docs/concepts/tool_calling/

[giscus[bot]]

docs/concepts/tool_calling/

• Tools

https://python.langchain.com/docs/concepts/tool_calling/

[mcavdar]

model_with_tools = model.bind_tools([tools_list])

it seems like bind_tools requires list of list. I guess this might be better:
model_with_tools = model.bind_tools(tools_list)

[thomasthm]

True

[Bhargav2525]

model_with_tools = model.bind_tools(tools_list) this means we have to pass list containing tools into model
Ex:- model_with_tools = model.bind_tools([multiply,add,....])

bind_tools will take list only not list of list

[mcavdar]

it was fixed in https://github.com/langchain-ai/langchain/pull/28222/files

[smithlai]

## tool 1:
tavily_search_tool = TavilySearchResults(max_results=2)
## tool 2:
from langchain_core.tools import tool
@tool(name_or_callable="magic_words")
def magic_words(username: str) -> str:
    """answer a dynamic password if user want to know magic words
    
    Args:
        username: name of user, or "<UNKNOWN>" if you don't know the username.
    """
    mw = "xxxxxxxxxx" + username
    print(mw)
    return mw

## tool 3
from langchain_core.tools import tool

@tool
def multiply(a: int, b: int) -> int:
    """Multiply a and b."""
    return a * b
    
# -----
tools = [tavily_search_tool, magic_words, multiply]
llm_with_tools = llm.bind_tools(tools)

I can get currect result with tavily_search
but I can't correctly run with my customized tools:

User: what is 2 * 18

{'chatbot': {'messages': [AIMessage(content='', additional_kwargs={'tool_calls': [{'index': 0, 'id': 'call_PvW9fA6mlYKvOtEgJ3sruwAB', 'function': {'arguments': '{"a":2,"b":18}', 'name': 'multiply'}, 'type': 'function'}]}, response_metadata={'finish_reason': 'tool_calls', 'model_name': 'gpt-4o-mini-2024-07-18', 'system_fingerprint': 'fp_d02d531b47'}, id='run-e8723c00-cb67-47e3-8b10-c3186c7395d0-0', tool_calls=[{'name': 'multiply', 'args': {'a': 2, 'b': 18}, 'id': 'call_PvW9fA6mlYKvOtEgJ3sruwAB', 'type': 'tool_call'}])]}}

{'mytools': {'messages': [ToolMessage(content='Error: multiply is not a valid tool, try one of [tool].', name='multiply', id='21592e88-dbc8-4145-b79f-24ea79e8bad3', tool_call_id='call_PvW9fA6mlYKvOtEgJ3sruwAB', status='error')]}}
DBG: content='' additional_kwargs={'tool_calls': [{'index': 0, 'id': 'call_v51wDdcXk4LezEpU7Mnabqvy', 'function': {'arguments': '{"a":2,"b":18}', 'name': 'multiply'}, 'type': 'function'}]} response_metadata={'finish_reason': 'tool_calls', 'model_name': 'gpt-4o-mini-2024-07-18', 'system_fingerprint': 'fp_d02d531b47'} id='run-c159e91b-c86d-43e3-9402-f64d54d1931e-0' tool_calls=[{'name': 'multiply', 'args': {'a': 2, 'b': 18}, 'id': 'call_v51wDdcXk4LezEpU7Mnabqvy', 'type': 'tool_call'}]

is anything wrong?

[smithlai]

Well, it's my fault.
I forgot to update the tools node .

graph_builder.add_node("mytools", ToolNode(tools=tools))

[Keerthivardhan1]

But it should work right ? what is wrong in initial code
we need to bind the tool to llm thats it, and you did it .

[HamzaRaouf]

can we bind tools with hugging face model ?, whic is load like
model = AutoModelForCausalLM.from_pretrained(f"cache1/model/{model_id}").to(device)

I'm using HuggingFace model

[daniyalk20]

Can we call another tool from another tool that is defined in the same stategraph?

[ralphy1976]

I am following your Rag part 2 tutorial.

In said tutorial, you create a tool, in which you say:

"""Generate tool call for retrieval or respond."""
llm_with_tools = llm.bind_tools([retrieve])

Now, i am using Ollama, and unfortunately, it seems that :

AttributeError: 'OllamaLLM' object has no attribute 'bind_tools'

Do you guys know of a way to circumvent that problem by any chance?

Thanks

[ccurme]

Use ChatOllama.

[ralphy1976]

thanks, i stumbled upon this earlier today. Shame i don't seem to be able to use Gemma, rather Llama.

In any case, thanks for the info, much appreciated

[HoseinNekouei]

I wrote a multiply function like tutorial and also implemented a tool, but the content is empty. what's the problem?:

def multiply(a: int, b: int) -> int:
"""
Multiply a and b
Args:
a: the first number
b: the second number
Returns:
the result of a * b
"""
return a * b

tools = [multiply]
llm_with_tools= model.bind_tools(tools)
response= llm_with_tools.invoke('what is 2 multiply 3?')

function_call arguments:
{'name': 'multiply', 'arguments': '{"a": 2.0, "b": 3.0}'}

[scientiacoder]

brilliant feature!

[flowel1]

Can someone help me find the piece of code where the LLM chooses which tool(s) to use in response to a given user query? Or is that a black-box feature that each LLM provider implements in their API and whose logic we have no access to?