blurBS

Inspiration

A lot of us are chronically online and get emotionally provoked by blatant falsehoods on the internet. In the face of such injustices, we wanted a remedy which can benefit us and others.

What it does

Our browser extension automatically identifies potential health misinformation on a webpage. It then highlights the questionable text and provides a concise, accurate correction generated from a knowledge base of credible medical sources.

How we built it

First, we manually scraped information from the Mayo Clinic to serve as a reference for the LLM. We then drafted several user interfaces, ultimately settling on a minimalistic design that we plan to improve in the future. In our implementation, the front-end sends the DOM to a back-end API endpoint, which then extracts the relevant text.

From a prototyping standpoint, we initially developed a proof of concept using unfiltered LLM responses. Subsequently, we implemented a RAG approach. This involved using Hugging Face to create embeddings from the extracted text. These embeddings were then used to rank and select the most relevant information to provide as context for the LLM, which in turn generated blurbs to correct the detected misinformation.

Challenges we ran into

Challenges we ran into included: sorting and querying the embedding vector data base, getting an LLM that would process the inputs in a timely manner, manually amassing all the sources to inform the LLM, and a lack of front end experience.

Accomplishments that we're proud of

We are proud of the minimalist and light weight user interface and the functional prototype we created in the short time frame.

What we learned

This hackathon was a valuable lesson in rapid, collaborative development. We learned to effectively integrate front-end and back-end components under pressure and gained practical experience with RAG pipelines and vector embeddings. Most importantly, we learned how to quickly merge our individual skills into a cohesive and productive team.

What's next for blurBS : Medically Accurate Browsing

One limitation we plan to address is that the extension currently only runs on the initial page load. This means it may not detect misinformation that appears dynamically after the page has loaded. To solve this, we intend to implement a feature that provides live analysis in response to page updates.

Looking ahead, we also plan to expand user controls. Furthermore, we will increase the extension's scope to cover topics like nutrition and other medical conspiracies by extending the RAG knowledge base with more relevant documents. Another key priority is to support video websites by analyzing their transcripts or captions.