Text this: Chain of Evidence: Pixel-Level Visual Attribution for Iterative Retrieval-Augmented Generation