A new retrieval-augmented generation system ditches text extraction and reads the web the way a human does — as images.
Researchers introduced PixelRAG, a pipeline that stores and retrieves web pages as screenshots rather than parsed text. The system indexes 30 million Wikipedia screenshot images, runs retrieval using a fine-tuned version of Qwen3-VL-Embedding, and feeds the resulting images directly into a vision-language model — no HTML linearization, no layout discarded in translation. On question-answering benchmarks like NQ and SimpleQA, PixelRAG beat both no-retrieval baselines and conventional text-based RAG. On multimodal and noisy-corpus benchmarks, accuracy gains reached 18.1% over text alternatives.
The result matters because it challenges an assumption baked into nearly every RAG system built in the last three years: that you have to convert web content into text before a model can reason over it. If visual retrieval holds up at production scale, it could simplify pipelines that currently depend on brittle scraping and parsing stacks. There is also an efficiency angle — image compression let the researchers cut token costs by up to 3x at lower resolutions without killing accuracy.
The catch is that 30 million screenshot images is a manageable research corpus, not the open web. Whether the approach survives contact with dynamic pages, paywalls, and the general chaos of real crawls remains an open question.