Render retrieved documents as images instead of text: RAGOCR beats naive RAG by 15% on one-eighth the input tokens

arXiv 2608.00765·medium signalXBlueskyLinkedInCopy link

RAGOCR compresses retrieved documents into compact visual representations conditioned on the input query, sidestepping the standard trade-off where hard compression is query-aware but weak and soft compression is strong but needs costly offline encoding. It exceeds naive RAG accuracy by over 15% while cutting input tokens to one-eighth, and consistently beats both hard and soft compression baselines across varying retrieval depths. The transferable mechanism is a query-aware dynamic resolution pass that renders highly relevant passages at higher visual granularity while aggressively downsampling peripheral documents — a relevance-proportional budget allocation any RAG pipeline can adopt even without the visual encoder.