If you use “공유나” (in Korean) which means “sharing or” in the context or prompt, Gemini api does not respond. Also “이지은” which is a person name it also does not response. “공유나” can a person’s name in Korean, too. This is very annoying and this combination can happen any Korean context. We’ve tried this with all the Gemini models, We’ve spent a day to find out the cause.
it shoud be fixed cause those words can be included and prompt. Just copy the word and send it to api then it stop respond.
1 Like
Welcome to the dev forum.
I can confirm that I was able to reproduce this gimmick.
I believe that it’s likely because of some prohibited-content flagging happening under the hood.
For this test, I disabled the safety settings and yet it was blocked.
@keertk-google might find this interesting.
This seems to be a bug, I have had an error before when using AI where I word in Russian or Spanish means one thing, but in English it means another thing that is not allowed and it gets blocked, but what I am talking about doesn’t seem to be the case.
Thx for the check!
It’s not just those two words i mentioned above.
There are more but I couldn’t figure out which words causing the same issue cause we are using the api for document loading and parsing which i have to find them from massive amount of text.
FYI, It works fine with Gemini Service.
1 Like
It seems that the AI Studio and API’s content filter become oversensitive when there is a name in the prompt. Almost every time Gemini (Studio/API) stops in the middle and there is a human name in the prompt, I can replace it with a fake name like ‘AFakeName’ and it gives me the result. The aggressive content filter makes it almost impossible for me to use Gemini, though.