lessnotes | Devpost

the home page
an example of conversation, where the ingested documents include lecture notes from COMP360, algorithm design class at McGill
how the file system for uploading files works

Inspiration

When thinking about ideas for this hackathon, we wanted to work on something for which the main users would be ourselves, and by ourselves, we mean us 4, but also students in general. We wanted to work on a tool which would be more than a line on a cv, but that would be something that we would actually use on a daily basis afterwards.

What it does

Lessnotes helps students boost their efficiency when studying by minimizing the time spent researching information in their class notes. Lessnotes allows students to input documents (these documents can be lecture slides, assignments they worked on, or typed notes they took during lectures).

When they ask a question, lessnotes will go through their documents, find the most relevant ones (based on context and meaning), and provide them to an LLM chatbot (Llama3.2) which then bases their response on the input documents.

This allows our users to use the tool both as a search app to search for details hidden deep within their notes, as well as have a better version of chatgpt, tailored to their learning experience.

Tech stack

Chroma: We used Chroma as our Vector Database, allowing us to efficiently store and retrieve user documents.
mxbai: The embeddings model behind the VectorDB. This embeddings algorithm from Mixed Bread is one of the most powerfull open-source embeddings algorithms on the market.
Llama 3.2 3B: We needed a powerfull LLM in order to achieve accurate results.
LangChain: to tie it all together.

Challenges

During this hackathon, we had and still have issues with treating the different file types. As student notes can have various file types, making sure that a wide variety of them are accepted is something we worked very hard on. It is not perfect, as some documents can not be processed (ex.: EndNote libraries)

Next Steps

On top of working on supporting more files, Lessnotes would greatly benefit from the use of something like Google's Layout Parser in order to process the layout of documents more accurately and extract information such as table layouts.