The problem DocumentGPT solves

ChatGPT (and other similar LLMs) are cool. They can answer anything from who was the 1st person on moon to how does backpropogation train nerual networks! But what if you asked it who was the 11th person on the moon? or what are the mathematical equations of backpropogation?

Not only will it crumble when ask about the specifics, it'll also hallucinate by providing you wrong information with complete confidence!

This is where DocumentGPT comes in. We aim to provide extra information(through documents/websites/videos) to LLMs so they can have the proper context of specific information needed to answer in a given question.

You want to know who the 11th person on moon was? We'll feed it the Moon Landing wikipedia page. You wan tto understand backpropogation mathematically? We'll feed it notes from graduate Machine Learning Students!

We've added support to accept multiple types of note taking documents(docs, pdfs, markdowns, etc), online websites, online videos, handwritten notes and basically anythingi that can feed the correct need information to answer the question at hand.

Students can use this in multiple ways like submit multiple pdfs and ask question to get answers based on informationi contained on all of them, quickly summarize hours long youtube videos, or highlight a text from complex research papers to get beginner friendly explanation.

Professionals can use this to stay up to date with cutting edge research happening in a specific field, summarize long documents, etc.

Challenges we ran into

LLMs are new and a rapidly evolving technology. To start with, there weren't enough resources for us to understand the tech behind it better. Some of the libraries we found which were related to our project were in heavy development phase and had breaking changes within span of hours. All the aspects of the library were not properly documented either so we had to treat the libraries source code as our documentation.

Allowing users to upload multiple files, files with handwritten notes and writing generalised functions to handle each case was an interesting challenge.

Making sure we provide responses which resonate with a variety of audience required some great testing.

Completing all the features we had planned in a 24 hour span was not an easy task and required great team cooperation.