UH ITS AskUs Chatbot - HACC 2023
The Problem Space
The present UH ITS Ask Us system is keyword search-based, requiring users to enter particular keywords for information retrieval, making it less user-friendly. To solve this, we want to use LLaMA2 to train a conversational AI using the ITS team’s enormous knowledge base drawn from hundreds of public web pages. This program intends to not only reduce workload but also to improve user experience and help process efficiency. To ensure smooth operations, we intend to build a reactive, intuitive, and functional user interface at the same time.
What I learned and contributed
I learned a lot about working in a collaborative environment and how to use GitHub to manage a project. Though the multiple meetings and IDPM we were able to work together to create a project that we are all proud of. In the technical aspect, I learned how to tune and tweak a model to get the best results. We experimented with many different prompts and hyper-parameters to get the best output from the model.
Once the model was trained, I had to figure out how to make the model communicate with the front end. I learned how to use FAST API to connect the front end to the model via requests which allowed us to get the model’s output and display it on the front end. This was a challenge because I had never used FAST API before, but I was able to learn it quickly and implement it into our project. This was a struggle since the FAST API documentation from LangChain was not very clear, and we had to fiddle with the request header for a while until we got it to work.
With the basis of the model completed the next step was to continually fine-tune the responses so that we could achieve an output that can satisfactorily answer the user’s questions. This was a challenge because with every single change, the model would have to be reloaded and it wasn’t always clear what changes should be done. We had to experiment with many different prompts and hyper-parameters to get the best output from the model.
Overall, I learned a lot about how to work in a team and how to manage a project and I am very proud of what we were able to accomplish in such a short amount of time.
Source Code & Demo
- Demo: https://its.monkes.net/ Down
- Github: https://github.com/bruhgroup/its-knowledge-base
This code is not currently being maintained, and thus the demo is not guaranteed to work.