Langgraph: how to use user input with supervisor agent?

[Xakim1c]

Checked other resources

• I added a very descriptive title to this question.
• I searched the LangChain documentation with the integrated search.
• I used the GitHub search to find a similar question and didn't find it.

Commit to Help

• I commit to help with one of those options 👆

Example Code

items_agent = create_agent(llm, [create_items], "You are items creater.")
items_node = functools.partial(agent_node, agent=items_agent, name="Items Creater")

customer_agent = create_agent(llm, [create_customer], "You are customer creator")
customer_node = functools.partial(agent_node, agent=customer_agent, name="Customer Creater")

workflow = StateGraph(AgentState)
workflow.add_node("Items Creater", items_node)
workflow.add_node("Customer Creater", customer_node)
workflow.add_node("supervisor", supervisor_chain)

Description

I'm using example of creating Agent Supervisor: https://github.com/langchain-ai/langgraph/blob/main/examples/multi_agent/agent_supervisor.ipynb

The problem here that for Customer Creater agent I need to ask user's Firstname, Lastname and phone number. But I don't know how to put user input between Customer Creater agent and Supervisor or how to make an agent that can wait user's answer before go to Supervisor

As a solution I tried to modify last node of agent (before was supervisor, I changed to "END" + I tried to add memory to agents:
for member in members:
# We want our workers to ALWAYS "report back" to the supervisor when done
workflow.add_edge(member, END)

but this solution isn't reliable

System Info

langchain==0.1.0
langchain-community==0.0.11
langchain-core==0.1.9
langchain-google-genai==0.0.5
langchain-openai==0.0.2
langchainhub==0.1.14
langchainplus-sdk==0.0.20

[dosubot[bot]]

🤖

Hey there, @Xakim1c! Great to see you back with another interesting challenge. Hope you've been doing well since our last chat!

To incorporate user input into the Customer Creator agent before it proceeds to the Supervisor in the LangChain framework, you can modify the create_conversational_retrieval_agent function to accept an additional parameter, say user_input. This user_input can then be used to create a custom system message or be added to the extra_prompt_messages list.

Here's how you can modify the function:

def create_conversational_retrieval_agent(
    llm: BaseLanguageModel,
    tools: List[BaseTool],
    user_input: str,  # new parameter for user input
    remember_intermediate_steps: bool = True,
    memory_key: str = "chat_history",
    system_message: Optional[SystemMessage] = None,
    verbose: bool = False,
    max_token_limit: int = 2000,
    **kwargs: Any,
) -> AgentExecutor:
    """A convenience method for creating a conversational retrieval agent.

    Args:
        llm: The language model to use, should be ChatOpenAI
        tools: A list of tools the agent has access to
        user_input: The user input to incorporate into the agent
        remember_intermediate_steps: Whether the agent should remember intermediate
            steps or not. Intermediate steps refer to prior action/observation
            pairs from previous questions. The benefit of remembering these is if
            there is relevant information in there, the agent can use it to answer
            follow up questions. The downside is it will take up more tokens.
        memory_key: The name of the memory key in the prompt.
        system_message: The system message to use. By default, a basic one will
            be used.
        verbose: Whether or not the final AgentExecutor should be verbose or not,
            defaults to False.
        max_token_limit: The max number of tokens to keep around in memory.
            Defaults to 2000.

    Returns:
        An agent executor initialized appropriately
    """

    if remember_intermediate_steps:
        memory: BaseMemory = AgentTokenBufferMemory(
            memory_key=memory_key, llm=llm, max_token_limit=max_token_limit
        )
    else:
        memory = ConversationTokenBufferMemory(
            memory_key=memory_key,
            return_messages=True,
            output_key="output",
            llm=llm,
            max_token_limit=max_token_limit,
        )
    _system_message = system_message or _get_default_system_message()
    prompt = OpenAIFunctionsAgent.create_prompt(
        system_message=_system_message,
        extra_prompt_messages=[MessagesPlaceholder(variable_name=memory_key), user_input],  # add user input here
    )
    agent = OpenAIFunctionsAgent(llm=llm, tools=tools, prompt=prompt)
    return AgentExecutor(
        agent=agent,
        tools=tools,
        memory=memory,
        verbose=verbose,
        return_intermediate_steps=remember_intermediate_steps,
        **kwargs,
    )

In this modification, user_input is added to the extra_prompt_messages list. This means that the user input will be incorporated into the prompt that the agent uses to generate responses.

You can find more details about this function in the libs/langchain/langchain/agents/agent_toolkits/conversational_retrieval/openai_functions.py file in the LangChain repository.

Please note that this modification assumes that the user input is available at the time of creating the agent. If the user input is not available at that time, you might need to modify the workflow to include a step for collecting user input before creating the agent.

Sources

• libs/langchain/langchain/agents/agent_toolkits/conversational_retrieval/openai_functions.py
• templates/rag-ollama-multi-query/rag_ollama_multi_query/chain.py

---

This response is meant to be useful and save you time. It is not meant to be a precise solution, but rather a starting point for your own research.

Help me be more useful! Please leave a 👍 if this is helpful and 👎 if it is irrelevant.

If you want to continue the conversation, start your reply with @dosu-bot.

[Xakim1c]

@dosu-bot thank you for your reply, but I think it's not a good idea to change functions from libraries

[dosubot[bot]]

🤖

Based on your question, it seems like you want to incorporate user input into the Customer Creator agent before it proceeds to the Supervisor. To achieve this, you can use the create_conversational_retrieval_agent function and modify it to accept an additional parameter for user input. This user input can then be used to create a custom system message or be added to the extra_prompt_messages list.

Here's how you can modify the function:

def create_conversational_retrieval_agent(
    llm: BaseLanguageModel,
    tools: List[BaseTool],
    user_input: str,  # new parameter for user input
    remember_intermediate_steps: bool = True,
    memory_key: str = "chat_history",
    system_message: Optional[SystemMessage] = None,
    verbose: bool = False,
    max_token_limit: int = 2000,
    **kwargs: Any,
) -> AgentExecutor:
    """A convenience method for creating a conversational retrieval agent."""

    if remember_intermediate_steps:
        memory: BaseMemory = AgentTokenBufferMemory(
            memory_key=memory_key, llm=llm, max_token_limit=max_token_limit
        )
    else:
        memory = ConversationTokenBufferMemory(
            memory_key=memory_key,
            return_messages=True,
            output_key="output",
            llm=llm,
            max_token_limit=max_token_limit,
        )
    _system_message = system_message or _get_default_system_message()
    prompt = OpenAIFunctionsAgent.create_prompt(
        system_message=_system_message,
        extra_prompt_messages=[MessagesPlaceholder(variable_name=memory_key), user_input],  # add user input here
    )
    agent = OpenAIFunctionsAgent(llm=llm, tools=tools, prompt=prompt)
    return AgentExecutor(
        agent=agent,
        tools=tools,
        memory=memory,
        verbose=verbose,
        return_intermediate_steps=remember_intermediate_steps,
        **kwargs,
    )

In this modification, user_input is added to the extra_prompt_messages list. This means that the user input will be incorporated into the prompt that the agent uses to generate responses.

You can find more details about this function in the LangChain repository.

Please note that this modification does not override any private methods or add new methods to existing classes within the repository. It simply extends the functionality of the create_conversational_retrieval_agent function to accept user input and incorporate it into the agent's prompt.

I hope this helps! If you have any further questions, feel free to ask.

Sources

• libs/langchain/langchain/agents/agent_toolkits/conversational_retrieval/openai_functions.py
• templates/rag-ollama-multi-query/rag_ollama_multi_query/chain.py

---

This response is meant to be useful and save you time. It is not meant to be a precise solution, but rather a starting point for your own research.

Help me be more useful! Please leave a 👍 if this is helpful and 👎 if it is irrelevant.

If you want to continue the conversation, start your reply with @dosu-bot.

[harsh020]

@Xakim1c Even I wanted to do something similar but I couldn't find anything. So, here is how I did it (modifying your code)

...

def user_node(state):
    user_input = input("Enter your response:")
    state.messages.append(HumanMessage(content=user_input))
    return state

workflow = StateGraph(AgentState)
workflow.add_node("Items Creater", items_node)
workflow.add_node("Customer Creator", customer_node)
workflow.add_node("User Input", user_node)
workflow.add_node("supervisor", supervisor_chain)

workflow.add_edge("Customer Creator", "User Input")
workflow.add_edge("User Input", "supervisor")

I might have made a mistake with the state, but I'm uncertain because I'm not utilizing the langgraph stateful graph. I had very specific memory management requirements. However, I believe you understand the situation.

Another approach could involve using the user input tool with your agent, but unfortunately, it didn't align with my needs.

However, I'm not sure if you'd be able to develop a frontend for it now, especially if your models are running on the backend. I don't believe we can achieve the request-response cycle when using graphs (let me know if you get something on this).

@hwchase17 @hinthornw This might not be the optimal solution for meeting the requirement. Is there a better approach?

[Xakim1c]

@harsh020 Hi, thank you for responding. I tried to use "from langchain.tools import HumanInputRun" but I couldn't make it work with langgraph agents

[harsh020]

@Xakim1c Try using the above approach, this is working for me.

[igormis]

@Xakim1c it works but when I add the User Input twice
`workflow = StateGraph(AgentState)
workflow.add_node("Introduction", intro_node)
workflow.add_node("Explain", explain_node)
workflow.add_node("User Input", user_node)
workflow.add_node("Reason", reason_node)
#workflow.add_node("supervisor", supervisor_chain)

workflow.add_edge("Introduction", "User Input")
workflow.add_edge("User Input", "Explain")
workflow.add_edge("Explain", "User Input")
workflow.add_edge("User Input", "Reason")
workflow.add_edge("Reason", END)`

I get `The above exception was the direct cause of the following exception:

InvalidUpdateError Traceback (most recent call last)
/usr/local/lib/python3.10/dist-packages/langgraph/pregel/init.py in _apply_writes(checkpoint, channels, pending_writes, config, for_step)
746 channels[chan].update(vals)
747 except InvalidUpdateError as e:
--> 748 raise InvalidUpdateError(
749 f"Invalid update for channel {chan}: {e}"
750 ) from e

InvalidUpdateError: Invalid update for channel sender: LastValue can only receive one value per step.`

[harsh020]

@igormis Based on the error, I think it's because your User Input node has two outgoing unconditional edges. If you think in terms of a state graph or NFA, it is invalid. If your User Input node can transfer the state to both Explain and Reason nodes, unconditionally how would it decide where to go?

If that's the issue, you can solve it by connecting User Input node to multiple nodes using conditional edges or create separate User Input nodes for each state transfer.

[igormis]

OK, got it, but what I need is to have interruption between agents with user input and I can have a lot of nodes.
Instead of state I am trying to impute the user input in this way
for s in graph.stream( { "messages": [ HumanMessage( content="Hello?" ) ], }, # Maximum number of steps to take in the graph {"recursion_limit": 150}, ): ''' print(state) state_name = next(iter(state)) messages = state[state_name] print(messages) user_input = input("Please enter something: ") print(first_key) state[state_name] ''' print(s) user_input = input("Enter your response:") mess = [] for key, value in s.items(): mess = value["messages"] mess.append(HumanMessage(content=user_input, name=key)) s[key] = mess print(s) print("----")
But the folowing error occurs:
AssertionError: Can't modify channel Introduction of type EphemeralValue

[Jeomon]

decision=lambda state: 'next' if isinstance(state['agent_outcome'],AgentFinish) else 'continue'

workflow=StateGraph(State)

workflow.add_node('developer',developer_node)
workflow.add_node('executor',executor_node)
workflow.add_node('tools',tools)

workflow.set_entry_point('developer')
workflow.add_conditional_edges('developer',decision,{'continue':'tools','next':'executor'})
workflow.add_edge('tools','developer')
workflow.add_conditional_edges('executor',decision,{'continue':'tools','next':END})
workflow.add_edge('tools','executor')

This is the graph I made it is having conditions for each agent but the when i going to the executor am getting the same error (InvalidUpdateError: Invalid update for channel input: LastValue can only receive one value per step.) you guys faced if am coming from the developer agent . Is there a way..

[rsk2327]

This is caused by two parallel nodes trying to set the same state variable. Im assuming in your State description you have a field that is static and doesnt have a logic to accommodate two parallel processes trying to set it. This causes the error.

My initial AgentState looked like this, which threw an error

class AgentState(TypedDict): messages: Annotated[Sequence[BaseMessage], operator.add] sender: str

On updating to the following, I was able to address the error

class AgentState(TypedDict): messages: Annotated[Sequence[BaseMessage], operator.add] sender: Annotated[str, operator.setitem]

[Jeomon]

if possible could you please also look into this issue as well it's related to the tool usage were the LLM is giving parsing error but don't know why its happening.
click here