For evaluation metrics, retrieval-layer metrics (such as Recall@k and MRR, based on canonical IDs) can be reported separately from generation-layer metrics (fluency and language naturalness).