Guardrails Configuration

Learn how to configure and apply Guardrails to protect your agents with intelligent content filtering. This guide covers everything from understanding the default guardrail to creating and managing custom guardrails.

Prerequisites

Before configuring Guardrails, ensure you have:

Agent Created: You need an existing agent to apply guardrails to

Admin Access: Only project Admins and Owners can configure guardrails

Requirements Identified: Know what content needs protection

Guardrails are managed in the Guardrails section of the agent dashboard.

Understanding the Default Guardrail

Every agent in PLai Framework comes with a default INPUT guardrail automatically active:

Default Guardrail Coverage

Sexual Content

Blocks explicit sexual material, inappropriate content, or sexual advances

Hate Speech

Blocks discrimination, prejudice, hateful content, or targeted harassment

Insults & Abuse

Blocks personal attacks, abusive language, or aggressive insults

Politics & Religion

Blocks political debates, partisan content, religious disputes, or divisive topics

Important: Default guardrail is optional enabled. It provides baseline protection for all agents without requiring any configuration.

Guardrail Types

Guardrails can be configured with different directions and actions:

1. Direction: INPUT vs OUTPUT

INPUT Guardrails
OUTPUT Guardrails
Both (INPUT + OUTPUT)

Applied to user messages before AI processingPurpose:

Protect AI model from harmful inputs
Filter malicious prompts
Mask sensitive user data
Block prohibited topics

Use Cases:

User-facing chatbots
Public interfaces
Customer service applications
Community platforms

Processing Flow:

User Message → INPUT Guardrail → AI Model → Response

INPUT guardrails run before the AI model sees the content, providing first-line defense against inappropriate inputs.

Applied to AI-generated responses before deliveryPurpose:

Ensure safe AI outputs
Prevent harmful responses
Mask PII in responses
Maintain brand compliance

Use Cases:

Content generation
Public communications
Regulated industries
Brand-sensitive applications

Processing Flow:

User Message → AI Model → AI Response → OUTPUT Guardrail → User

OUTPUT guardrails validate AI-generated content before users see it, ensuring safe and compliant responses.

Comprehensive protection on both sidesPurpose:

Maximum safety coverage
End-to-end protection
Full compliance assurance
Complete audit trail

Use Cases:

Highly regulated industries
High-risk applications
Maximum security requirements
Sensitive domains

Processing Flow:

User Message → INPUT Guardrail → AI Model → 
AI Response → OUTPUT Guardrail → User

Using both INPUT and OUTPUT guardrails provides maximum protection but adds latency to each interaction. Balance security needs with performance requirements.

2. Action: Block vs Mask

Block
Mask (Anonymize)

Completely prevent content from passingWhen to use:

Harmful content (hate speech, violence)
Prohibited topics (politics, religion)
Policy violations
Security threats
Inappropriate requests

Behavior:

Content Detected → BLOCKED → Safety Message Displayed

Example Configuration:

{
  "type": "INPUT",
  "action": "BLOCK",
  "categories": [
    "hate_speech",
    "violence",
    "sexual_content"
  ]
}

User Experience:

Request is not processed
Polite safety message displayed
User prompted to rephrase
Interaction logged for monitoring

Redact sensitive information while allowing flowWhen to use:

PII protection (emails, phones, SSN)
Financial data (credit cards, accounts)
Personal identifiers (names, addresses)
Health information (PHI)
Confidential data

Behavior:

Content Detected → PII MASKED → Processing Continues

Example Configuration:

{
  "type": "INPUT",
  "action": "MASK",
  "pii_types": [
    "email",
    "phone",
    "credit_card",
    "ssn",
    "name",
    "address"
  ]
}

Masking Examples:

Original: "My email is john.doe@example.com"
Masked:   "My email is [EMAIL_REDACTED]"

Original: "Call me at 555-123-4567"
Masked:   "Call me at [PHONE_REDACTED]"

Original: "My credit card is 4532-1234-5678-9012"
Masked:   "My credit card is [CREDIT_CARD_REDACTED]"

Original: "I'm John Smith from 123 Main St"
Masked:   "I'm [NAME_REDACTED] from [ADDRESS_REDACTED]"

Masking allows conversations to continue while protecting sensitive information from being processed, stored, or included in responses.

Creating Custom Guardrails

Custom guardrails are created on-demand through Amazon Bedrock Guardrails service to meet your specific requirements.

When to Create Custom Guardrails

Industry-Specific Compliance

Healthcare (HIPAA):

Mask protected health information (PHI)
Block medical advice outside scope
Prevent patient data disclosure

Financial Services (PCI-DSS):

Mask financial account details
Block unauthorized financial advice
Protect transaction information

Legal:

Prevent unauthorized legal advice
Protect privileged information
Maintain confidentiality

Organization-Specific Policies

Custom prohibited topics
Brand-specific content rules
Internal data protection
Proprietary information safeguards
Employee information protection

Advanced PII Protection

Custom PII types (employee IDs, patient numbers)
Industry-specific identifiers
Regional data protection (EU vs US)
Multi-language PII detection

Special Use Cases

Content generation safety
Academic integrity
Child safety protections
Community guidelines enforcement
Custom safety categories

Custom Guardrail Creation Process

Define Requirements

Document your specific needs:Required Information:

Purpose: What should this guardrail protect?
Direction: INPUT, OUTPUT, or both?
Action: Block or Mask?
Content Categories: What to filter?
PII Types: What to mask (if applicable)?
Scope: General or organization-only?

Request Creation

Contact your PLai Framework administrator or account manager:Provide:

Requirements document
Use case description
Compliance regulations
Timeline needs
Testing requirements

Methods:

Support ticket
Account manager email
Admin dashboard request
API (for enterprise customers)

Guardrail creation typically takes 2-5 business days depending on complexity and testing requirements.

Review and Testing

Once created, thoroughly test the guardrail:Test Scenarios:

Positive cases (should trigger)
Negative cases (should not trigger)
Edge cases
Performance impact
False positives
False negatives

Testing Checklist:

✅ Blocks/masks intended content
✅ Allows safe content through
✅ No excessive false positives
✅ Acceptable latency impact
✅ Works across different phrasings
✅ Handles edge cases appropriately
✅ Logs triggers correctly

Apply to Agents

Add the guardrail to your agents through:

Agent dashboard UI
API endpoint
Bulk application (multiple agents)

See “Applying Guardrails” section below for details.

Monitor and Refine

After deployment:

Monitor trigger rates
Review blocked content
Check for false positives
Adjust configuration as needed
Collect user feedback

Applying Guardrails to Agents

Via Dashboard UI

Navigate to Guardrails Section

Open your Agent Dashboard
Select the Guardrails tab
Select the guardrail(s) you want to apply.

Configure Settings

For each selected guardrail:Direction:

◉ INPUT only
◯ OUTPUT only
◯ Both INPUT and OUTPUT

Priority (if multiple guardrails):

Higher priority guardrails run first
Range: 1 (highest) to 10 (lowest)

Next Steps

Best Practices

Learn expert tips for optimal guardrail implementation

API Reference

Explore the complete Guardrails API documentation

Overview

Review Guardrails concepts and capabilities

Analytics

Monitor guardrail effectiveness and performance

Additional Resources

Amazon Bedrock Guardrails Documentation

Guardrails in PLai Framework are powered by Amazon Bedrock Guardrails. For technical details on the underlying service:

AWS Bedrock Guardrails Overview
Detection capabilities and models
PII types supported
Language support
Technical specifications

Compliance Resources

GDPR Compliance:

PII masking for EU users
Data protection requirements
Right to be forgotten

HIPAA Compliance:

PHI protection requirements
HIPAA Security Rule
HIPAA Privacy Rule

PCI-DSS Compliance:

Payment card data protection
Cardholder data environment
Security assessment procedures

Support and Assistance

For custom guardrail creation:

Contact your account manager
Submit detailed requirements document
Expected turnaround: 2-5 business days

Getting Started

🤖 Agents

🔧 Tools

🗄️ Datasources

📋 Batches

📋 Experiments

📊 Monitor

Advanced

Guardrails Configuration

Guardrails Configuration

Prerequisites

Understanding the Default Guardrail

Default Guardrail Coverage

Sexual Content

Hate Speech

Insults & Abuse

Politics & Religion

Guardrail Types

1. Direction: INPUT vs OUTPUT

2. Action: Block vs Mask

Creating Custom Guardrails

When to Create Custom Guardrails

Custom Guardrail Creation Process

Applying Guardrails to Agents

Via Dashboard UI

Next Steps

Best Practices

API Reference

Overview

Analytics

Additional Resources

Getting Started

🤖 Agents

🔧 Tools

🗄️ Datasources

📋 Batches

📋 Experiments

📊 Monitor

Advanced

​Guardrails Configuration

​Prerequisites

​Understanding the Default Guardrail

​Default Guardrail Coverage

Sexual Content

Hate Speech

Insults & Abuse

Politics & Religion

​Guardrail Types

​1. Direction: INPUT vs OUTPUT

​2. Action: Block vs Mask

​Creating Custom Guardrails

​When to Create Custom Guardrails

​Custom Guardrail Creation Process

​Applying Guardrails to Agents

​Via Dashboard UI

​Next Steps

Best Practices

API Reference

Overview

Analytics

​Additional Resources

Guardrails Configuration

Prerequisites

Understanding the Default Guardrail

Default Guardrail Coverage

Guardrail Types

1. Direction: INPUT vs OUTPUT

2. Action: Block vs Mask

Creating Custom Guardrails

When to Create Custom Guardrails

Custom Guardrail Creation Process

Applying Guardrails to Agents

Via Dashboard UI

Next Steps

Additional Resources