That use-case seems inefficient to solve like that in the long run as well, like if you really would have to use a million tokens to do every small query you require on your data it would be prohibitively costly except doing as an experiment.