by Justin Lacoste, Thomas Lemoine, and Henri Lemoine

The inspiration for the project was our desire to make studying and learning more efficient and accessible for students and educators. Utilizing advancements in technology, like the increased availability and lower cost of text embeddings, to make the process of finding answers within educational materials more seamless and convenient.

Wise Up is a website that takes many different types of file format, as well as plain text, and separates the information into "pages". Using text embeddings, it can then quickly search through all the pages in a text and figure out which ones are most likely to contain the answer to a question that the user sends. It can also recursively summarize the file at different levels of compression.

With blood, sweat and tears! We used many tools offered to us throughout the challenge to simplify our life. We used Javascript, HTML and CSS for the website, and used it to communicate to a Flask backend that can run our python scripts involving API calls and such. We have API calls to openAI text embeddings, to cohere's xlarge model, to GPT-3's API, OpenAI's Whisper Speech-to-Text model, and several modules for getting an mp4 from a youtube link, a text from a pdf, and so on.

We had problems getting the backend on Flask to run on a Ubuntu server, and later had to instead run it on a Windows machine. Moreover, getting the backend to communicate effectively with the frontend in real time was a real challenge. Extracting text and page data from files and links ended up taking more time than expected, and finally, since the latency of sending information back and forth from the front end to the backend would lead to a worse user experience, we attempted to implement some features of our semantic search algorithm in the frontend, which led to a lot of difficulties in transferring code from Python to Javascript.

Since OpenAI's text embeddings are very good and very new, and we use GPT-3.5 based on extracted information to formulate the answer, we believe we likely equal the state of the art in the task of quickly analyzing text and answer complex questions on it, and the ease of use for many different file formats makes us proud that this project and website can be useful for so many people so often. To understand a textbook and answer questions about its content, or to find specific information without knowing any relevant keywords, this product is simply incredibly good, and costs pennies to run. Moreover, we have added an identification system (users signing up with a username and password) to ensure that a specific account is capped at a certain usage of the API, which is at our own costs (pennies, but we wish to avoid it becoming many dollars without our awareness of it.

As time goes on, not only do LLMs get better, but new methods are developed to use them more efficiently and for greater results. Web development is quite unintuitive for beginners, especially when different programming languages need to interact. One tool that has saved us a few different times is using json for data transfer, and aws services to store Mbs of data very cheaply. Another thing we learned is that, unfortunately, as time goes on, LLMs get bigger and so sometimes much, much slower; api calls to GPT3 and to Whisper are often slow, taking minutes for 1000+ page textbooks.

What's next for Wise Up is to make our product faster and more user-friendly. A feature we could add is to summarize text with a fine-tuned model rather than zero-shot learning with GPT-3. Additionally, a next step is to explore partnerships with educational institutions and companies to bring Wise Up to a wider audience and help even more students and educators in their learning journey, or attempt for the website to go viral on social media by advertising its usefulness. Moreover, adding a financial component to the account system could let our users cover the low costs of the APIs, aws and CPU running Whisper.

amazon-web-services css flask gpt-3.5 html javascript openai python ubuntu whisper

wise-up.tech