OHPT | Devpost

Inspiration

We were inspired by the teacher assistant (TA) strike that occurred at UC Berkeley in Fall 2022. The strike brought to light a major issue in UC Berkeley's education system: departments were being insufficiently funded, leading to less available TAs to help students. This led to long office hour (OH) queues, with many students having their questions unanswered. To address this problem, we decided to work on OHPT!

What it does

OHPT is a chatbot that answers conceptual and logistical questions for courses.

How we built it

We built OHPT using a Jupyter Notebook initially as our environment, and leveraged softwares such as OpenAI (chatting), FeatureForm (orchestration), Pinecone (vector database), and HuggingFace (embedding). We also added course content (i.e. course textbook) and logistical posts on EdStem in the form of .csv files to serve as context for the chatbot to use when answering questions. This way, the answers will be personalized to the class and be much more accurate.

Challenges we ran into

Our main challenges involved balancing a relatively new software with challenging WiFi network capabilities.

Accomplishments that we're proud of

We are very proud of building a full-stack pipeline using new softwares to inject context to improve upon OpenAIs chatting capabilities.

What we learned

Primarily, we gained valuable experience in using different softwares and learning how to orchestrate them together to build a functioning and improved assistant.

What's next for OHPT

Next, we will work on making OHPT include more functions, training it to be able to help students with homework as a tutor rather than giving answers. This will involve making our system more robust to students cheating by adding functions that prevent the chatbot from directly giving answers.