Discover Topics with Embeddings + Clustering

The pipeline

Combine two ideas you've already met: embeddings for meaning, clustering for grouping.

1. Embed each document into a vector.
2. Cluster the vectors (k-means or HDBSCAN).
3. Label each cluster by asking an LLM to summarize its members.

vectors = [embed(doc) for doc in docs]
labels = KMeans(n_clusters=8, n_init="auto").fit_predict(vectors)

Where it helps

Triaging support tickets, analyzing survey responses, or mapping a research field. It's unsupervised, so it surfaces themes you didn't know to look for — a fast way to understand a pile of text.