Controlled Evaluation of Graph and Multimodal Augmentation in RAG for Document Question Answering
Results show that graph augmentation is most effective when it changes retrieval under evidence deficits, while multimodal evaluation requires explicit verification that answers are unavailable through text.