ResearchGPT

This is a fork of the original ResearchGPT. The current version modifies several settings from the original:

• Use GPT-3.5-turbo instead of GPT-3.
• Add Chinese language support.
• Improved results for journal papers (See compare results below).
• Modification of the frontend.

I use this repository to study OpenAI's API, Microsoft's Azure, GitHub Pages, and CI/CD. The partial original README is below.

This is a flask app provides an interface to enable a conversation with a research paper. You can enter a link to a pdf hosted online or upload your own pdf. The app will then extract the text from the pdf, create embeddings from the text and use them with the openai api to generate a response to a question you ask. It will also return a source for the part of the text it used to generate the response and the page number.

Example

This web app supports query in multiple languange. Here is an example of a query in both English and Chinese:

This app also supports pdf files in the Chinese language. Here is an example of a Chinese pdf file:

Installation

git clone https://github.com/MrPeterJin/researchgpt
cd researchgpt
pip install -r requirements.txt

Also, you need to have an OpenAI API key and set it as the environment variable 'OPENAI_API_KEY'.

Usage

Run locally

The local version would save the pdf embeddings in the embedding folder to save OpenAI API usage. You can run the app locally with:

export OPENAI_API_KEY=YOUR_API_KEY
python local.py

And then open http://127.0.0.1:8080/ in your browser.

Microsoft Azure Deployment

The online version does not save any data. Follow the instructions here. Once you have the azure cli set up with az login, you can deploy with streamed logs:

az webapp up --runtime PYTHON:3.9 --sku B1 --logs

Limitations

Due to the PDF to text conversion and embedding construction technique, the web app is limited to handle detailed query. Also, when a paper has distinguished pattern from the ordinary paper, this application also may not able to handle it. Comparing to the original version, this version has specifically targeted at this problem and making improvements. We are continuing working on improving the app to give better respond. At this time, you are encouraged to try this app on papers less than 20 pages and give us feedback.

The improved results can be seen as follows (Click for full-sized image):

This fork	Original