PixelRAG beats text parsers on accuracy and cuts AI agent token costs 10x

Researchers introduced PixelRAG, which replaces HTML-to-text parsing in RAG by indexing rendered webpage screenshots and feeding retrieved image tiles to a vision-language model. Across six benchmarks on Wikipedia-scale data, it beat text RAG by up to 18.1% and cut agent prompt tokens about 10x. Visual chunking remains unresolved, and hybrid use is advised.