Hi,
I am using a Prompt that allows me to identify themes (financial domain) from the underlying text data that I provide. The text is usually several paragraphs long.
In my observation when I used Gemini 2.5 Pro, the categories that came back were much more meaningful and limited in number and made more sense. When I explored using the Gemini 3 Pro Preiview, the exact same prompt resulted in almost a 5-7x more number of distinct categories.
Is there an explanation to this difference in behaviour?
Hi @Harshad_Chavan10
It could be because Gemini 3 model possess superior reasoning and a higher sensitivity to nuance, achieving the more constrained output style of Gemini 2.5 requires tightening your instructions. Specifically, implement constraints like: 'Provide only the top 3-5 primary themes that encapsulate the main narrative of the text. Do not include any supporting or less important sub-themes’
Please take a look at this https://ai.google.dev/gemini-api/docs/gemini-3#prompting_best_practices
Hey @Pannaga_J . Thanks a lot for these insights. This makes a lot of sense. Will have a look at it.