Hi everyone,
I’m currently building a small assistant using the Gemini API that helps UAE users with common banking queries, specifically around FAB balance inquiry steps through the mobile app and online banking. The goal is to let users ask natural questions like “How do I check my FAB balance without logging in?” or “Why is my available balance different from ledger balance?” and return clear step by step guidance.
The issue I’m facing is inconsistent responses when the same prompt is sent multiple times. For example, when I ask about checking balance via SMS banking in the UAE context, sometimes the model gives accurate steps relevant to First Abu Dhabi Bank, and other times it generates generic advice about US banks or even suggests calling numbers that are not related to FAB. I am already providing system instructions that specify UAE banking context and FAB specifically, but the output still drifts.
I also noticed that when users phrase the question slightly differently, like asking about blocked amounts or pending card transactions affecting available balance, the model sometimes hallucinates internal banking policies instead of staying at a general guidance level. For a financial assistant, this is risky because incorrect instructions can confuse users who are already worried about missing funds.
Has anyone dealt with maintaining strict regional and brand specific grounding when using Gemini for financial guidance bots? Would you recommend stronger system prompts, function calling with predefined responses, or grounding with external data through retrieval?
I would really appreciate suggestions on how to make the responses more stable and fact aligned for a UAE banking use case.