ParsePal

Inspiration

The burgeoning popularity of Large Language Models and their wide array of applications drove us to leverage this technology in order to simplify the tedious process of manually searching through digital documents.

What it does

Using the given course catalog for the University Course Catalog Data Extraction and Query Challenge, users can ask a chatbot to search through this catalog based on queried phrases. Users can also use the chatbot to search through their own uploaded documents as well. To further tailor this experience specifically for each individual user, the user's chat history is stored and connected to the user's account for future viewing.

How we built it

We used the Next.js framework to create the frontend of the website using react and Flask to implement the backend. Users log in using PropelAuth technology in order to protect their information and so that their chat history can be stored using a Neon database and retrieved at a later date. We then used various libraries to parse information from the inputted PDFs and preprocess this data; vector embeddings were created for each sentence and stored in a Pinecone vector database. Vector embeddings were also created for user queries to the chatbot and compared with the vector database, returning the five closest matches with their respective page and sentence numbers. This information is used to retrieve the image of the corresponding page, which is then sent to the GPT-4 vision model in order to find the answer that best matches the user's query.

Challenges we ran into

We ran into some issues relating to confusion surrounding the exact parameters of the Xficient challenge. Issues with storing the images to be sent to the GPT-4 vision model were also challenging.

Accomplishments that we're proud of

Successfully parsing information from the course catalog document to be used for the chatbot was a huge accomplishment for our team. Given the setbacks that we faced with this process due to some misunderstandings of the challenge instructions, finishing this capability proved to be a monumental achievement.

What we learned

We learned how to successfully host both the frontend and backend for our web application using the Next.js framework.

What's next for ParsePal

Looking ahead, ParsePal has two key improvements. First, we'll smooth out the chatbot's user interface to ensure a comfortable user experience. This will make interacting with ParsePal more intuitive and enjoyable. Second, we'll increase storage capacity for document images processed by the GPT-4 vision model. This allows us to serve a wider range of users by accommodating larger document volumes. Additionally, we will also work to streamline the document search and querying process so that users can more quickly find the information they need.