How AI describes VisRAG

VisRAG 2.0 (EVisRAG) is an evidence-guided Vision Retrieval-Augmented Generation framework that equips vision-language models to answer multi-image questions by first observing retrieved images to collect per-image evidence and then reasoning over those cues to respond. It is trained with Reward-Scoped GRPO, applying fine-grained token-level rewards to jointly optimize visual perception and reasoning. The VisRAG pipeline embeds documents as images using a vision-language model and retrieves them to enhance generation, aiming to preserve information in the original documents better than traditional text-based RAG.