Skip to main content

Guardrails

Configure safety features to protect your ZenSearch deployment from harmful inputs and outputs.

Overview​

Guardrails provide:

  • Input validation and filtering
  • Output safety checks
  • Content moderation
  • Hallucination detection

Accessing Guardrails​

  1. Click Settings in the sidebar
  2. Select the Guardrails tab
note

Guardrail configuration requires Admin or Owner role.

Input Guardrails​

Protect against harmful or inappropriate inputs.

Content Moderation​

SettingDescription
EnabledEnable content moderation
StrictnessLow, Medium, High

Filters for:

  • Hate speech
  • Violence
  • Adult content
  • Harassment

Prompt Injection Detection​

SettingDescription
EnabledDetect injection attempts
SensitivityDetection sensitivity

Protects against:

  • Jailbreak attempts
  • System prompt extraction
  • Role confusion attacks

PII Detection​

SettingDescription
EnabledDetect PII in queries
ActionWarn, Block, or Redact

Detects:

  • Email addresses
  • Phone numbers
  • Social security numbers
  • Credit card numbers

Length Validation​

SettingDescription
Max Query LengthMaximum characters
Max File SizeMaximum upload size

Output Guardrails​

Ensure AI responses are safe and accurate.

Hallucination Detection​

MethodDescription
LexicalWord overlap checking
SemanticMeaning comparison
HybridCombined approach

Settings:

  • Detection threshold
  • Action on detection (Warn/Block)

Toxicity Filtering​

SettingDescription
EnabledFilter toxic outputs
ThresholdToxicity score threshold

Relevance Checking​

SettingDescription
EnabledCheck response relevance
ThresholdMinimum relevance score

Ensures responses address the query.

Configuration​

Enable/Disable Guardrails​

Toggle individual guardrails:

┌─────────────────────────────────────┐
│ Content Moderation      [ON]  [OFF] │
│ Prompt Injection        [ON]  [OFF] │
│ PII Detection           [ON]  [OFF] │
│ Hallucination Detection [ON]  [OFF] │
│ Toxicity Filtering      [ON]  [OFF] │
└─────────────────────────────────────┘

Adjust Sensitivity​

Each guardrail has sensitivity settings:

LevelBehavior
LowMinimal filtering
MediumBalanced approach
HighStrict filtering

Set Actions​

Configure what happens when triggered:

ActionResult
AllowLog but allow
WarnShow warning to user
BlockPrevent the request
RedactRemove sensitive content

Monitoring​

Guardrail Events​

View triggered guardrails:

  • Event timestamp
  • Guardrail type
  • Trigger reason
  • Action taken

Analytics​

Track guardrail metrics:

  • Trigger frequency
  • False positive rate
  • Most common violations

Best Practices​

Configuration​

  1. Start with default settings
  2. Monitor false positives
  3. Adjust sensitivity gradually
  4. Review events regularly

Security​

  1. Enable all input guardrails
  2. Use hallucination detection
  3. Monitor for new attack patterns
  4. Keep guardrails updated

User Experience​

  1. Balance security with usability
  2. Provide helpful error messages
  3. Allow appeals for false positives
  4. Document guardrail policies

Troubleshooting​

Too Many False Positives​

  1. Lower sensitivity settings
  2. Review trigger patterns
  3. Add exceptions if needed
  4. Contact support for tuning

Guardrails Not Triggering​

  1. Verify guardrails are enabled
  2. Check sensitivity settings
  3. Test with known bad inputs
  4. Review configuration

Next Steps​