Overview
Guardrails let Operators define rules that inspect AI traffic flowing through a gateway. When a guardrail is triggered, the AI request is blocked before it reaches the model. Guardrails are useful to prevent sensitive or non-compliant data (such as national IDs, bank accounts, or credit cards) from being sent to the AI, helping you enforce data protection and compliance policies. Each guardrail is:- Assigned to one or more gateways, so it only applies to the traffic of those gateways.
- Evaluated on every AI request handled by the assigned gateways.
- Blocking: when the guardrail matches, the request is rejected.
For now, guardrails support regular expressions (regex) only. Additional guardrail types may be added in the future.
How guardrails work
A guardrail contains one or more regex patterns. On each AI request routed through an assigned gateway:- The request content is checked against the guardrail’s regex patterns.
- If any pattern matches, the guardrail is triggered.
- The AI request is blocked and does not reach the model.
Create a guardrail
To create a guardrail:- Open Guardrails in the Platform Backoffice.
- Click Add guardrail.
- Enter a name and an optional description.
- Select the gateways this guardrail should apply to.
- Add one or more regex patterns. Use Add regex to define multiple patterns in the same guardrail.
- Save the guardrail.
Manage guardrails
From the guardrails list you can review each guardrail’s name, description, type, effect, and the gateways it is assigned to.- Edit a guardrail to update its patterns or reassign it to different gateways.
- Delete a guardrail to stop it from blocking requests.