Generally, fine-tuning is to teach the model to reply in a specific way. Fine-tuning won’t be sufficient to teach the model “new content”.
That aside, what you’ve described for the fine-tuning examples, and the retrieval approach some sensible.
You could prototype with the assistants api first, though you won’t be able to use your finetuned model.