@hwchase17 Love this use case and agree that it’s tricky. It would be cool if you could create a dataset of the questions and csv files. Then others could experiment/contribute. I realize privacy is an issue, but even getting a realistic distribution over input queries would be useful.