How do you adjudicate between joint end-to-end training...

Joint end-to-end training requires re-encoding the entire corpus every time the model is updated. If the corpus turns over 5% per week, the marginal cost of a model update jumps significantly (e.g., from 83 seconds to 27 minutes). Query-side-only training is a viable compromise that avoids the need for full index rebuilds. If document-side representation errors are suspected, perform a controlled ablation to compare the recall gains against the operational cost of the index rebuild.