I like the idea, just that the examples are reproduced from the training data set.

How does it handle unknown queries?