startrail-org/pixelrag

Most search systems work by extracting text from a document and indexing that text. Tables, charts, diagrams, and anything that depends on visual layout are often reduced to fragments or lost entirely during that extraction step. PixelRAG takes a different approach: it converts documents into screenshot tiles and indexes the images directly, so visual structure stays intact for the AI model that reads the retrieved results. The project ships with a pre-built index of 8.28 million Wikipedia pages that you can query immediately through a hosted API or by running the search server locally. It also includes a command-line tool called pixelshot that renders any web page or PDF into screenshot tiles, and a pipeline for building a searchable index from your own document collections. The retrieval model is a fine-tuned version of Qwen3-VL-Embedding, trained specifically on screenshot data so that images with similar visual content are placed close together in the embedding space. The tool also ships as a plugin for Claude Code called pixelbrowse. When the plugin is active, Claude can screenshot a web page and read the image rather than fetching raw HTML, which means it can interpret charts, tables, and formatted layouts the way a person would. The pipeline is split into stages: rendering, chunking, embedding, index building, and serving. You install only the stages you need. A separate training environment is included for fine-tuning the embedding model on your own data. The pre-trained adapter weights and the training dataset are both available on Hugging Face. The project comes from Berkeley AI research groups and is released under Apache-2.0.