MedDiscover: A Domain-Specific Retrieval-Augmented Generation Framework for Evidence-Grounded Knowledge Extraction in...

Retrieval-augmented generation (RAG) can improve biomedical question answering, but claims about domain-specific retrievers require transparent, reproducible evaluation. We present MedDiscover, an open-source RAG implementation and benchmark instantiated for metabolomics and metabolic-disorder literature, with a 2-tier dataset design: a Gold set of 10 papers and 30 expert-curated question-answer (QA) pairs with traceable references, and a Silver set of 100 papers curated using International Statistical Classification of Diseases and Related Health Problems 10th Revision metabolic disorder codes E70 to E88 with ~600 synthetic QA pairs and retrievability metadata (Ada n = 300, MedCPT n = 300). On the Gold set, retrieval augmentation improves grounding compared with a non-RAG baseline. On the Silver benchmark, MedCPT and Ada show comparable retrieval-centric performance (faithfulness $P=0.85$ , context recall $P=0.50$ , context precision $P=0.49$ , and answer relevancy $P=0.82$ ), while MedCPT yields higher answer correctness (Mann-Whitney U $P=5.72\times {10}^{-11}$ , Cliff's $\delta =0.303$ , and common-language effect $=0.651$ ). We release code, evaluation scripts, and document lists to support reproducibility.