Hello everyone,
I’m working on building a bot using the real-time API, and I’ve encountered a couple of challenges when it comes to maintaining a strict flow and ensuring compliance with predefined instructions. I’d like some guidance or suggestions on how to address these issues.
The Context:
The bot is designed to follow a strict step-by-step conversation flow to reach a specific output. The flow is crucial for the bot to function properly and deliver accurate results. In the system prompt, I clearly define these rules:
The bot must adhere to a structured flow and not skip any steps.
Certain types of questions or interactions should be explicitly ignored or rejected.
The Challenges:
Flow Deviation:
When the conversation proceeds as intended, the bot follows the step-by-step flow correctly. However, if the conversation is tweaked (e.g., the user tries unconventional phrasing or jumps between steps), the bot sometimes skips steps or fails to maintain the flow, even though the system prompt strictly emphasizes adhering to the defined path.
Prompt Breaking:
Despite explicitly instructing the bot not to answer certain types of questions, a user can still exploit the conversation to bypass these restrictions. For example, reframing or indirectly approaching a restricted topic may lead to unintended answers.
My Goal:
I’m looking for a way to ensure that:
The bot always follows the strict step-by-step conversation flow, regardless of how the conversation evolves.
The bot reliably resists prompt-breaking techniques and strictly adheres to predefined restrictions.
What I’ve Tried:
Crafting detailed and specific system prompts to emphasize the importance of the flow and restrictions.
Experimenting with rephrasing rules and conditions in the system prompt.
However, these efforts have not fully resolved the issues, especially when a determined user tries to exploit the conversation.