●builderYou should optimize your RAG pipeline for paragraph-level citations rather than sentence-level to maintain higher model accuracy.
●researcherThis suggests that current evaluation metrics may penalize models for losing context during fine-grained grounding tasks.