I’ve been testing an assistant for a bilingual company.
The instructions were in french and we instructed the assistant to answer in the language spoken, then emphasized on it… But the assistant continued to respond in french unless explicitly prompted “english please” type of things.
It got to the point where we changed the instruction in english on each run. To our surprise, no effect. We confirm the asssistant was getting the instruction (also by explicitly mentioning a given number in it’s reply).
We then changed the assistant instruction thru an update. We did that many times before. So now, everything the assistant ever saw should have been in English but it still responded in French.
This leads us to conclude that other maybe non-disclosed issues are present. Is the assistant fine-tuned in the background on past conversation??? what’s going on with this
Hi - can you share details of the instructions you provided? That would help forming a better view of the potential issues.
You are an agent for Sol-R employees.
Each thread will provide you with some information about the employee, including his name and position.
Your goal is to make them enthusiastic about you.
Keep your replies very short.
Language:
If an employee addresses you in a language, always respond in the same language (this is very important).
Also, if you converse in French, use Quebecois French.
Email:
If you are asked to write an email, keep it short. Never add a signature. Put it in a markdown code block.
If you are asked to show an email, if possible respond with the content in HTML.
Finally, it’s VERY IMPORTANT to always respond in the same language as the employee.
Here is a brand new thread. 1 message. Bam in French. How can that be? It’s certain that with that name you want to speak French?
As a few general pointers, what seems to be helping is to be as specific as you can in your instructions. So for instance here, it seems that the user name triggers the assistant to think that the user is French. You could specify that it should focus solely on the language in which the user communicates and not make the choice of the response language dependent on other user characteristics such as the name. You could also add an example in the instruction, i.e. if user uses language X, then assistant should respond in language X.
So in your case, maybe rephrasing the instructions specific to language to the following might help:
You should always communicate with the employee in the same language in which the employee communicates with you. For example, if the user starts a conversation in French, your responses should be consistently in French, unless the user changes to a different language during the course of the conversation. Never determine the choice of language based on other user characteristics, such as the user’s origin of name. If you converse in French, always use Quebecois French.
No guarantee that this will work but sometimes it is about trial and error and my experience has been that small changes in instructions can have a material impact on the assistant’s performance.
Ok. In the first reply message I sent. THIS is the full instruction (about 600 charaters). In it, we have this: “If an employee addresses you in a language, always respond in the same language (this is very important).” and at the end we have this: “Finally, it’s VERY IMPORTANT to always respond in the same language as the employee.” My picture was a new thread of these very instructions. It was a short message: “Hi, I am Jean-Sebastien and I am a President at Sol-R and my email is [my email]\n Jean-Sebastien has already had 72 conversations with talkToGPT”. … and it STILL answered in French. (yes, there was some detail in our instrution that IF it was conversing in French, to use the local version)
Yes, I saw the full instructions only at second sight - sorry for this oversight. I have now adjusted my response to accommodate for this.