Executor calling CustomTool with no parameter needed insists to call the tool with a parameter coming from nowhere

[jeromemassot]

System Info

langchain==0.0.232
Google Colab

Who can help?

No response

Information

• The official example notebooks/scripts
• My own modified scripts

Related Components

• LLMs/Chat Models
• Embedding Models
• Prompts / Prompt Templates / Prompt Selectors
• Output Parsers
• Document Loaders
• Vector Stores / Retrievers
• Memory
• Agents / Agent Executors
• Tools / Toolkits
• Chains
• Callbacks/Tracing
• Async

Reproduction

def get_current_oil_price():
    ticker_data = yf.Ticker("CL=F")
    recent = ticker_data.history(period='1d')
    return {"price": recent.iloc[0]["Close"], "currency": ticker_data.info["currency"]}

def get_oil_price_performance(days):
    past_date = datetime.today() - timedelta(days=int(days))
    ticker_data = yf.Ticker("CL=F")
    history = ticker_data.history(start=past_date)
    old_price = history.iloc[0]["Close"]
    current_price = history.iloc[-1]["Close"]
    return {"percent_change": ((current_price - old_price) / old_price) * 100}

class CurrentOilPriceTool(BaseTool):
    name = "get_oil_price"
    description = "Get the current oil price. No parameter needed from input"


    def _run(self):
        price_response = get_current_oil_price()
        return price_response


    def _arun(self):
        raise NotImplementedError("get_oil_price does not support async")

class CurrentOilPerformanceTool(BaseTool):
    name = "get_oil_performance"
    description = "Get the current oil price evolution over a given number of days. Enter the number of days."


    def _run(self, days):
        performance_response = get_oil_price_performance(days)
        return performance_response


    def _arun(self):
        raise NotImplementedError("get_oil_performance does not support async") 

llm = ChatOpenAI(model="gpt-3.5-turbo-0613", temperature=0)

tools = [CurrentOilPriceTool(), CurrentOilPerformanceTool()]

agent = initialize_agent(
    tools=tools,
    llm=llm,
    agent=AgentType.OPENAI_FUNCTIONS,
    verbose=True
)

agent.run("What is the oil price?")

Expected behavior

> Entering new AgentExecutor chain...

Invoking: `get_oil_price`

But instead, the executor insists on adding a parameter to the tool function !!!

> Entering new AgentExecutor chain...

Invoking: `get_oil_price` with `USD`

provoking an obvious error:

TypeError: CurrentOilPriceTool._run() takes 1 positional argument but 2 were given

[dosubot]

Answer generated by a 🤖

Answer

I understand that you're experiencing an issue where the executor is insisting on passing a parameter to your get_oil_price tool, even though it doesn't require any parameters. This is causing a TypeError because the _run method in your CurrentOilPriceTool class is not expecting any arguments.

The root of this issue lies in how the BaseTool class in LangChain handles tool inputs. When a tool is run, it tries to convert the tool input into arguments and keyword arguments for the tool's _run or _arun method. If the tool input is a string, it is passed as a positional argument. This is why the executor is trying to pass "USD" as a parameter to your get_oil_price tool.

One potential solution to this issue is to override the _to_args_and_kwargs method in your CurrentOilPriceTool class to always return an empty tuple for arguments and an empty dictionary for keyword arguments, regardless of the tool input. This will ensure that no parameters are passed to the _run method.

Here's how you can implement this solution:

class CurrentOilPriceTool(BaseTool):
    name = "get_oil_price"
    description = "Get the current oil price. No parameter needed from input"

    def _to_args_and_kwargs(self, tool_input: Union[str, Dict]) -> Tuple[Tuple, Dict]:
        return (), {}

    def _run(self):
        price_response = get_current_oil_price()
        return price_response

    def _arun(self):
        raise NotImplementedError("get_oil_price does not support async")

This modification overrides the _to_args_and_kwargs method from the BaseTool class, ensuring that no parameters are passed to the _run method, regardless of the tool input.

Please note that this is just one potential solution. If you believe this is a bug that could impact other users, you're encouraged to make a pull request with this change. If this solution doesn't align with your experience or if you're still encountering issues after implementing it, I would recommend upgrading to the latest version of LangChain to see if that resolves the issue.

Sources

• langchain/tools/base.py

---

This response is meant to be useful, save you time, and share context. 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.

[jeromemassot]

Superb and helpful. Good bot, good bot :)

[lmaddox]

I'm still having issue with the input args. It seems to happen where there are more than one no-argument functions. In this case, I have two.

"""
Question: What is the purpose of the my_name_tool?

Thought: The my_name_tool is a tool that allows me to learn more about myself. It can provide information about my name, the name of the currently running program, and other details about how I operate.

Action:

{                                                                                                                                                                                                   
  "action": "introspect",                                                                                                                                                                           
  "action_input": ""                                                                                                                                                                                
}                                                                                                                                                                                                   

"""

Traceback (most recent call last):                                                                                                                                                                  
  File "/home/lmaddox/.local/lib/python3.10/site-packages/langchain/agents/agent.py", line 1224, in _iter_next_step
    yield self._perform_agent_action(
  File "/home/lmaddox/.local/lib/python3.10/site-packages/langchain/agents/agent.py", line 1246, in _perform_agent_action
    observation = tool.run(
  File "/home/lmaddox/.local/lib/python3.10/site-packages/langchain_core/tools.py", line 452, in run
    raise e
  File "/home/lmaddox/.local/lib/python3.10/site-packages/langchain_core/tools.py", line 406, in run
    parsed_input = self._parse_input(tool_input)
  File "/home/lmaddox/.local/lib/python3.10/site-packages/langchain_core/tools.py", line 299, in _parse_input
    key_ = next(iter(input_args.__fields__.keys()))
StopIteration

Update: workaround

@typechecked
class MyName(BaseModel): # TODO
    unused:str # just give it *something* for that input_args iterator in the sub-layers
    pass

@typechecked
class MyNameTool(StructuredTool):
    name       :str  = "What is my name?"
    description:str  = '''
            the name of the currently running program
            (useful to learn your name)

            https://stackoverflow.com/questions/4152963/get-name-of-current-script-in-python
            '''
    args_schema:Type = MyName

    def _to_args_and_kwargs(self, tool_input: Union[str, Dict]) -> Tuple[Tuple, Dict]:
        return (), {}

    def _run(self, *args:P.args, **kwargs:P.kwargs)->str:
        return str(get_path_executed_script())

    def _arun(self)->str:
        raise NotImplementedError("my_name does not support async")