Gemini-2.5-flash vs Gemini-2.5-lite preview output discrepancy in ADK WEB

[pixjobs]

I'm using Secret Manager to set my Gemini version so I can hot swap in adk web in Workbench.

I've noticed that when I use the new Gemini-2.5-flash, the output seems to be fine, but when I try Gemini-2.5-flash-lite preview (excuse my numbering) the LLM seems to lose fidelity.

It will retrieve some data and then just say that it's already displayed the data (in my case job listings).

when I switch back to Gemini-2.5-pro or Gemini 2.5-flash this gets reverted.

[DeanChensj]

Hi @pixjobs , This seems to be issue related to the model quality that is out of ADK framework's control.

[pixjobs]

@DeanChensj cool, thank you, it is a preview model after all. Some models also seem to be less forgiving with complex instructions, e.g. when using adk web the response might be ok in the console but adk web doesn't render the information because it isn't wrapped in a certain way.

[hangfei]

If it's related to adk-web issue? Could you share what's the text diffs?

cc: @wyf7107

[Frozenace]

Sorry I didn’t get time to test it in detail but a direct swap of my project (workmatch) model with lite caused the thing to not display anything in the adk web even though the console returned a result.

My repository has the instruction to replicate the project on GCP with the secrets and the prompts.

I’m a bit burnt out as I coded this thing on top of my day job…

Maybe I can dig up the langfuse traces over the weekend. Thanks for looking.