I have like super detailed flow descriptions that are a couple of lines themselves. I even added examples. My main issue is that I want this to be working with smaller LLMs. When using GPT-4o I got no issues. But something like phi-3 or gemma2 and it really struggles when the user input is a bit ambiguous. But when I tested with a bunch of examples I could think up, it always ended in pattern_classification. So my idea was to have a detailed set of instructions for the LLM to rephrase the user input, and then use that to identify which flow.
I could see how to implement this when using APIs, but any solution that can be used with inspect and shell would be much appreciated