Guardrails control what your AI models can and cannot do. Set them up to prevent unwanted outputs and enforce your rules.
How guardrails work
Company guardrail – Applied to all models by default, unless a custom or excluded guardrail is set.
Custom guardrail – Applies to specific models and has higher priority than the company guardrail.
Excluded models – Models in the exclusion list are not affected by any guardrail.
Priority order: Excluded → Custom → Company
Select from the validators library based on your security requirements. Click here for more info on Regex. Click here for more info on No PII.
Controls content based on pattern matching. Configure as follows:
Name
Assign a descriptive identifier for the rule
When Regex matches
Search: looks for pattern existence in content
Extraction: identifies and extracts specific pattern matches
Regex rule
Define the pattern to match (e.g., credit card numbers, internal codes, sensitive identifiers)
Trigger on
Input: validates user prompts and requests Output: validates AI model responses Both: validates both input and output
Guardrail behavior
Fix: automatically removes or replaces sensitive data Filter: removes sensitive content with generic replacement Exception: blocks the request entirely and returns error
Detects and anonymizes Personally Identifiable Information. Configuration includes:
Name
Assign identifier for the PII protection rule
Entities
Select specific PII types to detect (credit card numbers, crypto addresses, etc)
Trigger on
Choose Input, Output, or Both
Guardrail behavior
Fix: Replaces PII with generic placeholders
Exception: Blocks content containing PII
Last updated
