Embeddings

An embedding is a numeric vector representing the meaning of a piece of text, image or product, positioned so that similar things sit close together in the same space.

Also known as: vector embeddings

Embeddings are produced by a model that has learned, from very large corpora, which things appear in similar contexts. The practical consequence is that distance becomes a proxy for similarity of meaning — "waterproof jacket" lands near "rain coat" without either word being declared a synonym.

For search, both the corpus and the query are embedded with the same model, and retrieval becomes a nearest-neighbour lookup. Because the comparison is on meaning rather than characters, embeddings handle paraphrase, description and intent well, and exact identifiers badly.

Embeddings are model-bound: change the model and every stored vector must be regenerated. That re-indexing cost is the main operational consideration when adopting them.