Embedding model

An embedding model maps a piece of text to a list of numbers — typically hundreds to a few thousand dimensions — positioned so that semantically similar texts land near each other. Comparing vectors then answers questions keyword search cannot: which documents mean roughly the same thing as this query.

Embeddings are the retrieval half of RAG and power semantic search, deduplication, recommendation, and classification. Multimodal embedding models extend the same idea to images and mixed content.

Pricing is input-only and cheap relative to generation — often orders of magnitude less per token than frontier chat models — so the dominant costs in embedding-heavy systems are usually vector database storage and re-embedding corpora when switching models, not the embedding calls themselves.