LangChain, terminating a chain on specific tool output

Question

Im using a conversational agent, with some tools, one of them is a calculator tool (for the sake of example).

Agent initializated as follows:

conversational_agent = initialize_agent(
        agent='chat-conversational-react-description',
        tools=[CalculatorTool()],
        llm=llm_gpt4,
        verbose=True,
        max_iterations=2,
        early_stopping_method="generate",
        memory=memory,
        # agent_kwargs=dict(output_parser=output_parser),
    )

When the CalculatorTool is being activated, it will return a string output, the agent takes that output and process it further to get to the "Final Answer" thus changing the formatting of the output from the CalculatorTool

For example, for input 10*10, the tool run() function will return 100, which will be propagated back to the agent, that will call self._take_next_step() and continue processing the output.

It will create a final output similar the result of your prompt of 10x10 is 100

I dont want the added formatting by the LLM, just the output of 100.

I want to break the chain when the CalculatorTool is done, and have it's output returned to the client as is.

I also have have tools that return serialized data, for a graph chart, having that data re-processed by next iterations of the agent will make it invalid.

Answer 1

In order to end the chain when the a langchain Tool has completed running, the tool must be configured with return_redirect boolean flag declared in BaseTool to True

class CustomTool(BaseTool):

    name = "Custom tool"
    description = "Custom tool for some task"

    def __init__(self):
        super().__init__()
        self.return_direct = True

    def _run(self, work_order_id: str):
        raise NotImplementedError("implement run function")

    def _arun(self, radius: Union[int, float]):
        raise NotImplementedError("This tool does not support async")

The chain will be terminated by the Agent using the tool, when the tool is finished processing a given input. Returning the tool's output through AgentFinish.