https://arxiv.org/abs/2606.28344. The end of web parsing. The beginning of scalable pixel-native search. link: https://pixelrag.ai/
Search the web, then see the pages
Use Parallel Search MCP to discover pages beyond PixelRAG's pre-built Wikipedia index, then optionally render them as screenshot tiles with PixelRAG. Anonymous search uses Fast mode; search and page fetching are free for exploration and light use, with no account or API key. Free-tier rate limits apply.
From the repository root, install the isolated demo environment:
uv sync --project demos/parallel_search --locked
uv run --project demos/parallel_search python demos/parallel_search/run.py \
"Find explanations of retrieval augmented generation" \
--query "retrieval augmented generation overview" --limit 2 --fetch
The demo discovers web_search and web_fetch over Streamable HTTP at
https://search.parallel.ai/mcp. It saves ranked source URLs and search excerpts
in demos/parallel_search/output/results.json; --fetch adds page excerpts.
It does not read saved credentials or API keys.
Inspect the URLs first, then add --render to the command to capture the selected
pages with PixelRAG's public render_urls API. Chrome is required for screenshots:
on Linux x64, install zstd and Chrome's system libraries (on Debian/Ubuntu,
sudo apt install zstd and npx playwright install-deps chromium), then run
uv run --project demos/parallel_search pixelshot install-chrome;
elsewhere install Chrome/Chromium or set CHROME_PATH as described in the root
README. Each run writes to --output (default above), replacing results.json.
Use a separate output directory for each search manifest you want to retain.
Each screenshot capture uses a fresh subdirectory under output/tiles/, so a
new manifest cannot pick up tiles left over from an earlier page.
Screenshot paths are recorded in results.json. These are inputs for visual
inspection or the existing chunk/embed/index pipeline; this demo does not build
an index or generate answers. Search and text fetching work without Chrome.
Some sites block screenshot access or require login. Page content, excerpts and
screenshots are untrusted data, never instructions for an agent to follow.
The example runs only when invoked and leaves PixelRAG's existing search, renderer defaults and agent configuration unchanged.
uv run --project demos/parallel_search pytest demos/parallel_search/test_run.py