CiteSmart

CiteSmart is a lightweight, eco-friendly, and privacy-focused assistant that helps users find accurate answers and citations for queries about complex documents.

Comment

Inspiration

We were inspired by our experience as international students unfamiliar with Canadian laws. We wanted a reliable assistant that could cite its sources and handle questions out of its domain. This led to building CiteSmart, focused on answering questions about the Canadian Environmental Protection Act of 1999, combining precision with eco-friendly and privacy-preserving attributes.

What it does

CiteSmart answers questions about the Canadian Environmental Protection Act by retrieving relevant portions from a vectorized document and citing its sources. If the query is out of its scope, it uses the DuckDuckGo API to fetch relevant answers from the web, including citation links.

How we built it

In python, we split the document into chunks, embedded them using Sentence Transformers, and stored them in a vectorized database. We used the Llama 3.2 model to generate answers from relevant document chunks. A conditional router directs unanswered queries to DuckDuckGo, and a UI was created using Gradio.

Challenges we ran into

Resource limitations were a major hurdle, requiring us to optimize for efficiency on low compute. Tuning the model for accuracy and creating an intuitive UI also posed challenges.

Accomplishments that we're proud of

Creating a compact, private, and eco-friendly assistant is something we're proud of, especially one that operates effectively on limited resources.

What we learned

We deepened our knowledge of RAG, LLMs, and API integrations, particularly how to balance privacy with the need for reliable information retrieval.

What's next for CiteSmart

We aim to scale CiteSmart to handle broader datasets and make it customizable for various legal documents worldwide.

Built With

Share this project:

Updates