PixelRAG swaps HTML parsing for full-page screenshots, then builds a visual search index over those images. You install it with pip and run two main commands. First, “pixelshot” captures any URL or PDF as tiled images. Then the “pixelrag” pipeline—chunk, embed, build-index—turns those tiles into FAISS or Qdrant vectors. The team ships a pre-built index of 8.28 million Wikipedia articles, accessible via a live API (no key required) or locally by downloading ~217 GB from Hugging Face. Under the hood, they fine-tuned a Qwen3-VL-Embedding model to convert screenshots into vectors that preserve tables, charts and layout. A Claude Code plugin, “pixelbrowse,” hooks into Claude so it literally screenshots pages and reads diagrams and tables the way a person would. You can ask Claude: “screenshot https://news.ycombinator.com and summarize the top stories,” and it returns visuals and text together. The tool works on Linux (CUDA) or macOS (MPS) without a GPU. You create a YAML config pointing to local docs or URLs, set model Qwen/Qwen3-VL-Embedding-2B, and run `pixelrag index build` followed by `pixelrag serve`. Later, you query via HTTP POST with JSON, e.g. `{"queries":[{"text":"What is the capital of France?"}], "n_docs":5}`. For scale or production, you can swap FAISS for Qdrant. Qdrant offers disk-backed storage, payload filters, and vector quantization (int8, scalar quantile, etc.). You install with `pip install 'pixelrag[serve,qdrant]'`, spin up Qdrant via Docker, and point PixelRAG at it. That setup lets multiple servers share one collection and tune memory vs. recall.