Graph-Augmented Intelligence

Year

2025

Tech & Technique

KG-RAG (Baranzini Lab, UCSF), GPT-4, Llama, BiomixQA (evaluation benchmark)

Description

Knowledge-graph retrieval-augmented generation (KG-RAG) for biomedical question answering. Built on the open-source KG-RAG framework from the Baranzini Lab, UCSF.

What I did:

Evaluated GPT-4 and Llama on the BiomixQA benchmark using the KG-RAG framework.

Credit: the KG-RAG framework was created by the Baranzini Lab, UCSF. BiomixQA is an existing benchmark, not my dataset.

My Role

Project — applied and evaluated an existing framework

Evaluated GPT-4 and Llama on BiomixQA with the KG-RAG framework (Baranzini Lab, UCSF).

KG-RAG Schematic

Schematic of the KG-RAG framework (credit: Baranzini Lab, UCSF): prompt → disease-entity recognition → graph retrieval → prompt-aware context → LLM.

Example Question Source

FDA news page on setmelanotide for Bardet-Biedl syndrome — the example used in the KG-RAG project README (Baranzini Lab, UCSF).