My assistant was doing well, but out of nowhere, it started to gives replies in a different input language.
It does it for a single reply and then goes back to speaking english (or the chat language).
It is an intermitten issue, maybe affecting 1 chat for every 20 thread, but serious enough to affect user experience.
I tried putting instruction such as “Reply only in English, unless requested by the user”. But despite the user clear use of english, once in a while, it went off course and speak another language.
The model is so uncertain of what to write (high perplexity), it can’t even predict a token set in the right language.
That there are no sampling parameters in assistants to constrain a bad token from steering the generation of language off-course makes it that much worse.
gpt-3.5-turbo-0613 is better if you’d prefer lower cost and lower damage done to the AI’s brain (just incompatible with retrieval).
Thanks. Will look into that. For my use case, it seems that gpt-3.5 was having trouble to follow the instructions closely. Gpt-4 is much better but this unexplained language switching is driving me crazy.