[BUG] safety settings Gemini broken

[darthalex2014]

Description

Everything was fine yesterday, but today the Censorship (SAFETY) Level switch for Gemini models is completely broken on v2-dev. It doesn't seem to affect anything anymore. All of my prompts that used to work now trigger an error "[Gemini Issue]: Generation stopped due to SAFETY: HARM_CATEGORY_...". (I verified by running an old fork of v2-dev, and it's working fine there.)

Device and browser

Mobile/Desktop, last v2-dev

Screenshots and more

No response

Willingness to Contribute

• 🙋‍♂️ Yes, I would like to contribute a fix.

[darthalex2014]

Hello, we tried to solve the issue.

Failed to compile.
./src/modules/llms/server/gemini/gemini.router.ts
Error: x 'import', and 'export' cannot be used outside of module code
,-[/vercel/path0/src/modules/llms/server/gemini/gemini.router.ts:22:1]

:(

[darthalex2014]

I for nothing raised panic.
This is fault of model googleai-models/gemini-2.0-flash-exp.
On all other models such problem is not. Most likely Google something is muddying.

[darthalex2014]

https://discuss.ai.google.dev/t/flash-2-0-doesnt-respect-block-none-on-all-harm-categories/59352

found a fix instead of BLOCK_NONE we now must use OFF for Flash 2-0, but ONLY for Flash 2-0. if you send OFF with any other model you will get an error

[enricoros]

Wow, great find @darthalex2014 . What do you think could be a generic way of solving this in Big-AGI?

[enricoros]

You can now update to enjoy the "OFF" filtering. Will work on newer Gemini models, and may trip the older ones. I left both options in, and hopefully Gemini will converge quickly to one or another.