Embedding Atlas: a scalable way to explore text embeddings with DuckDB

There have been several projects over the past few years to make text embeddings visually explorable (notably Nomic.ai's Nomic Atlas). However, Apple's just released a tool that makes this kind of analysis super accessible and insanely interactive.

Under the hood it's powered by Mosaic[0], a dataviz library built on top of DuckDB that's designed to handle coordinated interactive plots over huge datasets, the kind of thing where you interact with one plot and the rest all respond, which requires going back to the database to recalculate all the aggregations.

I've been fanboying Mosaic for the past year but finally have this to point to as an illustration of what's possible with it.

[0]: https://idl.uw.edu/mosaic