Its just the reponse from the llm. time to first token is probably very fast, but the entire response must be generated before it is sent