https://journofinder.com/news-graph

Each one is embedded with OpenAI's text-embedding-3-large, clustered into stories with hnswlib, laid out in 2D with UMAP so related topics sit near each other, grouped into topic regions with recursive DBSCAN, and labeled by an LLM. The map is rebuilt from scratch every hour.

I deliberately skipped progressive loading. The whole map loads upfront, ~18MB of data, and after that everything runs client-side with zero network requests: pan, zoom, search, cluster panels. deck.gl handles all 133k points fine, even on mobile.

Currently showing ~133,000 stories across 1,081 topics.

Happy to answer any questions.