Gemini API Quiz App: Issue with "Dummy" Report Generation instead of analyzing collected leads

What is the best way to ensure the model “anchors” its report on the specific data provided? Should I be using Structured Outputs (JSON) to feed the data back, or is there a specific System Instruction that prevents the model from defaulting to “general knowledge” patterns?

Also, if I want to store these leads in a real database (like Firebase or SQL) for long-term reporting, what is the recommended architecture for a Gemini-powered quiz app?

Thanks for any insights!