Out-of-Scope Intent Handling
A useful concept I learned is Out-of-Scope (OOS) Intent Detection, sometimes called Out-of-Domain (OOD) Intent Handling.
Imagine a card-dispute assistant receives:
“What is the salary of employee X?”
The LLM may identify the candidate intent as employee_salary_inquiry, but the workflow engine should check whether that intent is supported.
User message
|
v
Extract candidate intent
|
v
Is intent supported by this workflow?
|
+--- No ---> Controlled out-of-scope response
|
+--- Yes --> Continue with validation and authorizationAn example controlled response is:
“I can help with card transactions and disputes, but I can't assist with that request here.”
Out-of-scope checks and access control are not the same thing. A salary inquiry might be out of scope for a dispute assistant. In an HR assistant, it might be in scope but still require role-based authorization. Both checks matter.
Can you explain it without the notes?
Out-of-Scope Intent Handling
Use your own words. Explain an example and one limitation before checking the explanation.
Check your explanation
A useful concept I learned is Out-of-Scope (OOS) Intent Detection, sometimes called Out-of-Domain (OOD) Intent Handling.
Turn this concept into a clear answer.
Practice this conceptAdapted from 19_deterministic_workflow_engine_learnings.md · “7. Out-of-Scope Intent Handling”. Examples are study fragments; framework behavior can depend on your version.