langgraph/concepts/functional_api/

[giscus[bot]]

langgraph/concepts/functional_api/

Build language agents as graphs

https://langchain-ai.github.io/langgraph/concepts/functional_api/

[donatoaz]

I am getting a None returned by interrupt when I invoke my entrypoint, whereas if I stream my entrypoint I get a correct result...

@entrypoint(checkpointer=MemorySaver())
def agent(messages: list[BaseMessage]):
    # we will do an initial call to the llm, maybe the user's initial message contains all necessary info
    llm_response = call_llm(messages).result()

    while True:
        # we will iterate until we have a tool call
        if not llm_response.tool_calls:
            messages = add_messages(messages, llm_response)
            clarification = interrupt({"follow_up": llm_response})
            # call llm again with the clarification
            messages = add_messages(messages, clarification)
            llm_response = call_llm(messages).result()
            # loop ...
        else:
            # oh, there is finally a tool call!
            tool_result_futures = [
                call_tool(tool_call) for tool_call in llm_response.tool_calls
            ]
            tool_results = [fut.result() for fut in tool_result_futures]

            messages = add_messages(messages, [llm_response, *tool_results])
            break

    return messages


if __name__ == "__main__":
    thread_id = "123"
    config = {"configurable": {"thread_id": thread_id}}

    msgs = [
        ("user", "I need help with my account"),
    ]

    new_state = agent.invoke(msgs, config=config)
    print(f"{new_state=}")

Prints new_state=None

Whereas

    ...

    for chunk in agent.stream(msgs, config=config):
        print(chunk)
        print("\n")

Gives me interrupt result...

{'call_llm': AIMessage(content="Certainly! I'd be happy to help you with your account...", additional_kwargs={}, response_metadata={'ResponseMetadata': {'RequestId': '1d257fd3-3823-41c3-a7fb-a4d4b4838a56', 'HTTPStatusCode': 200, 'HTTPHeaders': {'date': 'Wed, 05 Feb 2025 13:28:39 GMT', 'content-type': 'application/json', 'content-length': '676', 'connection': 'keep-alive', 'x-amzn-requestid': '1d257fd3-3823-41c3-a7fb-a4d4b4838a56'}, 'RetryAttempts': 0}, 'stopReason': 'end_turn', 'metrics': {'latencyMs': [3079]}}, id='run-a730263b-8e3d-4e29-ba7c-dc9213572e5d-0', usage_metadata={'input_tokens': 582, 'output_tokens': 118, 'total_tokens': 700})}


{'__interrupt__': (Interrupt(value={'follow_up': AIMessage(content="Certainly! I'd be happy to help you with your account...", additional_kwargs={}, response_metadata={'ResponseMetadata': {'RequestId': '1d257fd3-3823-41c3-a7fb-a4d4b4838a56', 'HTTPStatusCode': 200, 'HTTPHeaders': {'date': 'Wed, 05 Feb 2025 13:28:39 GMT', 'content-type': 'application/json', 'content-length': '676', 'connection': 'keep-alive', 'x-amzn-requestid': '1d257fd3-3823-41c3-a7fb-a4d4b4838a56'}, 'RetryAttempts': 0}, 'stopReason': 'end_turn', 'metrics': {'latencyMs': [3079]}}, id='run-a730263b-8e3d-4e29-ba7c-dc9213572e5d-0', usage_metadata={'input_tokens': 582, 'output_tokens': 118, 'total_tokens': 700})}, resumable=True, ns=['agent:469edd7a-e8ea-e92c-dabc-c0bee21c62db'], when='during'),)}

[donatoaz]

langchain-core==0.3.33
langchain-community==0.3.9
langchain-aws==0.2.9
langgraph==0.2.69

[vbarda]

@donatoaz this is the expected behavior at the moment -- we're thinking about ways to improve this devx. this is actually not limited to the functional API, StateGraph API has the same behavior. for now i would recommend relying on .stream() to surface the interrupt information

[donatoaz]

Ok, cool, thanks for the heads up... I might have missed it on the docs

[vbarda]

another thing you can do is inspect the state via agent.get_state(config) -- you should have information about interrupts contained in the returned StateSnapshot as well (under tasks)