That's what a very common ada embedding scheme decided to use. Different schemes have different vector counts. But they essentially distill key "hidden features" or "most important characteristics" of language, images etc so that they can be compared and cross referenced at runtime to create a usef...