What is a Similarity Search?

The technique underlies semantic search but is more general: given a query vector (an embedding of text, an image, audio, or any encoded object) similarity search finds the nearest points in a vector space using a distance metric like cosine similarity or Euclidean distance. It applies well beyond text — image similarity, audio fingerprinting, and recommendation systems ("users who liked this also liked") all rely on the same nearest-neighbor principle. At scale, exact nearest-neighbor search against millions or billions of vectors is computationally expensive, so production systems use approximate nearest-neighbor (ANN) algorithms and specialized index structures (HNSW, IVF) that trade a small amount of accuracy for dramatically faster lookup. This is the core capability vector databases are built around. The quality of similarity search is bounded by the quality of the embeddings it operates on — a similarity search over a poorly-trained or mismatched embedding model will return results that are mathematically nearby but not meaningfully similar, which is why embedding-model choice matters as much as the search algorithm itself.