Large Language Models (LLMs) offer transformative potential in healthcare, yet
Medical Question Answering (QA) remains challenging due to strict requirements
for specialized knowledge and safety. While Retrieval-Augmented Generation (RAG)
mitigates ha...
Large Language Models (LLMs) offer transformative potential in healthcare, yet
Medical Question Answering (QA) remains challenging due to strict requirements
for specialized knowledge and safety. While Retrieval-Augmented Generation (RAG)
mitigates hallucinations, standard approaches struggle with retrieval noise and information
loss within long contexts. Existing solutions relying on fine-tuning or specialized
retrievers face high barriers in clinical settings due to data privacy regulations
and infrastructure costs.
To address these limitations, we propose a Medical RAG framework utilizing
Post-Retrieval Context Engineering, designed to maximize performance without
modifying model parameters. By keeping the retriever and LLM fixed, this approach
focuses on the selection, compression, and reordering of the post-retrieval context.
This study aims to secure system accuracy and robustness solely through context
optimization, avoiding the need for high-cost training.
The proposed framework integrates recipes tailored to the medical domain, including
filtering based on medical entities and PICO structures, deduplication strategies,
and reordering algorithms using cues like definitions and MeSH terms. Additionally,
adaptive strategies such as evidence sketching and dynamic search expansion
are employed to enhance logical completeness.
Experiments on the BioASQ Task 13b benchmark demonstrate significant performance
improvements across all question types compared to a baseline RAG. Notably,
the method reduced the performance gap with an Oracle setting by over half
for Factoid questions without model changes. Furthermore, it achieved higher accuracy
while reducing input tokens by approximately 50%, proving cost-efficiency and
effectively suppressing hallucinations through transparent citation generation.
In conclusion, this study proves that substantial improvements in medical RAG
are achievable through sophisticated context engineering alone, challenging the need
for high-cost training. The proposed framework offers a practical solution for medical
institutions constrained by data security and cost, contributing to reliable medical
AI design.