Scope Control
Runs on incoming messages
Stops the assistant from answering questions about topics you choose. For example, a maternal-health assistant will not give legal or financial advice.
What happens
⊖ Blocks — sends set reply — The message never reaches the AI assistant. The user gets the set reply you've written in the settings.
Example
| User | Is it okay to take this medicine? |
| Assistant replies | Sorry, I can't help with that topic. I can only answer questions related to this programme. |
These questions are stopped before the assistant answers them.
Your set reply is automatically translated by AI to match the language the user wrote in — a Hindi message gets the reply in Hindi, an English one in English.
Without this guardrail
The assistant answers questions it should stay away from. Risk: users may get wrong advice on serious topics like health or money.
How well it works
| Accuracy | Response time |
|---|---|
| 82% | 1–2 sec |
Tested on 355 real-style messages across education and healthcare topics, in English, Hinglish and Hindi.
Technical details
| Method | Accuracy | Precision | Recall | F1 | Sample size |
|---|---|---|---|---|---|
| LLM | 89% | 90% | 76% | 82% | 355 |
Dataset: Evaluation dataset (355 samples) ↗
Evaluated on 355 real-style messages across education and healthcare topics, in English, Hinglish and Hindi. Precision (90%) means false alarms are rare — about 1 in 10 blocked messages was actually in scope; recall (76%) means roughly 1 in 4 out-of-scope questions still gets through. F1 (82%) balances the two.