embeddings

Definition

Real-valued vectors assigned to discrete things — words, sentences, images, users — such that semantic proximity becomes geometric proximity. Once the mapping exists, retrieval, recommendation, and clustering all reduce to nearest-neighbour search in that space, and the same vectors serve as conditioning for generative models. What matters is the training objective, not the dimensionality: two embedding models of identical size trained differently are not interchangeable, and mixing their vectors in one index is meaningless.