Which is the correct and honest answer. LLMs, broadly speaking, cannot provide confidence levels for anything in the way statistical models can. Asking for a confidence level from the model here is a "vibe check" at best. If a vibe check is still useful to you as a fuzzy, non-deterministic confidence level, you could try adding something like "If you don't have over X% confidence in the identified field value, append '[Low Confidence]' to the return value.". Gives you something to search / trigger workflows on.

You can get into approaches involving "LLM-as-judge" which have a second model score responses from the first based on pre-defined grading rubrics and such, but that's slow and expensive and you need good grading rubrics for every scenario. Doesn't really make sense for something general purpose by design like Smart Fields.