PixelRAG

PixelRAG is a retrieval framework that indexes document collections as rendered page images rather than extracted text, then searches over those images using vision-language models. The core workflow: pages are captured as screenshots, indexed for visual retrieval, and at query time the system returns the relevant page images alongside the matched context — which an LLM or VLM then reads directly. The vendor describes both a hosted API endpoint (no key required) and a pip-installable open-source package for self-hosted deployments.

The differentiating bet is that rendering before indexing preserves what parsers destroy. A table rendered in a PDF, a chart embedded in a Wikipedia article, a multi-column layout with sidebars — these survive intact as pixels in a way they do not survive HTML stripping or OCR passes. For document QA tasks where the answer lives inside a table cell or requires reading a figure caption in context, this matters.

PixelRAG fits into pipelines that feed retrieved context to Claude or other VLMs capable of reading image inputs. The vendor positions it as a way to reduce token usage while improving accuracy on visually dense content — the claim being that a clean page screenshot sent to a VLM outperforms a noisy, structure-stripped text chunk. Where it does not fit: collections where text extraction is already reliable and vision inference overhead is not justified, or production environments that require documented throughput benchmarks and active community support for debugging. The publicly visible evidence base centers on Wikipedia and custom document collections; behavior at significantly larger scale or with non-Latin scripts is not documented on the project pages.