How do you compare K-Means with modern embedding-based clustering approaches?

Embedding-based clustering captures semantic structure, while K-Means relies on geometric distance.

Modern systems often use neural embeddings (e.g., from transformers) where semantic similarity is encoded in vector space. K-Means then acts as a simple grouping layer on top. However, embeddings already contain learned structure, making K-Means only a coarse segmentation tool.