In this theoretical lesson, Matts presents to us the importance of the message parts that our LLM is capable of producing and it also show us the difference between the messages that are produced during a stream and the ones we get when we leverage the onFinish callback right inside the streamText().toUIMessageStream() method.

The toUIMessageStream should not be a “new entry” for us, we already met it in the sixth lesson of our first module AI SDK Basic. Basically with this method we are able to get a stream of each step that our LLM is taking in order to fulfill our request.

The full response is going to be an array of messages, to be more specific we will have array of UIMessageChunk . This type is not commonly used, but it served me well to show you which kind of structure we, or better yet the useChat hook, is dealing with.

On the other hand, the onFinish callback give us access to a single messages (type of UIMessage) object that has a standard parts prop that is holding an array of messages that are more concise if we compare it with the full stream we talked earlier. One of the most interesting thing about this kind of response is that all the tool chunks are no condensed inside a single object where its type is a string builded following this pattern: tool-<toolName>.