med.ai

Inspiration

We got the inspiration from the challenge of KI group, which wasn't full, but we fully defined in our own scope.

What it does

It is a multi-agent chatbot, built on top of chatGPT, who takes some question about healthcare or some medicine etc. while possibly being able to upload its medical history and get close-to-expert answer.

How we built it

From the software perspective we build Front-end of a chatbot and Back-end with several endpoints. From AI perspective we built a multi-agent system with 3 models and altogether 8 pre-trained or pre-prompted chatGPT models. It consists of

Donatello: main creative model, which has three to five agents inside of it: * First is looking for the information in the RAG database and if the haven't found anything than the second one is doing general reasoning and is the main source of LLM hallucination inside of Donatello * Then from one to three others they do some expertise check and reduce hallucination.
Leonardo: model which hallucination check according to recently developed FactSelfCheck, which is ran on knowledge graphs and sampling, with feedback loop to Donatello, which is prompted to regenerate in case it hallucinates
## Challenges we ran into
hallucinations of the model
knowledge graph creation and kg-fusion
multi-agent communication orchestration
optimization of the execution time of each part of the model
uploading and processing pdf by the openai model
showing status message on the front-end side ## Accomplishments that we're proud of
we trained very complex multi-agent system with knowledge graphs and fact checking
we created decent front-end - back-end infrastructure ## What we learned
how to build multi-agent systems
how to extensively interact with openai API