Mayo Clinic’s secret weapon against AI hallucinations: Reverse RAG in action
1 min read
Summary
Mayo Clinic in the US is using a backwards retrieval-augmented generation (RAG) model to improve the accuracy of its large language models (LLMs) when providing medical advice.
RAG pulls information from specific, relevant data sources, but the process has limitations, including a high incidence of hallucinations.
The clinic pairs the RAG process with the CURE clustering algorithm and vector databases to check the accuracy of data retrieval twice.
Each fact extracted from a summary generated by the LLM is matched back to the source document, scored based on how well it aligns, and then verified.
Mayo is using the AI technology to improve the administrative burden on doctors when reviewing complex and disparate patient records and has received “incredible interest” from doctors across the clinic.
The technology can also be used to support more advanced use cases such as improved image encoding for X-rays and genomic modelling to determine the best course of arthritis treatment.