Ai-hackathon for Uchicago DSI
We built Genesis.AI, a retrieval-augmented chatbot that securely accesses patient records and answers general medical questions. To enable a seamless, end-to-end medical Q&A chatbot using entirely open-source tools. Our stack includes:
- LangChain for orchestration
- PostgreSQL (with vector embeddings) as both relational and vector databases
- Streamlit for the front-end
- Google Cloud Vertex AI to host and fine-tune Llama 3-8B
Key innovations:
- Dynamic SQL Generation
- Fine-tuned Llama 3-8B to translate natural-language medical queries into SQL on the fly.
- Executes those SQL queries against patient tables and metadata.
- Vector Retrieval
- Indexes medical reports in PostgreSQL’s vector store.
- Uses the same SQL-driven pipeline to retrieve relevant unstructured data.
- Accurate Response Generation
- Further fine-tuned Llama 3-8B to synthesize structured and vector-retrieved information into precise, context-aware answers.
Demo:
