Prompting vs. Function Calling

[pietz]

To my understanding marvin will use the generic way of sending the data to the LLM. Since OpenAI also supports function calling, I was wondering if anybody collected experience on how they compare.

I build some fairly complex agents in langchain using their generic way of talking to the LLM (zero shot reAct). These agents have a high probability of failing and I assume this comes from the fact that normal LLMs aren't great at using functions to stick to programmatic instructions. My hypothesis is that function calling would improve that quite a bit, because suddenly we're using a tool that the model was trained to use.

Do you have any experience on that?

[pietz]

I think I misunderstood some internal functionality