Grouping items by similarity with no labels. Run over embeddings it produces topic groups, near-duplicate sets and query groups that nobody defined in advance.
Clustering groups items by similarity without labels. Run over vector embeddings it turns a pile of pages, queries or entities into groups nobody defined in advance, which is how topic maps, near-duplicate detection and query grouping are built.
Every method needs a distance. For text embeddings that is usually cosine similarity, which compares direction and ignores magnitude. K-means partitions items into a fixed number of groups by moving centroids until they settle, and k has to be chosen up front. Hierarchical clustering merges the closest pairs repeatedly and yields a tree that can be cut at any level. Density methods such as DBSCAN and HDBSCAN find groups of any shape and leave sparse points unassigned, which suits real corpora where some pages belong nowhere.
There is no ground truth, so quality is measured internally: the silhouette score compares each item's distance to its own cluster against the nearest other cluster. The outcome depends on the embedding as much as on the algorithm, because clustering can only find structure the embedding space already encodes. The manifold hypothesis is why it works at all: real data occupies a low-dimensional surface inside the high-dimensional space, and clusters are dense regions of that surface.
