Find answers to common questions about configuring, training, and troubleshooting input guardrails in Optimizely Opal. For step-by-step instructions, see Create and train an input guardrail.
Do input guardrails filter Opal's responses?
No. Input guardrails screen the requests users send to Opal, not the responses Opal returns. To evaluate or control Opal's output, use output evaluation or execution guardrails.
Can more than one guardrail apply at the same time?
Only one guardrail is active at the instance level at a time. Per-agent guardrails operate independently, so a specialized agent enforces its own guardrail whether or not an instance-level guardrail is active.
Why did my guardrail training fail?
Training fails when your rule decomposes into more than five atomic policies. Narrow or simplify the rule text, then train again.
Training also fails after repeated transient errors. Opal retries a job up to three times across 15 minutes, then marks the version Failed. When a version shows Failed, train it again.
How long does training take?
Training runs in the background and takes several minutes. Opal runs one training job per instance at a time, so a later job waits for the current one to finish.
How do I fix a guardrail that blocks too many messages?
Lower the sensitivity for the policy that is over-blocking. Alternatively, review the training data and correct or remove mislabeled examples.
Sensitivity changes take effect immediately when you click Save changes. Training-data changes require a retrain.
How do I fix a guardrail that misses messages it should block?
Raise the sensitivity for the relevant policy, or add training examples that show the messages you want blocked, then retrain. Test the guardrail with edge cases to confirm the change.
My guardrail is not enforcing. What should I check?
Confirm the guardrail is in effect. An instance-level guardrail must be selected in Settings > Chat. A per-agent guardrail must be assigned to the agent.
Confirm the version status is Ready, because only trained versions enforce. Guardrails also fail open, so classifier errors let messages through rather than block them.
What happens when a classification error occurs?
Guardrails fail open by design. When Opal cannot classify a message, it allows the message through rather than blocking the user. A classification error never prevents someone from using Opal.
What happens when I edit a rule or a training example?
Editing a rule, block message, or training example creates a new untrained draft version. Retrain the guardrail for your changes to take effect. Sensitivity changes are the exception, they apply immediately without retraining.
What happens when I delete a guardrail?
Deleting a guardrail removes it and all of its versions, and you cannot restore it from version history. Enforcement stops immediately. When an instance or a specialized agent still references the deleted guardrail, Opal no longer screens that input, and the requests pass through unscreened. To keep enforcement in place, assign a different guardrail.
See Delete a guardrail.
Article is closed for comments.