Datamaps are 2D maps which visualize hyperdimensional latent spaces

[I was talking to Joe at the end of the recent gathering at Fenwick & West on the 29th, “AI Trust & Security in AI: Demo Night.” He said I should submit a datamap demo proposal. So, here it is.]

Datamaps are a new visualization technique custom designed for vector databases, RAG systems, or more generally any collection of embedding vectors, that is any “AI-ready data.” The best-of-breed datamap tooling is all permissively licensed open source such as UMAP, HDBSCAN, Tomonymy, DataMapPlot, etc.

This lightning talk explains what datamaps are by providing a live interactive demo of Ai2 Semantic Scholar’s curated library of scientific papers. The demo shows to how to DIY code up datamaps via open source. The (pre-cooked) example datamap would be a single static HTML page which each attendee can simultaneous load onto their phone and interact with. Here is the datamap they would experience:

https://connoiter.com/datamap/55948af4-903c-41cb-997e-a1dfda45eb5f/