Streaming: Values and Updates

Where we left off

Every lesson so far has used .invoke(), which runs the whole graph and hands you back the final state only once it's completely done. For a graph with several steps, that means waiting for everything before seeing anything. .stream() instead gives you a chunk of output after every single node finishes, which matters a lot once graphs (or model responses) take more than an instant to run.

What we're building

Lesson 3's four-node text pipeline (clean, split, count, format), streamed two different ways so you can see exactly what each stream_mode gives you.

The code, piece by piece

for chunk in app.stream(initial_state, stream_mode="values"):
    print(chunk)

stream_mode="values" yields the entire current state, as it stands, after each node finishes. Every chunk is a full snapshot, so if your state has five fields and one node just changed one of them, you still get all five fields back, just with that one field's new value. This is the easiest mode to reason about because each chunk is "complete," but it can repeat a lot of unchanged data on graphs with many fields.

for chunk in app.stream(initial_state, stream_mode="updates"):
    print(chunk)

stream_mode="updates" yields only the diff: a dictionary shaped like {"node_name": {the partial dict that node returned}}. Instead of the whole state, you get exactly what changed and which node changed it. This is the mode to reach for when you care about "what just happened" rather than "what does everything look like right now," useful for logging progress or driving a UI that shows per-step activity.

When to use which

Use "values" when you want to display or act on the state as a whole at each step (say, showing the current draft of a document as it's built up). Use "updates" when you want a clean, structured log of which node did what, without extra fields cluttering each entry. Both modes stream once per node execution, they differ only in what shape of information you get for each step, not in when you get it.