Feels like you are still stuck in using completion as the engine to retrieve the suggestion instead of database. Am I missing something in your use case? Do you have the suggestions data ready or do you expect the AI to generate you a suggestion from scratch?
I was reading again my messages and I think i was not clear enough to you all about this point: I do not have a pre-defined number of categories. It means that if a customer likes ‘Hiking’ and ‘Restaurant’, I’d be glad to the model suggest something like ‘Healthy restaurants to go after your hiking activities’. This would be a new topic that was not designed at first. If the customer likes it, i will also be glad to use ‘Healthy restaurants to go after your hiking activities’ as a new input
If the second, how do you plan to overcome the hallucination issue?
I didn’t think about this, to be honest. Maybe feedback might help to show the model about absurd topics and avoid this situation. For example, a customer loves ‘Hiking’ and ‘Museum’, then it comes a absurd recomendations like ‘Hiking spots inside museums’. Customer will tag it with a dislike, then I expect it not to come through this theme again. How does it sound to you?