Clustering Embeddings for Topic Discovery

The reduce-then-cluster pipeline that works on high-dimensional embeddings, why k-means is the wrong first choice, and how to turn clusters into labels people can read.