Systematic Study of Retrieval Pipeline Design for Retrieval-Augmented Medical Question Answering
arXiv·medium signal
Provides a rigorous comparison of RAG pipeline architectures for medical QA, testing combinations of retrievers, rerankers, chunk sizes, and prompt strategies. Finds that pipeline design choices (not just the LLM) account for significant variance in answer quality — and that commonly used defaults are suboptimal for medical domains. Practical design guidance for anyone building RAG systems for specialized knowledge-intensive applications.