What is embedding?

An embedding is a list of numbers, typically a few hundred to a few thousand of them, that represents the meaning of a piece of text. An embedding model turns text into that list so that texts with similar meanings get similar numbers, measured by the distance between the lists. "Can I get my money back?" and a page titled "Refund policy" share no words, but their embeddings sit close together.

It matters because embeddings are what make retrieval work on meaning rather than on exact words. Each chunk of the knowledge base is embedded once when it is imported; each visitor question is embedded as it arrives; the chunks whose embeddings are nearest to the question's are retrieved. This is vector search, also called semantic search, and it is the first half of retrieval-augmented generation.

Where the idea runs out: similarity of meaning is not relevance to an answer. A product code, an error message or a part number matches badly on meaning and well on exact text, which is why many systems add keyword search beside vector search. Embeddings from different models are not comparable, so changing the embedding model means re-embedding the whole knowledge base. And the embedding of a question no page answers still has nearest neighbors; retrieval always returns something, which is why the model must also be told when "nearest" is not "relevant" and should say it does not know.