ReMAG-KR: Retrieval and Medically Assisted Generation with Knowledge Reduction for Medical Question Answering

Large Language Models (LLMs) have significant 002 potential for facilitating intelligent end-user ap-003 plications in healthcare.However, hallucinations 004 remain an inherent problem with LLMs, mak-005 ing it crucial to address this issue with extensive 006 medical knowledge and data.In this work, we 007 propose a Retrieve-and-Medically-Augmented-008 Generation with Knowledge Reduction (ReMAG-009 KR) pipeline, employing a carefully curated 010 knowledge base using cross-encoder re-ranking 011 strategies.The pipeline is tested on medical 012 MCQ-based QA datasets as well as general QA 013 datasets.It was observed that when the knowl-014 edge base is reduced, the model's performance 015 decreases by 2-8%, while the inference time im-016 proves by 47%.