Large language models (LLMs) are based on the transformer architecture, a complex deep neural network whose input is a sequence of token embeddings.