Introduction

Guardrails is a powerful content filtering system in Fuel iX Copilots that helps organizations maintain control over user interactions with AI assistants. This feature is exclusively available to organization administrators and must be explicitly enabled for your organization.

🚧

All guardrail features are disabled by default and must be activated by an organization administrator.

Key Benefits

• Prevent sharing of sensitive information
• Maintain professional communication standards
• Protect organization security and compliance
• Enhance AI interaction quality
• Reduce risk of inappropriate content

Administrator Access

Requirements

• Must have administrator role in your organization
• Access to organization settings
• Authorization to modify security settings

Accessing Guardrails Settings

1. Log in to your Fuel iX Copilots account with administrator privileges
2. Navigate to Settings in the main navigation
3. Select "Guardrails" from the settings menu
4. The Guardrails configuration page will display all available options

Types of Guardrails

Basic

Ban Substrings

The most configurable guardrail type, allowing custom string definitions.

Configuration Options:

• Enable/Disable toggle
• Custom string entries (unlimited number)
• Each string limited to 40 words

Default Settings:

• Match Type: Word-based matching
• Case Sensitivity: Disabled (case-insensitive)
• Redaction: Disabled
• Contains All: Disabled (matches any substring)

❗️

Adding more strings will increase processing time and may impact user experience through increased latency. Consider this trade-off when configuring ban substrings.

Advanced

Secrets

Examines user inputs and prevents the exposure of sensitive information.

Types of Secrets:

• API tokens
• Private Keys
• Passwords
• And more

Invisible Text

Detects and removes non-printable, invisible Unicode characters from the input.

Regex Scanner

Evaluates and blocks the prompts matching specific regular expression patterns predefined by the admin.

Premium

Premium guardrails leverage GPU-accelerated machine learning models to provide advanced content protection. These features require additional infrastructure and are available as an add-on service.

Availability: Contact your Fuel iX account representative or [email protected] to enable premium guardrails for your organization.

Performance: Premium guardrails utilize GPU processing to maintain fast response times despite the computational complexity of ML-based analysis.

1. Prompt Injection

Detects and blocks attempts to manipulate or override the copilot's system prompt through user inputs.

Configuration: Simple enable/disable toggle with preset detection models.

How it works: Uses machine learning to identify patterns commonly associated with prompt injection attacks, protecting your copilot's intended behavior and guardrails.

2. Toxicity Detection

Evaluates and blocks prompts containing toxic, offensive, or inappropriate content.

Configuration: Simple enable/disable toggle with preset configurations.

Current Default Settings:

• Threshold: 0.75
• Match Type: Sentence

How it works: ML models analyze user inputs at the sentence level to detect potentially harmful or inappropriate content before it reaches the AI model.

3. Anonymization

Automatically detects and redacts personally identifiable information (PII) from user inputs.

Configuration: Simple enable/disable toggle with preset PII detection types.

Current Protected Information Types:

• Credit Card Numbers
• Cryptocurrency Addresses
• IBAN Codes
• IP Addresses
• US Social Security Numbers
• US Bank Account Numbers
• UUIDs

How it works: Pattern recognition and ML models identify sensitive information in user inputs and automatically redact it before processing.

🚧

Future updates will allow organization-specific configuration of PII types. Currently, these defaults cannot be modified.

Important Considerations

Model Inherent Guardrails

👍

Each AI model comes with its own built-in guardrails. These may interact with or conflict with your organization's guardrail settings. Always test your configurations to ensure they work as expected with your chosen models.

Multi-Language Support

Language Considerations

[!IMPORTANT]
Fuel iX Copilots supports both English and French interfaces, requiring careful consideration when configuring guardrails to ensure a consistent experience across both languages.

Ban Substrings in Multiple Languages

Best Practices for Bilingual Support:

1. Create Parallel Lists
   • Include both English and French versions of each banned term
   • Consider common variations in both languages
   • Account for regional variations (e.g., Canadian French)

Example Configuration:

Language-Specific Considerations:

• Accented Characters

  • Ban substrings are case-insensitive but preserve accent matching
  • "é" will match both "é" and "e" in French terms
  • Consider including both accented and non-accented versions for critical terms

• Word Order

  • French and English may use different word orders
  • Consider phrase variations in both languages
  • Test matching behavior for compound terms

Language Impact on Guardrails

Ban Substrings:

• Requires manual configuration for both languages
• Consider maintaining a bilingual reference document
• Test effectiveness in both language contexts

Recommended Workflow for Bilingual Organizations

1. Initial Setup

   • Create a spreadsheet mapping terms across languages
   • Include common variations and misspellings
   • Document regional considerations

2. Testing Process

   • Test configurations with both English and French inputs
   • Verify detection accuracy in both languages
   • Confirm error messages display in the user's selected language

3. Maintenance

   • Regular review of terms in both languages
   • Update both language versions simultaneously
   • Document any language-specific exceptions

👍

Consider creating a bilingual review team to ensure guardrail effectiveness across both languages.

Content Processing Scope

Guardrails apply differently to various content types:

System Prompts

• Guardrails are applied to all user inputs
• System prompts and model configurations are exempt from guardrail checks
• Administrative prompts and system configurations bypass guardrail restrictions

Image Processing

• Text contained within uploaded images is not processed by guardrails
• OCR (Optical Character Recognition) is not performed on images
• Users should exercise caution when uploading images containing sensitive information

Best Practices

Configuration Recommendations

1. Start with minimal ban substrings

   • Begin with most critical terms
   • Monitor performance impact
   • Gradually add more strings as needed

2. Testing

   • Test new configurations in a controlled environment
   • Verify impact on response times
   • Check for conflicts with model guardrails

Monitoring and Maintenance

1. Regular Review

   • Monitor system performance
   • Review blocked content patterns
   • Adjust configurations as needed

2. Performance Optimization

   • Regularly clean up unused ban substrings
   • Consolidate similar terms
   • Document performance baselines

Troubleshooting

Common Issues

1. Unexpected Blocking

   • Check for partial word matches
   • Review case sensitivity impact
   • Verify model guardrail interactions

2. Performance Degradation

   • Review number of ban substrings
   • Check string length and complexity
   • Monitor system logs for bottlenecks