I'm looking into ways to compress databases into ML models (I'm failing so far). The idea is to train the model on a dataset and teach it to answer questions ab

I'm looking into ways to compress databases into ML models (I'm failing so far). The idea is to train the model on a dataset and teach it to answer questions about the dataset (like SQL over table rows). I imagine that extracting individual data entries might work, but ensuring that the model doesn't just make up entries that didn't exist in the original dataset seems impossible.