How to implement content refresh functionality in RAG-based systems

I need help with building a presentation generator that uses Azure’s GPT-3.5-turbo-16k model. My users want a refresh button that creates different content when they don’t like the first result.

When someone clicks refresh, the system should create new content that’s more technical and uses bullet points instead of paragraphs. I’m working with two types of presentations:

  • Brief presentations: Maximum 4 slides for sales teams
  • Extended presentations: Up to 12 slides for executives

I already tried adjusting the temperature setting, but I’m wondering what other methods exist to make sure the regenerated content is actually different from the original output?

Here are some strategies to make your content refresh work better. Beyond tweaking temperature, try changing your system prompt each time you refresh. Ask for different formats or angles - this pushes the model to create truly different outputs. You can also add a “variation seed” to your prompt by requesting an alternate take or asking for a different analytical angle. Keep notes on what the model already generated and tell it to avoid those points. This prevents you from getting the same ideas with slightly different wording.