[Mbranconier]

I don't have a good answer to this, but I suspect that second approach with the API call would work well. Haven't done this in production to know. I do think any solution in this situation where you can't be 100% confident in the output would benefit from citations. Then if your feb 29th example does return the wrong data, the user would say "ah the model wasn't wrong, I was wrong because I was unclear".

[mambrogi]

Yea in my mind citations in this case would be the agent's reasoning and action trace. Like "I looked at all expenses in February and summed them."

But this adds another layer of complexity figuring out how to properly generate this user facing trace. This is something I like from PromptQL's approach.