david-andrew / pdf-chatter

LLM powered Q&A over extracted PDF text

PDF Chatter

Question Answering over PDFs using Nougat-OCR and GPT-4.

Getting Started

Prerequisites

• Python 3.9 or later
• a NVIDIA GPU with CUDA support
• environment variable OPENAI_API_KEY set to your OpenAI API key

Installation

pip install pdf-chatter

Usage

pdf-chatter path/to/pdf

which opens a REPL where you can ask questions, and GPT-4 will answer them based on the content of the PDF.

Note: pdf-chatter will save a .mmd (multi-markdown) next to the target pdf. This contains the extracted text from the PDF, and is used as a cache so the same PDF doesn't need to be re-processed every time you run pdf-chatter.

Additionally you can run the summarize command to get a summary of the PDF before entering the REPL.

pdf-summarize path/to/pdf

Example

Tips & Notes

• Nougat-OCR doesn't extract images, so any questions about images in the document will not be answered
• Nougart-OCR works best on documents similar to scientific papers, reports, etc.

How it works

1. Extract text from the PDF using Nougat-OCR
2. The entire document is fed to GPT-4 as part of its chat history via the OpenAI API
3. A simple REPL collects the user's questions and feeds them to GPT-4, which streams the answer back.