Preventing a tool from being used

[cadr]

Hello,

I wrote a simple tool to look up ham radio users by their call sign. It works really well. However, the llm wants to use it in the inverse case as well (looking up the call sign of a ham radio user). I could enhance the tool to do both, but I am interested in how to prevent a tool from being used except in the case where it can be used as intended.

My tool's current description is

"Use this tool when you need to look up the name of ham radio operator by their call sign."

I have tried variations to try to tell it when not to use the tool, or to only use it when it has a call sign, etc. But none seem to be doing the trick.

Do you have any suggestions around how to craft the description to get what I want here?
Alternatively, do you think langchain needs to be enhanced to better handle the negative case?

Thank you for your help!

[Hamas-ur-Rehman]

You could try adding negative words as well right this tool prompt like

this is not able to do this .... and using this tool for other purpose will fail and you will not be able to give the user the correct answer

[cadr]

Thanks for the response!

Yes, I've tried a lot of variations of negative phrases, and haven't been able to get any to work. I started experimenting with changing the prompt that langchain itself uses (the one in prompt.py), but haven't had any luck there, either.

[adriens]

I had not thought about this use case ie. "negative prompt" : thanks a lot for sharing this inspiring pattern 🙏

[willsmithorg]

Maybe make a structured tool and make it clear that the input string is a callsign?
https://python.langchain.com/en/latest/modules/agents/tools/custom_tools.html

Or maybe your description of the tool should be clearer about inputs and outputs:
""Use this tool when you need to look up the name of ham radio operator by their call sign. The input is the call sign. The output of the tool will be the ham radio operator name."

Or if the mapping is 1:1, maybe create the opposite tool as well and describe as above, so the AI has a clearer way to solve the inverse mapping, i.e from name to call sign.

[adriens]

Making description pattern as "Input is.... Output is..." seems to be an efficient pattern. BTW, how can we get the score on how the LLM choses the tool (trust rate) so we could fine-tune description ?

[cadr]

Thanks for the suggestion - I can look into structured tools.
I actually have tried many variations of creating the opposite tool and talking about the input/output. All variations seem to want to choose this tool regardless of input. I started going down the road of varying the prompt that langchain itself imbeds this in, but haven't gotten very far in it.

And, again, even if I get my exact case working after a lot of trial and error around the prompts, this seems like a larger concept involving having more fine-tuned control in helping the llm chose the tools.

[adriens]

Please keep posting your experiments, I find them really useful @cadr 🙏