Does langgraph support streaming of reasoning content?

[danni-boii]

Hi, I am currently developing some chat bot using langgraph. I am currently studying models with reasoner supports. Like deepseek R1 ,OpenAI o1, ...etc.

I knew that the langchain package seems to support the message chunking for deepseek R1's reasoning by detecting the reasoning_content key in the api response.

And currently, I can see that langgraph did serve the reasoning_content in its api. However, it seems that the reasoning content is only available after the runs is finished.

I wonder if langgraph has the ability to support the streaming of reasoning context just like how deepseek chat did.

[danni-boii]

I found it. If the langgraph is set to message stream mode, the stream would indeed contain the reasoning content.

Here is an example message object

{
    "event": "messages",
    "data": [
        {
            "content": "",
            "additional_kwargs": {
                "reasoning_content": "Alright"
            },
            "response_metadata": {},
            "type": "AIMessageChunk",
            "name": null,
            "id": "run-b8e386dc-4639-4875-a57e-2fa1e9c0acc7",
            "example": false,
            "tool_calls": [],
            "invalid_tool_calls": [],
            "usage_metadata": null,
            "tool_call_chunks": []
        },
        {
            "created_by": "system",
            "thread_name": "Explain rocket science",
            "graph_id": "graph",
            "assistant_id": "b54c0545-9d79-5688-bc81-269c46247a52",
            "run_attempt": 1,
            "langgraph_version": "0.2.76",
            "langgraph_plan": "developer",
            "langgraph_host": "self-hosted",
            "llm_provider": "deepseek_r1",
            "system_message": "You are a helpful assistant.",
            "langgraph_auth_user_id": "",
            "checkpoint_id": "1eff4c0a-fb34-690a-8003-1c5850baddfc",
            "run_id": "1eff4c21-f9fa-69d1-ac7b-8eba62654de2",
            "thread_id": "5754f72e-29c8-439d-9487-ed04f11b2edb",
            "user_id": "",
            "langgraph_step": 4,
            "langgraph_node": "supervisor",
            "langgraph_triggers": [
                "start:supervisor"
            ],
            "langgraph_path": [
                "__pregel_pull",
                "supervisor"
            ],
            "langgraph_checkpoint_ns": "supervisor:bedbf9f3-4549-53b7-994b-ba9bc38f1ca2",
            "checkpoint_ns": "supervisor:bedbf9f3-4549-53b7-994b-ba9bc38f1ca2",
            "ls_provider": "openai",
            "ls_model_name": "deepseek-reasoner",
            "ls_model_type": "chat",
            "ls_temperature": null
        }
    ]
}

[danni-boii]

Hi, Looks like you are using the python sdk. I am only using js sdk for my project. But I guess the reasoning_content is located under `chunk_msg.additional_kwargs.reasoning_content` Let me know if you need any more info. Best Regards, Danny Jinseok Hong ***@***.***> 於 2025年3月11日週二 下午1:43寫道：

…

Hi, did you resolve it? I cannot get streaming result for reasoning content as below code. for chunk_msg, metadata in grpah.stream({"question": query}, stream_mode="messages"): print(chunk_msg.reasoning_content) the graph contains openai format r1 response — Reply to this email directly, view it on GitHub <#3547 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ANJNBVWIATQHTEKJ3R7WKBT2TZZWJAVCNFSM6AAAAABXSVKQEWVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTENBVG42TOOI> . You are receiving this because you modified the open/close state.Message ID: ***@***.*** .com>

[mrbaowei]

Well, it's strange. When using the Doubao 1.6thinking model, the reasoning_content can be normally obtained in the return of the OpenAI SDK, but it cannot be found in Langgraph, and it is not in additional_kwargs.

events = agent_executor.astream(
            {"messages": messages},
            config,
            stream_mode="messages"
        )
        
        async for event in events:
            print(f"event: {event} \n")
            if "messages" in event:
                event["messages"][-1].pretty_print()

[826327700]

metoo! If you find out how to receive reasoning_content, and how to set thinking: {
type: "disabled"
}, Can you reply to me?

[Rinzyy]

Having the same issue, I'm using js sdk. Even the normal OpenAI model doesnt output the summary of the reasoning