Letter to the Editor: Toward Retrieval-Grounded Evaluation for Conversational LLM-Based Risk Assessment (Preprint)

UNSTRUCTURED Abstract This letter provides a methodological commentary on a recently published study describing a conversational large language model–based system for pediatric COVID-19 risk assessment. We discuss how evaluation based solely on large language model–only pipelines and aggregate discrimination metrics may overestimate reliability in conversational clinical applications when factual verifiability is not explicitly assessed. Drawing on recent empirical evidence from retrieval-augmented generation in medical tasks, we highlight the importance of evidence grounding for accuracy interpretation, safety assessment, and subgroup-level auditing. We suggest that retrieval-grounded sensitivity analyses may strengthen the evaluation of conversational AI systems intended for clinical or public-facing use.