Chat with any PDF using GPT3.5-turbo, faiss vector database and InstructorEmbedding. Uses around 3GB of GPU VRAM for the InstructorEmbedding model.

Tested on Ubuntu 22.04

• Create a new conda env conda create -n chatpdf python=3.9 -y
• Activate the conda env conda activate chatpdf
• Install required packages pip install -r requirements.txt
• Install faiss vector database conda install -c conda-forge faiss-gpu
• Create a OpenAI account and get a OPENAI_API_KEY

python chat.py

Q) Why not use OpenAI Embeddings A) Please feel free to create a pull request to add support for it. I wanted to keep the project entirely open source but ultimately couldn't resist using GPT3.5-turbo

• Add support for other embeddings, chat models
• Improve the Prompts
• Show streaming response from GPT3.5-turbo

This code is based on two similar projects

• https://github.com/mayooear/gpt4-pdf-chatbot-langchain
• https://github.com/amrrs/LLM-QA-Bot