Similarity search is the operation of finding the stored items most similar to a query item, measured by a distance metric in vector space. It is the fundamental capability of a vector database: given a query vector, return the vectors closest to it, ranked from most to least similar.
What makes it powerful is that similar is defined by meaning, not by exact matching. Because items are represented as embeddings that encode their content, two things with similar meaning have nearby vectors, so similarity search surfaces conceptually related results even when they share no literal features. This is the engine behind semantic search, recommendations, deduplication, and retrieval-augmented generation.
At scale, similarity search is implemented with approximate-nearest-neighbour algorithms that find the closest vectors quickly without comparing against every item, trading a little accuracy for enormous speed. The choice of distance metric — cosine, Euclidean, dot product — defines what similar means for a given application, and should match how the embedding model was trained. Similarity search is, in essence, what a vector database exists to do.