Keeping Chatbots On Topic

“Stay on topic” is not a single model setting. A useful scope statement identifies the subject matter, sources the assistant may use, tasks it can perform, and cases that require a refusal or handoff. A prompt can communicate those boundaries, but it cannot ensure that every response follows them. The application should also limit its retrieval corpus, label source material, and constrain available tools to the work the bot is authorized to do. If the answer is not supported by the allowed material, the system should say so or route the user onward rather than fill the gap with a plausible guess. Controls belong at multiple points. Input checks can identify requests outside the service’s remit. Retrieval checks can reject irrelevant or untrusted chunks before they enter context. Conversation-flow rules can keep a multi-turn interaction within an approved process. Tool controls should validate each action and its arguments. Output checks can flag an answer that contradicts the stated scope or lacks required support. NVIDIA NeMo Guardrails documents these as separate input, retrieval, dialog, execution, and output rail types; the same design principle can be implemented with other frameworks or ordinary application code. These controls can conflict with usefulness: a narrow classifier may reject legitimate edge cases, while a broad scope can invite answers the organization cannot support. Design an explicit fallback, expose the relevant source or limitation, and route sensitive decisions to people with authority. Evaluate with ordinary in-scope questions, borderline cases, unsupported questions, adversarial attempts, and service outages. Track both inappropriate responses and unnecessary refusals. Topic controls reduce the range of likely failures, but they do not establish that an answer is factually correct or that a tool action is authorized.