CARAIO Convo

Inspiration

We are CARAIO, a biotech startup focusing on cytogenetics. Currently, we have a computer vision pipeline that automates karyotyping from picture to karyogram. This is very useful for medical applications, but we are also looking to add an educational component as well. So, there is a CARAIO trainer that helps cytogenetics students test their skills. However, cytogenetics is a complicated field, leaving some students without answers or explanations regarding their mistakes. So, we set out to build a chatbot that will answer students' questions about cytogenetics.

What it does

Our chatbot, CARAIO Convo, is specially trained to work with cytogenetics students to guide them along their learning process. It has access to our prepared learning material in order to answer questions as factually as possible.

How we built it

We built it using a streamlit front end and a langchain backend. The main flow functions as follows: The user is prompted to select a scope of knowledge. If the user selects the local scope, the chatbot will be prompted to only use information from the documents. If global scope is selected, the chatbot will be prompted to utilize all its resources to answer the questions. Regardless of selection, the chatbot will be geared to answer questions regarding cytogenetics. Once the user asks a question, the chatbot utilizes text embeddings to find the relevant chunks of documents to base knowledge. Then, it generates text using GPT-4 and provides that to the user.

Challenges we ran into

We ran into multiple problems regarding our vectorstore during this hackathon. We had to examine the steps being taken and their order to gain a better grasp of what was occurring when it came to our bugs. There were multiple time where we took things out of order and our vectors weren't initialized correctly, but lots of debugging allowed us to overcome this.

Accomplishments that we're proud of

We are proud of being able to build a functioning app on streamlit that will help students streamline their learning. Using LangChain is not easy, but we were able to utilize it well and develop the capabilities we needed our chatbot to have.

What we learned

We learned about different frameworks and libraries such as Langchain, Openai, and Streamlit. In addition, we learned how to use services such as Pinecone to optimize storage and performance.