docs/how_to/tool_results_pass_to_model/

[giscus[bot]]

docs/how_to/tool_results_pass_to_model/

This guide assumes familiarity with the following concepts:

https://python.langchain.com/docs/how_to/tool_results_pass_to_model/

[jameslmoser]

I am very new to using LangChain (and generative AI in general). I was able to follow this and get it working in a Jupyter notebook. I see it calling the tools (add, multiply) while debugging.

Out of curiosity I modified the functions to return the wrong values. Those values are in the ToolMessage content, however the final AIMessage contains the correct values. The model appears to be discarding the values and doing the calculations on its own.

Is this expected behaviour? A result of the contrived examples? If you wanted to, how would you go about ensuring that the values of your tools would be used? Would that be something you would specify in a prompt message or tool description?

[RovaHerve]

Hello Jameslmoser,
Yes, you are right. I also test like your comment and the model answer in its own probabilty if the return type is not good for as the model.
I found a method to break it, modify the tool (a function that you defined) to a good prompt and the model answer you the return type as you wish.
@tool
def my_tool(args):
"""I make here a good prompt to jealbreak model response"""
I hope this answer help you and other person.

[YMC2ATiberius]

Hi, I ran into the same problem.
In definition of tools, I deliberately swap the definiton of add and multi as this:
@tool
def add(a: int, b:int)->int:
"""Multiply a and b. Stick to this result."""
return a * b

@tool
def multiply(a: int, b: int) -> int:
"""Add a and b.Stick to this result."""
return a + b

The query I use is "What is 312? Also, what is 11 + 49? By the way, what is 66? Please use tools provided to get the result!"

Then I print out the tool_msg, which matches what I want:
ToolMessage(content='15', name='multiply', tool_call_id='call_uzftejK0vrbGQBnF4XGQ4UUV'), ToolMessage(content='539', name='add', tool_call_id='call_0j6c1mJLTrG1QEhhUpFQDOCq'), ToolMessage(content='12', name='multiply', tool_call_id='call_u7Ru6mhrB5zJgGI9Dzrm0rGZ')

However, LLM (I use gpt-4.1-mini) discart the result of my tools and gives the final response content like is:

• 3 * 12 = 36
• 11 + 49 = 60
• 6 * 6 = 36

[HelloWorldLTY]

It seems that the current answer is wrong:

AIMessage(content='36 * 12 = 432\n\n11 + 49 = 60\n\nThe answer is: 432 and 60', additional_kwargs={}, response_metadata={'token_usage': {'completion_tokens': 26, 'prompt_tokens': 1175, 'total_tokens': 1201, 'completion_time': 0.021666667, 'prompt_time': 0.174865897, 'queue_time': 0.19576452900000002, 'total_time': 0.196532564}, 'model_name': 'llama3-8b-8192', 'system_fingerprint': 'fp_6a6771ae9c', 'finish_reason': 'stop', 'logprobs': None}, id='run-4a84ee76-0f64-4fb2-915f-6f59113ab17a-0', usage_metadata={'input_tokens': 1175, 'output_tokens': 26, 'total_tokens': 1201})

[RovaHerve]

what is your issues, I don't understand your answer ?

[plawhatre]

I wanted to test if this would work with compound operations

system_message = SystemMessage("You are a helpful assistant who helps with any question asked of you.")
human_message = HumanMessage("Multiply 5.2 by 63, then add 23 to the result.?")
messages = [system_message, human_message]

and I got the following which is wrong

Tool called
Processing tool: multiply, 
{'name': 'multiply', 
'args': {'a': 5.2, 'b': 63}, 'id': '66396fc8-7408-4283-8b1b-e7d265020cf1', 
'type': 'tool_call'}

Processing tool: addition, 
{'name': 'addition', 
'args': {'a': 321.6, 'b': 23}, 'id': '2860be65-8304-44d2-aaba-78292396ebf3', 
'type': 'tool_call'}

final_output content: 327.6 + 23 = 344.6 (The multiplication of 5.2 and 63 results in 327.6, then adding 23 to the result gives us 344.6.)

Correct Answer = 5.2 * 63 + 23 = 327.6 + 23 = 350.6

[RovaHerve]

for your example, I see that you are not made the tools function , that's right ??
You should play the correct answers for addition and multiplication and guide the model to the right answers.