Survey assistant asking random questions not in th...
# support
n
I have a specific set of survey questions.. for a nultilingual squad. The english assistant sticks to the script.. but hindi / hingligh assistant goes off onto a tangent asking totally different questions. How can I make the assistant folow the script??
t
This usually happens when the prompt isn’t strictly enforced across languages. I can help you fix it by tightening the system instructions, adding “no deviation from script” constraints, and implementing a structured survey flow (step-by-step or slot-based) with validation. Also, we can add language-specific guardrails so Hindi/Hinglish follows the same schema as English. Happy to take a look and fix it quickly. @Nisar
s
Hi, thanks for reaching out.
Could you please provide us with the call IDs?
That would help us investigate this further.
n
@Shaunak call ids: perfect response: 019e3eed-c472-722d-8ce2-2fcba788d9af deviant responses: 019e3ef2-2dfe-711e-a34a-98a67aca9d26 019e3f4f-b28c-7000-8e24-2c5fefd38834
s
Hi, Thanks for sharing those call IDs - that made it easy to trace the issue.
After looking into both assistants, the cause is a configuration mismatch rather than a platform bug. Your English assistant has the full survey script in the system prompt, while the Hindi/Hinglish assistant currently has a system prompt of just "You are an assistant." - so the LLM has no survey structure to follow and improvises as a generic assistant instead.
The fix is straightforward: open the Hindi assistant (01d1c1f8-88ed-468f-9ee6-1e3cfce968eb) in your Vapi dashboard, replace the system prompt with the same survey script from your English assistant, and adapt the language to Hindi/Hinglish while keeping the same structure - identity, Q1 through Q5, and closing. Once that's in place the Hindi assistant should follow the script just like the English one does.
As a longer term suggestion, you may want to consider consolidating into a single multilingual assistant with language-switching instructions in one prompt. This avoids configuration drift where one language gets updated but others are missed, which is exactly what happened here.