Embedding
An embedding is a list of numbers that captures the meaning of a piece of text, so that texts about the same thing end up close together — the machinery behind semantic search and RAG.
Computers cannot compare meaning directly, so an embedding model turns a sentence, a paragraph or a whole document into a vector — a few hundred to a few thousand numbers. Texts that mean similar things get vectors that point in similar directions, whatever words they use. “How do I lower VRAM use?” and “reducing graphics memory” land close together; a recipe for pancakes lands far away.
That is the trick behind retrieval: index every passage of your documents as a vector once, embed the question at query time, and fetch the passages whose vectors are nearest. No keywords, no exact matches — meaning. Embedding models are small compared with chat models, often well under a billion parameters, so they run happily on a laptop CPU and can index a large archive locally.
Two things to know before relying on them. Vectors from different embedding models are not comparable, so an index has to be rebuilt if you switch models. And an embedding captures what a passage is about, not whether it is true — retrieval finds relevant text; the chat model still has to read it critically.
Where you see it on the radar
Embedding models are not on the board — the radar tracks chat and coding models, and an index built on your machine costs no context at all. What does cost context is the passages it retrieves, and the finder prices that per context length.
See how retrieved context affects fit →