Can I use astream_events like this?

[ivanzhu109]

Scenario:
When the ChatModel Node outputs tools, it returns the full parameters without streaming. When it outputs content, it uses streaming. The Tool Node directly returns the execution result.

Code example:

config = {'configurable': {'thread_id': f'{user_info.build_thread_id()}', 'user_info': user_info}}
    state = {'messages': [{'role': 'user', "content": user_input}]}
    async for event in app.state.agent.astream_events(state, config=config, stream_mode="messages"):
        event_type = str(event['event'])
        if event_type == 'on_chat_model_stream':
            data = event['data']['chunk']
            if data.tool_call_chunks or data.response_metadata.get('finish_reason') == 'tool_calls':
                continue
            yield ServerSentEvent(str(data.content)).encode()
        elif event_type == 'on_chat_model_end':
            data = event['data']['output']
            if data.response_metadata.get('finish_reason') == 'tool_calls':
                yield ServerSentEvent(str(data.tool_calls)).encode()
        elif event_type == 'on_chain_end':
            if event['name'] != 'tools':
                continue
            yield ServerSentEvent(str(event['data']['output'])).encode()

[dosubot[bot]]

Hi @ivanzhu109! I'm Dosu and I’m helping the langchain team.

Yes, your usage of astream_events matches LangChain's event model and best practices for streaming content and handling tool outputs. When the chat model streams content, you correctly yield content chunks from on_chat_model_stream events, and when it outputs tool calls, you wait for the full parameters in on_chat_model_end before yielding. Filtering for tool node results via on_chain_end with event['name'] == 'tools' is also standard practice. This approach is supported by both the codebase and documentation, including the handling of .tool_call_chunks and finish_reason to distinguish between streaming content and tool calls [docs].

A couple of caveats: not all providers support streaming tool calls, so behavior may vary depending on the underlying model [tool streaming docs]. Also, if you use Python ≤ 3.10 or custom runnables, make sure to propagate the config/callbacks to ensure events are emitted as expected [event streaming docs]. Your implementation is robust and idiomatic for the scenario you described.

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}