How to force a Tool call for an agent

OSS Product Help LangGraph
1

Hi,
I have a graph with some basic math tools, for example:

def add(a: int, b: int) -> int:
    """Adds a and b.

    Args:
        a: first int
        b: second int
    """
    return a + b

Say I’m receiving the parameters a and b from an API (so not from the user), how can I make a call to the graph using these parameters.

The full endpoint looks like this:

@app.post("/math_operations")
async def chat_with_math_payload(input: MathWidget) -> dict:
    """test to receive parameters to math operations
    """
    print(f"{input=}")

    payload = {
        "operator": str(input.operator), 
        "first number": str(input.num_1),
        "second number": str(input.num_2), 
    }

    # should match the names of the math_agent tools
    if input.operator == "+":
        operator_name = "add"
    elif input.operator == "-":
        operator_name = "subtract"
    elif input.operator == "*":
        operator_name = "multiply"
    else:
        operator_name = "divide"

    message = AIMessage(
            content="",
            tool_calls=[
                {
                    "name": operator_name,
                    "args": {"a": input.num_1, "b": input.num_2},
                    "id": "tool_call_id",
                    "type": "tool_call",
                }
            ],
        )
    res = graph.invoke({"messages": [message]},)
    return res

But I’m getting the following error:

openai.BadRequestError: Error code: 400 - {'error': {'message': "An assistant message with 'tool_calls' must be followed by tool messages responding to each 'tool_call_id'. The following tool_call_ids did not have response messages: tool_call_id", 'type': 'invalid_request_error', 'param': 'messages', 'code': None}}

Should I use a unique (or existing) id ?

thanks.

2

The error occurs because the model expects tool results for each tool call ID but did not receive them.

So your issue is not forcing a tool call, it is sending back the tool result to the model. So execute the tool calls yourself and send back ToolMessages with matching IDs before invoking the model again.

3

thanks a lot @Jameskanyiri

Below is my quick solution. I imported the tools available by the agent and executed by hand before with the proper args.

That might not work for all tool and it’s clearly a toy example but I see the mechanism.

Appreciate the help !

# ugly hack - FIX ME later
    res = ""

    # should match the names of the math_agent tools
    if input.operator == "+":
        operator_name = "add"
        tmp = add(input.num_1, input.num_2)
    elif input.operator == "-":
        operator_name = "subtract"
        tmp = subtract(input.num_1, input.num_2)
    elif input.operator == "*":
        operator_name = "multiply"
        tmp = multiply(input.num_1, input.num_2)
    else:
        operator_name = "divide"
        tmp = divide(input.num_1, input.num_2)

    print(f"{operator_name=},{input.num_1=}, {input.num_2=}")
    messages = []

    messages.append(AIMessage(
            content="",
            tool_calls=[
                {
                    "name": operator_name,
                    "args": {"a": input.num_1, "b": input.num_2},
                    "id": "call_18TdmTfJMty5bRiltHyZepbV",
                    "type": "tool_call",
                }
            ],
        ))
    messages.append(ToolMessage(content=str(tmp), name='multiply', id='e184c5b3-8527-4457-a53f-fab32ec4f06b', tool_call_id='call_18TdmTfJMty5bRiltHyZepbV'))
    res = graph.invoke({"messages": messages},)