Insurance Assistant

sample of conversation

Inspiration

In the beginning of the hackathon Luis mentioned that he injured his finger, and he finds it tedious to research how he can help find relief while trying to stay under the constraints of his insurance. He sent me 155-page document that would have taken several hours if not days to read. Nonetheless, attempt to find the answer to specific questions.

What it does

The User is able to upload their insurance policy. Then the user loads the pdf into the langchain_core.vectorstores.VectorStore where once the data is embedded the user is able to begin chatting.

How we built it

This code was built in a Huggingface space. It's uses the Cohere embedding model, and it uses the gpt-3.5-turbo-16k large language model as it handles the large context window in addition to maximizes the quality of the response.

Challenges we ran into

We ran into challenges configuring the hf-space / GitHub repository such that any contributor who pushes to main will also update the hf-space. This is managed, by using GitHub actions configured to the hf-space. Another issue was updating the hf-space from the original code as that space was made months ago.

Accomplishments that we're proud of

We are proud of completing the MVP and having it public with over 12 hours to spare to receive feedback.

What we learned

We learned the of code in terms of less is more. The more code, the harder it is for someone else to make additions. It's also important to take your time when writing code, especially for other people. It's important to document code maintain readability. We had someone join the team when the MVP, was 80% done, but they quit not knowing where the make contributions.

What's next for Insurance Assistant

Being able to input different policies to determine the pros and cons of picking one over the other. Gaining insight on what user's care most about with regard to insurance and health.