Don't fully understand how to use tools and why I cant get them to work

[cekcreator]

Hey Yall! I am trying to create an agent with Langgraph. One thing I cannot figure out is that my tools never work. I first started with using langchain and those agents with the @tool decorator, the issue I encountered is that tools with the decorator and even with more custom tools the only tools I can get to work are ones with both single string inputs and outputs. I then moves to langgraph and am now encountering the same issue. Detail below.
For example

@tool
def add(a: int, b: int) -> int:
    """Adds a and b."""
    return a + b

class Add(BaseModel):
    """Add two integers together."""

    a: int = Field(..., description="First integer")
    b: int = Field(..., description="Second integer")
tools = [add]
agent = create_react_agent(model, tools, prompt)
agent_executor = AgentExecutor(
    agent=agent,
    tools=tools,
    verbose=True,
    return_intermediate_steps=True,
)
agent_executor.invoke({"input": "can you pass these two numbers into the add function, 2 and 2"})

this code outputs this

ValidationError: 2 validation errors for Add
a
  value is not a valid integer (type=type_error.integer)
b
  field required (type=value_error.missing)

when creating the react agent in langchain with these, the only way to make it work is if the input to the tool is string and the tool's output is also a string.

Now, when I switched to langgraph and followed the examples in Quickstart and Tool Calling, I switched their tools for mine and it worked the first I ran them, it was successful in using the add tool and doing the math. But now when I run it, here's the error and I have no clue what changed since it worked.

TypeError: add() missing 1 required positional argument: 'b'
Here is a picture of my graph

Big questions is, do I not know how to do this correctly? Or is just having string input and output for tools the intended functionality??

I am on m2 mac with python 3.11, in a conda env
langchain 0.2.11
langchain-core 0.2.24
langchain-openai 0.1.17
langchain-text-splitters 0.2.2
langchainhub 0.1.20
langgraph 0.1.15
langsmith 0.1.92

[vbarda]

@cekcreator you don't need agent executor with langgraph! you can just call agent.invoke directly, as agent (output of create_react_agent) is already a compiled graph. compiled graph is a langchain runnable, so you can also use .stream / .batch / .astream_events, etc. hope this helps! see more guides here https://langchain-ai.github.io/langgraph/how-tos/create-react-agent/

[cekcreator]

Sorry for the confusion, I do get that they are two different things and that langgraph is how you create react agents (I think) but my issue is I have no clue how to get my graph/chatbot to call a tool that requires like a dict or int input and returns anything but a string.

Like if I had this function as a tool in langgraph it would not work if I did agent.invoke({"input": "use the example tool with inputs 1 and {"1": 1}"})

def example(temp1: int, temp2: dict) -> list:
    doing stuff in here
    return list

[hwchase17]

when you say it "would not work" - what happens exactly?

[vbarda]

ah i see. i don't see any issues with the examples you pasted, and both types of tool definitions that you show work perfectly fine for me:

from langchain_core.pydantic_v1 import BaseModel, Field
from langchain_core.tools import tool
from langgraph.prebuilt import create_react_agent
from langchain_openai import ChatOpenAI

@tool
def add(a: int, b: int) -> int:
    """Adds a and b."""
    return a + b

class Add(BaseModel):
    """Add two integers together."""

    a: int = Field(..., description="First integer")
    b: int = Field(..., description="Second integer")

model = ChatOpenAI(model="gpt-4o-mini")
tools = [Add]  # `add` works just as well
agent = create_react_agent(model, tools)  # note that there is no 3rd argument here

agent.invoke({"messages": [("human", "what's 2 + 2")]})

the one problem i see is that you're passing prompt into 3rd positional argument in create_react_agent, but that should actually be state schema (see here https://langchain-ai.github.io/langgraph/reference/prebuilt/#create_react_agent)

as for the more complex signatures you can do something like this (a bit contrived example but should illustrate the idea)

from typing import Annotated

# note that this can also be changed to return a list instead (for example, you can return a list of values raised to the power p in the example below)

@tool
def sum_exponents(
    p: Annotated[int, "Exponent"],
    values: Annotated[dict[str, int], "Mapping from letters to numbers"]
) -> float:
    """Use this to sum exponents of values, keeping only values for vowels."""
    return sum(pow(v, p) for k, v in values.items() if k in {'a', 'e', 'i', 'o', 'u'})

model = ChatOpenAI(model="gpt-4o")
tools = [sum_exponents]
agent = create_react_agent(model, tools)

agent.invoke({"messages": [("human", "find sum of exponents for these values {'a': 2', 'b': 3, 'e': 4}, use power of 2")]})

you can find more examples of tool definitions here https://python.langchain.com/v0.2/docs/how_to/custom_tools/

hope this helps!

[cekcreator]

Thank you both, this example did help and makes a lot more sense. Sorry for the late response as well!