PixelRAG: Retrieval-Augmented Generation Without Text Parsing

PixelRAG proposes a fundamentally different take on retrieval-augmented generation by skipping HTML parsing and text linearization altogether. Instead, it represents web pages in their native visual form and performs both retrieval and reading directly in pixel space, preserving layout, visual structure, and formatting that text-based pipelines typically discard.

The system is described as the first to operate over a full Wikipedia corpus in this format, managing a datastore of 30 million screenshot images through an efficient visual retrieval index. Built on top of Qwen3-VL-Embedding, the model is further fine-tuned on screenshot data using carefully curated contrastive training pairs. Retrieved screenshots are then fed as raw pixel inputs to a vision-language model, with no intermediate text conversion step.

Across evaluations, PixelRAG consistently beats both no-retrieval and text-based RAG baselines, including on text-heavy benchmarks like NQ and SimpleQA where text pipelines would traditionally have an edge. It also shows strong gains on multimodal open-domain QA, noisy news-based benchmarks, and agentic tasks, improving accuracy by up to 18.1% over text-based approaches. On top of that, image compression enables up to 3x lower token costs at reduced resolutions without sacrificing accuracy — giving engineers a new efficiency lever for RAG pipelines. The results push back against the assumption that text extraction is a necessary step for web retrieval, suggesting native visual RAG can be both more accurate and more efficient.