Agent Prompter calls AI tool with wrong parameter names

[sommer_sun]

Hello,

I’ve created an AI tool that has two input variables: userQuery (string) and number-chunks (integer). I connected the tool to the “Agent Prompter” node using the “Workflow to Tool” node.

When I execute the Agent Prompter node, I get the following error message:

ERROR CloseablePythonNodeProxy agents._tool:Missing configuration parameters: userQuery-1951, number-chunks-1952

Can someone help with this?

[sommer_sun]

I have now found that the problem only occurs in connection with the Gemini LLM Selector. If I connect the input of the Agent Prompter to the OpenAI LLM Selector, the error does not occur.

[mlauber71]

@sommer_sun there already is a Ticket for this (AP-24608).

Agentic AI - local LLMs and Ollama - problems with names of parameters

I am experimenting with using local LLMs instead of just OpenAI with the new agentic nodes by KNIME. I use this workflow by @MartinDDDD as a blue print (also see his article). As one can see in the screenshot the model “quen” manages to figure out the exact parameter name (“selectedProfitCenter-9”) instead of using the given name in the configuration “selectedProfitCenter” of the tool. So somehow the internal ‘technical’ name seems to be expected when using local LLMs. BasicAgent-Part - Local …

[MartinDDDD]

@sommer_sun

I have now found that the problem only occurs in connection with the Gemini LLM Selector. If I connect the input of the Agent Prompter to the OpenAI LLM Selector, the error does not occur.

I know that @ActionAndi played around with agents as well - initially unsuccessfully - however after tweaking the default prompt that comes with Agent Prompter node I believe it worked out.

There was quite some discussion on this parameter topic in the thread @mlauber71 linked further up - quality I think definitely depends on which model is being used (and probably also which provider the model is from).

If I had to guess I’d say that this functionality was optimized for OpenAI models (which explains why e.g. @ActionAndi had to modify the system message to make it work with Gemini) and that model quality play a crucial role - e.g. I was not able to make a more complex agent work with smaller models that run local - primarily because at some stage the get the parameter names confused due to the number suffix… very similar to what you are reporting above.