One of the Most powerful things with AIP Logic is the usage of Tools. For simple tasks the tools work fine, but when using multiple tools with complex logic, the LLM gets the thought process right, but makes mistakes with the syntax to use the tool.
In these cases there is a few back and forth between the model and the AIP.
I think if AIP logic can adopt the newer tool use/Function calling features that are available in the newer models and define tool use syntax in a structred way rather than free text, it might be helpful.
Hey Maddy,
Apologies for the delay in responding.
Often the an LLM is using the wrong syntax can come from a difference in the way a tool is referred to in the âUser providedâ prompts vs the generated prompt inserted by the LLM block, in particular function names are a common source of this, the LLM gets given something slightly different to the âdisplay nameâ shown in the platform.
You can diagnose if this is the case checking the raw text provided to the LLM, in the debugger when you hover the first LLM message you will see a âShow rawâ button.
I think if AIP logic can adopt the newer tool use/Function calling features that are available in the newer models and define tool use syntax in a structred way rather than free text, it might be helpful.
Logic is model agnostic and since different models have different capabilities we had to DIY our function calling syntax. For models that support it we are planning on special casing the implementation. This hasnât happened yet since we have some other exciting features we are working on.
1 Like