The model's output is a single array of numbers (a vector) rather than generated text, which can be efficiently compared with other vectors to measure similarity.