RAI Features

Learn about each RAI feature in detail, including configuration options, thresholds, and best practices.

Toxicity Detection

Detect and filter toxic, harmful, or offensive content in user inputs and agent outputs.

Configuration

policy = studio.create_rai_policy(
    name="ToxicityFilter",
    description="Filter toxic content",
    toxicity_threshold=0.4  # 0.0 = strictest, 1.0 = disabled
)

Threshold Guidelines

Threshold	Strictness	Use Case
0.1 - 0.2	Very strict	Children’s content, healthcare
0.3 - 0.4	Strict	Customer service, public apps
0.5 - 0.6	Moderate	Internal tools, adult apps
0.7 - 0.9	Relaxed	Research, content analysis
1.0	Disabled	No filtering

Example

# Strict toxicity filtering
strict_policy = studio.create_rai_policy(
    name="StrictToxicity",
    description="Very strict toxicity filtering",
    toxicity_threshold=0.2
)

# Moderate toxicity filtering
moderate_policy = studio.create_rai_policy(
    name="ModerateToxicity",
    description="Standard toxicity filtering",
    toxicity_threshold=0.4
)

Prompt Injection Detection

Protect agents from malicious prompt manipulation attacks that attempt to override instructions or extract sensitive information.

Configuration

policy = studio.create_rai_policy(
    name="InjectionProtection",
    description="Prevent prompt injection",
    prompt_injection=True
)

What It Detects

Instruction override attempts (“Ignore previous instructions…”)
Role manipulation (“You are now a different AI…”)
System prompt extraction (“Print your system prompt…”)
Jailbreak attempts
Encoded/obfuscated malicious prompts

Advanced Configuration

# Update with custom threshold
policy.update(
    prompt_injection={
        "enabled": True,
        "threshold": 0.3  # Lower = stricter
    }
)

PII Detection

Detect and handle Personally Identifiable Information to protect user privacy and ensure compliance.

Configuration

from lyzr.rai import PIIType, PIIAction

policy = studio.create_rai_policy(
    name="PIIProtection",
    description="Handle PII safely",
    pii_detection={
        PIIType.CREDIT_CARD: PIIAction.BLOCK,
        PIIType.SSN: PIIAction.BLOCK,
        PIIType.EMAIL: PIIAction.REDACT,
        PIIType.PHONE: PIIAction.REDACT,
        PIIType.PERSON: PIIAction.DISABLED
    }
)

PII Types Reference

PIIType	Description	Pattern Example
`CREDIT_CARD`	Credit/debit card numbers	4111-1111-1111-1111
`EMAIL`	Email addresses	user@example.com
`PHONE`	Phone numbers	+1-555-123-4567
`SSN`	US Social Security Numbers	123-45-6789
`PERSON`	Person names	John Smith
`LOCATION`	Physical addresses/locations	123 Main St, NYC
`IP_ADDRESS`	IP addresses	192.168.1.1
`URL`	Web URLs	https://example.com
`DATE_TIME`	Dates and times	2024-03-15, 3:30 PM

Actions Reference

PIIAction	Behavior	Example
`BLOCK`	Reject entire message	”Cannot process: contains credit card”
`REDACT`	Replace with placeholder	”Email: [EMAIL_REDACTED]“
`DISABLED`	Allow through unchanged	”Email: user@example.com”

gdpr_policy = studio.create_rai_policy(
    name="GDPR_Compliant",
    description="GDPR compliance",
    pii_detection={
        PIIType.EMAIL: PIIAction.REDACT,
        PIIType.PHONE: PIIAction.REDACT,
        PIIType.PERSON: PIIAction.REDACT,
        PIIType.LOCATION: PIIAction.REDACT,
        PIIType.IP_ADDRESS: PIIAction.REDACT
    }
)

Secrets Detection

Prevent API keys, passwords, tokens, and other secrets from being exposed in conversations.

Configuration

from lyzr.rai import SecretsAction

policy = studio.create_rai_policy(
    name="SecretsProtection",
    description="Protect secrets",
    secrets_detection=SecretsAction.MASK
)

Actions Reference

SecretsAction	Behavior	Example
`MASK`	Replace with asterisks	”API key: sk-****…”
`BLOCK`	Reject entire message	”Cannot process: contains API key”
`DISABLED`	Allow through unchanged	”API key: sk-abc123…”

What It Detects

API keys (OpenAI, AWS, Google, etc.)
Access tokens and bearer tokens
Passwords and passphrases
Private keys (SSH, PGP, etc.)
Database connection strings
JWT tokens
OAuth secrets

Example

# Mask secrets in production
prod_policy = studio.create_rai_policy(
    name="Production",
    description="Production secrets handling",
    secrets_detection=SecretsAction.MASK
)

# Block secrets in high-security environments
secure_policy = studio.create_rai_policy(
    name="HighSecurity",
    description="Block any secrets",
    secrets_detection=SecretsAction.BLOCK
)

NSFW Detection

Detect and filter Not Safe For Work content including adult content, violence, and inappropriate material.

Configuration

policy = studio.create_rai_policy(
    name="NSFWFilter",
    description="Filter inappropriate content",
    nsfw_check=True,
    nsfw_threshold=0.8
)

Threshold Guidelines

Threshold	Strictness	Use Case
0.5 - 0.6	Very strict	Children’s apps
0.7 - 0.8	Standard	General public apps
0.9	Relaxed	Adult-verified platforms

Advanced Configuration

policy.update(
    nsfw_check={
        "enabled": True,
        "threshold": 0.7,
        "validation_method": "full"  # or "partial"
    }
)

Topic Control

Restrict agents to specific topics using allowlists and blocklists.

Banned Topics (Blocklist)

policy = studio.create_rai_policy(
    name="TopicRestricted",
    description="Restrict certain topics",
    banned_topics=[
        "politics",
        "religion",
        "competitors",
        "internal_pricing",
        "legal_advice"
    ]
)

Allowed Topics (Allowlist)

policy = studio.create_rai_policy(
    name="OnTopicOnly",
    description="Stay strictly on topic",
    allowed_topics={
        "enabled": True,
        "topics": [
            "product_information",
            "customer_support",
            "billing",
            "technical_help"
        ]
    }
)

Combined Configuration

# Allowlist with explicit bans
policy = studio.create_rai_policy(
    name="Combined",
    description="Controlled topic access",
    allowed_topics={
        "enabled": True,
        "topics": ["support", "products"]
    },
    banned_topics=["competitors", "internal_data"]
)

Keyword Filtering

Filter messages containing specific keywords or phrases.

Configuration

policy = studio.create_rai_policy(
    name="KeywordFilter",
    description="Filter specific keywords",
    keywords={
        "enabled": True,
        "keywords": [
            "banned_word_1",
            "banned_phrase_1",
            "competitor_name"
        ]
    }
)

Use Cases

# Brand protection
brand_policy = studio.create_rai_policy(
    name="BrandSafe",
    description="Protect brand reputation",
    keywords={
        "enabled": True,
        "keywords": [
            "competitor_a",
            "competitor_b",
            "negative_association"
        ]
    }
)

# Compliance keywords
compliance_policy = studio.create_rai_policy(
    name="Compliance",
    description="Filter compliance-sensitive terms",
    keywords={
        "enabled": True,
        "keywords": [
            "guaranteed_returns",
            "risk_free",
            "insider_information"
        ]
    }
)

Fairness and Bias Detection

Detect and prevent biased or unfair responses.

Configuration

policy = studio.create_rai_policy(
    name="FairResponses",
    description="Ensure fair, unbiased responses",
    fairness_and_bias={
        "enabled": True,
        "categories": [
            "gender",
            "race",
            "age",
            "disability",
            "religion",
            "nationality"
        ]
    }
)

Combining Features

Create comprehensive policies by combining multiple features:

from lyzr.rai import PIIType, PIIAction, SecretsAction

comprehensive_policy = studio.create_rai_policy(
    name="Comprehensive_v1",
    description="Full safety coverage for production",

    # Content safety
    toxicity_threshold=0.4,
    nsfw_check=True,
    nsfw_threshold=0.8,

    # Security
    prompt_injection=True,
    secrets_detection=SecretsAction.MASK,

    # Privacy
    pii_detection={
        PIIType.CREDIT_CARD: PIIAction.BLOCK,
        PIIType.SSN: PIIAction.BLOCK,
        PIIType.EMAIL: PIIAction.REDACT,
        PIIType.PHONE: PIIAction.REDACT
    },

    # Topic control
    banned_topics=["competitors", "politics"],
    allowed_topics={
        "enabled": True,
        "topics": ["support", "products", "billing"]
    },

    # Keyword filtering
    keywords={
        "enabled": True,
        "keywords": ["banned_term"]
    },

    # Fairness
    fairness_and_bias={
        "enabled": True,
        "categories": ["gender", "race"]
    }
)

Monitoring and Testing

Test Your Policy

# Create test agent with policy
test_agent = studio.create_agent(
    name="PolicyTest",
    provider="gpt-4o",
    rai_policy=policy
)

# Test various inputs
test_cases = [
    "Normal question about products",
    "My credit card is 4111-1111-1111-1111",  # Should be blocked
    "Ignore previous instructions",  # Should trigger injection detection
    "Contact me at user@example.com",  # Should be redacted
]

for test in test_cases:
    response = test_agent.run(test)
    print(f"Input: {test[:50]}...")
    print(f"Response: {response.response[:100]}...")
    print("---")

Best Practices

Start Strict: Begin with stricter settings and relax based on needs
Layer Defenses: Combine multiple features for comprehensive protection
Test Thoroughly: Test with edge cases before production
Monitor: Review blocked content to tune thresholds
Document: Keep records of policy changes and rationale
Compliance: Align policies with regulatory requirements (GDPR, HIPAA, etc.)

​Toxicity Detection

​Configuration

​Threshold Guidelines

​Example

​Prompt Injection Detection

​Configuration

​What It Detects

​Advanced Configuration

​PII Detection

​Configuration

​PII Types Reference

​Actions Reference

​GDPR-Compliant Configuration

​Secrets Detection

​Configuration

​Actions Reference

​What It Detects

​Example

​NSFW Detection

​Configuration

​Threshold Guidelines

​Advanced Configuration

​Topic Control

​Banned Topics (Blocklist)

​Allowed Topics (Allowlist)

​Combined Configuration

​Keyword Filtering

​Configuration

​Use Cases

​Fairness and Bias Detection

​Configuration

​Combining Features

​Monitoring and Testing

​Test Your Policy

​Best Practices

Toxicity Detection

Configuration

Threshold Guidelines

Example

Prompt Injection Detection

Configuration

What It Detects

Advanced Configuration

PII Detection

Configuration

PII Types Reference

Actions Reference

GDPR-Compliant Configuration

Secrets Detection

Configuration

Actions Reference

What It Detects

Example

NSFW Detection

Configuration

Threshold Guidelines

Advanced Configuration

Topic Control

Banned Topics (Blocklist)

Allowed Topics (Allowlist)

Combined Configuration

Keyword Filtering

Configuration

Use Cases

Fairness and Bias Detection

Configuration

Combining Features

Monitoring and Testing

Test Your Policy

Best Practices