关于LLM流式输出的疑问

I really appreciate the author's framework. I think the different types of abstractions can serve as a standard for engineering development. I have a question I would like to ask the author. For example, if I want to implement real-time voice dialogue, I understand that the voice input is first converted into text, and then passed through a node. The text input is passed through a large model node, and the output of the large model stream is passed to the next node for voice synthesis and then returned. If we want to achieve high speed, we would definitely prefer all the intermediate steps to be in the form of streams. That is, the voice input is passed in the form of a stream, the asr function is called and the output is in the form of a stream, the text input to the large model is in the form of a stream, the corresponding response is in the form of a stream, and the tts input is in the form of a stream, and the corresponding voice output is returned. I understand that each step should be a node, but this seems to be different from the original node definition, and the original node, if it is strictly followed in three steps, will eventually return a corresponding default action identifier. This does not seem right. If a node contains the above steps, it feels too cumbersome.