Issues while using human in the loop.

[Harshs-09]

Can anyone point put what am I doing wrong here as my graph does not resume instead it runs twice when provide with clarification.

class State(TypedDict):
messages: Annotated[list, add_messages]
clarification: Annotated[
str,
"Request for clarification from the human. "
"The model should populate this field whenever human assistance is needed "
"because the query is ambiguous or missing details.",
]
human_response: Annotated[
str,
"Response from the human when asked for clarification."
]
human_required: Annotated[
bool,
"Set to True if human is required in the loop for additional clarification else False"
]

class LLMResponse(BaseModel):
clarification: str = Field(description="Request message for clarification from the human whenever human assisstance is required if the query is ambiguous or missing details.")
human_required: bool = Field(description='Set to True if human is required in the loop for additional clarification else False')
response: Any = Field(description="Response from the LLM.")

llm = llm.with_structured_output(LLMResponse, strict=True, method='json_mode')

def human_assistance(state: State):
"""Request assistance from a human."""
print("Requesting information from human.")
human_response = interrupt({"query": state.get("clarification")})

return {
    'human_response': human_response
}

def chatbot(state: State):
print("Initial state:", state)
# Append the new message to the existing messages in the state
messages = state.get("messages", [])
llm_response = llm.invoke(messages)
message = AIMessage(content=llm_response.response)
return {
"messages": messages + [message], # Append the new message
"clarification": llm_response.clarification, # Access attribute directly
"human_required": llm_response.human_required # Access attribute directly
}

def need_human(state: State) -> Literal['human_assistance', 'END']:
print("Need human:", state)
if state.get('human_required'):
return 'human_assistance'
else:
return "END"

def end(state: State):
print("Ending the graph.")

graph_builder = StateGraph(State)
graph_builder.add_node("chatbot", chatbot)
graph_builder.add_node("END", end)
graph_builder.set_entry_point("chatbot")
graph_builder.add_conditional_edges(
"chatbot",
need_human,
)
graph_builder.add_node("human_assistance", human_assistance)
graph_builder.add_edge("human_assistance", "chatbot")
graph_builder.add_edge("chatbot", "END")
graph_builder.add_edge("END", END)
memory = InMemorySaver()
graph = graph_builder.compile(checkpointer=memory)

thread_id = str(uuid.uuid4())
config = {"configurable": {"thread_id": thread_id}}

Include the SystemMessage in the initial list of messages

result = graph.invoke(
{"messages": [SystemMessage(content=system_message_content), HumanMessage(content=user_query)]}, config=config
)
print(f"Results: \n {result['messages']}")

human_response = input("Enter your response:")
result = graph.invoke(
Command(resume={'query': human_response}), config=config
)

[augustocesarperin]

You wired both conditional edges and an unconditional chatbot -> END.
Use END as a sentinel only (no node/edge) and return it only from the conditional.
Also Include the human reply in messages before invoking the LLM.

[Harshs-09]

Thanks a lot for the clarification and guidance.

Just to confirm my understanding: when the graph is resumed, the entire human_assistance node is executed again from the top. The difference is that since the human response is already available, the interrupt call doesn’t block this time — it immediately returns the provided value.

So effectively the node re-runs, but the flow is no longer paused at the interrupt, and execution continues smoothly back into the graph.

Thanks again for the help.

[augustocesarperin]

Yep, that's it! :)