langgraph/concepts/human_in_the_loop/

[giscus[bot]]

langgraph/concepts/human_in_the_loop/

Build language agents as graphs

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

[phfifofum]

I have a backend, fast api and front end streamlit architecture.... struggling to get this working well.

I can edit and resume the graph fine from the back end and fast api using checkpointing.

What I cant do well (or cleanly) is trigger a graph update from streamlit front end so the events stream nicely into my existing graph invoke code and event/UI handling.

Any insights or code examples for this?

[idotr7]

i have this problem a well, it's not clear how to make it work

[cris-m]

I have try this and I am still getting HumanMessage instead of AIMessage after resuming the state.

from langgraph.graph import StateGraph, MessagesState
from langgraph.types import interrupt, Command
from langgraph.checkpoint.memory import MemorySaver

def human_node(state: MessagesState):
    value = interrupt(f"What should I say in response to {state['messages']}")

    return {"messages": [{"role": "user", "content": value}]}

checkpointer = MemorySaver()
graph_builder = StateGraph(MessagesState)
graph_builder.add_node(human_node)
graph_builder.set_entry_point("human_node")
graph = graph_builder.compile(
    checkpointer=checkpointer
)

First call

thread_config = {"configurable": {"thread_id": "some_id"}}
graph.invoke({"messages": [{ "role": "user", "content": "Hello!"}]}, config=thread_config)

I go the following out put
{'messages': [HumanMessage(content='Hello!', additional_kwargs={}, response_metadata={}, id='dce77926-559a-4900-bf80-63e3879aecd9')]}

Now resume the graph:

graph.invoke(Command(resume="Hello, How can i help you?"), config=thread_config)

The output is 2 HummanMessage:

{'messages': [HumanMessage(content='Hello!', additional_kwargs={}, response_metadata={}, id='dce77926-559a-4900-bf80-63e3879aecd9'),
  HumanMessage(content='Hello, How can i help you?', additional_kwargs={}, response_metadata={}, id='e20ad3e8-86d9-4bd1-a7b1-c17f88a8a421')]}

[eyurtsev]

I have try this and I am still getting HumanMessage instead of AIMessage after resuming the state.

The code that you shared is expected to be doing that. There's no node for a chat model to add an AIMessage.

What you're doing is:

1. graph.invoke({"messages": [{ "role": "user", "content": "Hello!"}]}, config=thread_config) writing a human message to the chat history
2. value = interrupt(f"What should I say in response to {state['messages']}") pausing immediately to collect human input
3. graph.invoke(Command(resume="Hello, How can i help you?"), config=thread_config) followed by: return {"messages": [{"role": "user", "content": value}]}
   writing another human message to the chat history
4. ending the graph execution.

---

I recommend starting from reading the langgraph tutorial before doing any of the concept pages or how-to guides: https://langchain-ai.github.io/langgraph/tutorials/introduction/

HIL isn't needed if the graph only contains a single agent. (In the shared code there isn't even a single agent / llm call.)

[vigneshmj1997]

Hey i am using interrupt inside sub graph and when i saw the writes variable of the interrupt
it was only half filled Writes [('00000000-0000-0000-0000-000000000000', 'resume', '3')]
any idea why ?

[vigneshmj1997]

Hey My interrupt is going in a infinite loop any solutions ?

[eyurtsev]

Start here if you're looking for help: https://stackoverflow.com/help/minimal-reproducible-example