Predictions of recommendation systems (TwoTowers, NCF, MLP, DLRM, DCN)

How do we actually retrieve the recommendations once the model trained ?
It does not appear in any notebook.

Most of the recommendation systems models serves as ranking purpose. The prediction value, typically click-through-rate, is used as ranking score. So normally, there would exist an serving program which would utilize prediction value to sort candidates. When user request arrives, serving program call backend predicting service - trained model is applied - to get those sorting score. Of course, there is much more things happened underneaths, eg. how trained models are updated in model repository used by predicting service, etc.