PIXELRAG: Web Screenshots Beat Text for Retrieval-Augmented Generation

A new retrieval method called PixelRAG processes web pages as screenshots rather than parsed text, outperforming conventional text-based retrieval-augmented generation systems across multiple benchmarks, according to a paper posted to arXiv on June 1, 2026 [1][2]. The system, described by its authors as the first pipeline to operate over a full Wikipedia corpus in its native visual form, builds a datastore of 30 million screenshot images and uses an efficient visual retrieval index [1][2]. PixelRAG is built on the Qwen3-VL-Embedding model, which the researchers fine-tuned on screenshot data with contrastive training [1][2]. Retrieved screenshots are fed directly as pixel inputs to a vision-language model, bypassing the HTML parsing and text-extraction steps that dominate current web retrieval pipelines [1][2]. The paper reports that PixelRAG consistently outperforms both no-retrieval and text-based baselines, including on widely studied text-centric tasks such as Natural Questions and SimpleQA [1][2]. On multimodal open-domain question-answering benchmarks like MMSearch, noisy news corpora such as LiveVQA, and agentic benchmarks like MoNaCo, accuracy improved by up to 18.1% over text-based baselines [1][2]. The authors also demonstrate that image compression can reduce token costs by up to 3x at lower resolutions while maintaining accuracy [1][2]. The preprint was posted to arXiv, the open-access repository that hosts e-prints across physics, computer science, and other fields [6]. As of November 2024, arXiv receives roughly 24,000 submissions per month and has surpassed two million total articles [6]. Papers on the platform are moderated but not peer-reviewed, a distinction that has drawn attention in cases such as the 2023 LK-99 room-temperature superconductor claims, which were later found to be unsupported after replication attempts [6][8]. PixelRAG’s approach challenges the assumption that web retrieval requires text representations, suggesting that operating directly in the web’s visual form can improve both performance and efficiency [1][2]. The paper is available on arXiv under the computer science information retrieval category [1].