Open source ? GPT-4 through OpenAI’s API ? …
  • What is the best prompting method ? What can we do to know which action to take ?
  • …
• Implement a PDF loader to get a document and extract information (Théo)
  • Load file
  • Cut text in chunks
  • Compute embeddings
• Choose and setup vector database (Théo, Nicolas)
  • ChromaDB ? Pinecone ? MongoDB Atlas Vector Search ? …
• Create the main tool with the connection between the LLM and the database using functions from the ReAct framework (Everyone)
• Build a front to have a more valuable product using Gradio or Streamlit (not defined yet as this is not the most important part)

The last part as not been done due to time constraints. We decided to focus on the main tool to try to improve the results, which inevitably did not help much.

Usage of the main tool

Use poetry to install the dependencies.

poetry install

Download Ollama to run Llama2 locally (https://ollama.com/download)

ollama run llama2

Run main.py and follow the input steps

python main.py

Usage of the notebook

It is a test of the Langchain version.

Use poetry to install the dependencies.

poetry install

Then create a .env file with the following variables:

OPENAI_API_KEY=YOUR_API_KEY

You should then be able to run the langchain_react_rag.ipynb notebook.