Creating RAI Policies

Create RAI policies to define safety guardrails for your agents. Policies can be created, updated, and applied to multiple agents.

Quick Start

from lyzr import Studio
from lyzr.rai import PIIType, PIIAction, SecretsAction

studio = Studio(api_key="your-api-key")

# Create a policy
policy = studio.create_rai_policy(
    name="StandardSafety",
    description="Standard safety guardrails for production",
    toxicity_threshold=0.4,
    prompt_injection=True,
    pii_detection={
        PIIType.CREDIT_CARD: PIIAction.BLOCK,
        PIIType.EMAIL: PIIAction.REDACT
    }
)

print(f"Policy created: {policy.id}")

studio.create_rai_policy()

studio.create_rai_policy(
    name: str,
    description: str,
    toxicity_threshold: float = 0.4,
    prompt_injection: bool = False,
    secrets_detection: SecretsAction = SecretsAction.DISABLED,
    pii_detection: Dict[PIIType, PIIAction] = None,
    banned_topics: List[str] = None,
    nsfw_check: bool = False,
    nsfw_threshold: float = 0.8,
    allowed_topics: Dict[str, Any] = None,
    keywords: Dict[str, Any] = None,
    fairness_and_bias: Dict[str, Any] = None
) -> RAIPolicy

Parameters

Parameter	Type	Default	Description
`name`	str	Required	Policy name
`description`	str	Required	Policy description
`toxicity_threshold`	float	0.4	Toxicity detection threshold (0.0-1.0)
`prompt_injection`	bool	False	Enable prompt injection detection
`secrets_detection`	SecretsAction	DISABLED	How to handle secrets
`pii_detection`	dict	None	PII type → action mapping
`banned_topics`	list	None	Topics to ban
`nsfw_check`	bool	False	Enable NSFW detection
`nsfw_threshold`	float	0.8	NSFW detection threshold
`allowed_topics`	dict	None	Allowed topics config
`keywords`	dict	None	Keyword filtering config
`fairness_and_bias`	dict	None	Fairness detection config

Enums

PIIType

from lyzr.rai import PIIType

PIIType.CREDIT_CARD    # Credit card numbers
PIIType.EMAIL          # Email addresses
PIIType.PHONE          # Phone numbers
PIIType.SSN            # US Social Security Numbers
PIIType.PERSON         # Person names
PIIType.LOCATION       # Physical locations
PIIType.IP_ADDRESS     # IP addresses
PIIType.URL            # Web URLs
PIIType.DATE_TIME      # Dates and times

PIIAction

from lyzr.rai import PIIAction

PIIAction.BLOCK      # Block the entire message
PIIAction.REDACT     # Replace with [REDACTED]
PIIAction.DISABLED   # No action

SecretsAction

from lyzr.rai import SecretsAction

SecretsAction.MASK      # Replace with asterisks
SecretsAction.BLOCK     # Block the message
SecretsAction.DISABLED  # No action

Creating Policies

Basic Policy

policy = studio.create_rai_policy(
    name="Basic",
    description="Basic safety policy",
    toxicity_threshold=0.5
)

Full-Featured Policy

from lyzr.rai import PIIType, PIIAction, SecretsAction

policy = studio.create_rai_policy(
    name="Comprehensive",
    description="Full safety coverage",

    # Toxicity detection
    toxicity_threshold=0.3,

    # Prompt injection protection
    prompt_injection=True,

    # Secrets handling
    secrets_detection=SecretsAction.MASK,

    # PII handling
    pii_detection={
        PIIType.CREDIT_CARD: PIIAction.BLOCK,
        PIIType.SSN: PIIAction.BLOCK,
        PIIType.EMAIL: PIIAction.REDACT,
        PIIType.PHONE: PIIAction.REDACT,
        PIIType.PERSON: PIIAction.DISABLED  # Allow names
    },

    # Topic control
    banned_topics=["politics", "religion", "competitors"],

    # NSFW detection
    nsfw_check=True,
    nsfw_threshold=0.8
)

With Advanced Configuration

policy = studio.create_rai_policy(
    name="Advanced",
    description="Advanced configuration",

    # Allowed topics (whitelist mode)
    allowed_topics={
        "enabled": True,
        "topics": ["customer_support", "products", "billing"]
    },

    # Keyword filtering
    keywords={
        "enabled": True,
        "keywords": ["banned_word1", "banned_word2"]
    },

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

Getting Policies

Get by ID

policy = studio.get_rai_policy("policy_id")

print(f"Name: {policy.name}")
print(f"Description: {policy.description}")
print(f"Toxicity: {policy.toxicity_check}")
print(f"PII: {policy.pii_detection}")

List All Policies

policies = studio.list_rai_policies()

for policy in policies:
    print(f"{policy.id}: {policy.name}")
    print(f"  Description: {policy.description}")
    print(f"  Created: {policy.created_at}")

Updating Policies

policy.update()

policy = studio.get_rai_policy("policy_id")

# Update specific settings
policy = policy.update(
    description="Updated description",
    toxicity_check={
        "enabled": True,
        "threshold": 0.3
    }
)

Update via Module

updated_policy = studio.rai.update_policy(
    policy_id="policy_id",
    name="New Name",
    toxicity_check={"enabled": True, "threshold": 0.25},
    prompt_injection={"enabled": True, "threshold": 0.3}
)

Update Parameters

Parameter	Type	Description
`name`	str	New policy name
`description`	str	New description
`toxicity_check`	dict	Toxicity config
`prompt_injection`	dict	Injection detection config
`secrets_detection`	dict	Secrets handling config
`pii_detection`	dict	PII handling config
`nsfw_check`	dict	NSFW detection config
`allowed_topics`	dict	Allowed topics config
`banned_topics`	dict	Banned topics config
`keywords`	dict	Keyword filter config
`fairness_and_bias`	dict	Fairness detection config

Deleting Policies

policy.delete()

policy = studio.get_rai_policy("policy_id")
success = policy.delete()

if success:
    print("Policy deleted")

Delete via Module

studio.rai.delete_policy("policy_id")

Applying to Agents

At Creation

policy = studio.create_rai_policy(
    name="SafePolicy",
    description="Safety guardrails"
)

agent = studio.create_agent(
    name="Safe Bot",
    provider="gpt-4o",
    role="Assistant",
    rai_policy=policy  # Apply policy
)

After Creation

agent = studio.create_agent(name="Bot", provider="gpt-4o")
policy = studio.get_rai_policy("policy_id")

# Add RAI policy
agent.add_rai_policy(policy)

RAIPolicy Properties

Property	Type	Description
`id`	str	Policy ID
`name`	str	Policy name
`description`	str	Policy description
`toxicity_check`	dict	Toxicity detection config
`prompt_injection`	dict	Injection detection config
`secrets_detection`	dict	Secrets handling config
`pii_detection`	dict	PII handling config
`nsfw_check`	dict	NSFW detection config
`allowed_topics`	dict	Allowed topics config
`banned_topics`	dict	Banned topics config
`keywords`	dict	Keyword filter config
`fairness_and_bias`	dict	Fairness detection config
`user_id`	str	Owner user ID
`created_at`	str	Creation timestamp
`updated_at`	str	Last update timestamp

Examples

Customer Service Policy

cs_policy = studio.create_rai_policy(
    name="CustomerService",
    description="Safe customer service interactions",
    toxicity_threshold=0.4,
    prompt_injection=True,
    pii_detection={
        PIIType.CREDIT_CARD: PIIAction.BLOCK,
        PIIType.SSN: PIIAction.BLOCK,
        PIIType.EMAIL: PIIAction.REDACT
    },
    secrets_detection=SecretsAction.MASK,
    banned_topics=["competitors", "internal_pricing"]
)

# Apply to support agents
support_agent = studio.create_agent(
    name="Support",
    provider="gpt-4o",
    role="Customer support",
    rai_policy=cs_policy
)

Healthcare Compliance

hipaa_policy = studio.create_rai_policy(
    name="HIPAA_Compliant",
    description="HIPAA-compliant healthcare policy",
    toxicity_threshold=0.3,
    prompt_injection=True,
    pii_detection={
        PIIType.SSN: PIIAction.BLOCK,
        PIIType.PERSON: PIIAction.REDACT,
        PIIType.PHONE: PIIAction.REDACT,
        PIIType.EMAIL: PIIAction.REDACT,
        PIIType.DATE_TIME: PIIAction.REDACT,
        PIIType.LOCATION: PIIAction.REDACT
    },
    secrets_detection=SecretsAction.BLOCK,
    allowed_topics={
        "enabled": True,
        "topics": ["health_info", "appointments", "medications", "insurance"]
    }
)

Child-Safe Content

kids_policy = studio.create_rai_policy(
    name="ChildSafe",
    description="Appropriate for all ages",
    toxicity_threshold=0.2,  # Very strict
    prompt_injection=True,
    nsfw_check=True,
    nsfw_threshold=0.9,
    banned_topics=[
        "violence",
        "adult_content",
        "gambling",
        "drugs",
        "alcohol"
    ],
    keywords={
        "enabled": True,
        "keywords": ["inappropriate_word1", "inappropriate_word2"]
    }
)

Financial Services

finance_policy = studio.create_rai_policy(
    name="FinanceCompliant",
    description="Financial services compliance",
    toxicity_threshold=0.4,
    prompt_injection=True,
    pii_detection={
        PIIType.CREDIT_CARD: PIIAction.BLOCK,
        PIIType.SSN: PIIAction.BLOCK,
        PIIType.PERSON: PIIAction.REDACT
    },
    secrets_detection=SecretsAction.BLOCK,
    banned_topics=[
        "specific_investment_advice",
        "guaranteed_returns",
        "insider_information"
    ]
)

Best Practices

Start Strict, Then Relax

# Start with strict settings
policy = studio.create_rai_policy(
    name="Initial",
    description="Starting strict",
    toxicity_threshold=0.2,  # Very strict
    prompt_injection=True
)

# Relax based on monitoring
policy.update(
    toxicity_check={"enabled": True, "threshold": 0.4}
)

Separate Policies by Use Case

# Different policies for different agents
public_policy = studio.create_rai_policy(
    name="Public",
    description="For public-facing agents",
    toxicity_threshold=0.3
)

internal_policy = studio.create_rai_policy(
    name="Internal",
    description="For internal tools",
    toxicity_threshold=0.6  # More relaxed
)

Document Your Policies

policy = studio.create_rai_policy(
    name="Production_v2",
    description="""
    Production safety policy v2 (2024-03)

    Changes from v1:
    - Lowered toxicity threshold to 0.3
    - Added prompt injection detection
    - Added credit card blocking

    Compliance: SOC2, GDPR
    Owner: Security Team
    """
)

Getting Started

Agents

Knowledge Bases

Memory

Tools

Contexts

RAI Guardrails

Structured Outputs

File & Image Output

Streaming

Reference

​Quick Start

​studio.create_rai_policy()

​Parameters

​Enums

​PIIType

​PIIAction

​SecretsAction

​Creating Policies

​Basic Policy

​Full-Featured Policy

​With Advanced Configuration

​Getting Policies

​Get by ID

​List All Policies

​Updating Policies

​policy.update()

​Update via Module

​Update Parameters

​Deleting Policies

​policy.delete()

​Delete via Module

​Applying to Agents

​At Creation

​After Creation

​RAIPolicy Properties

​Examples

​Customer Service Policy

​Healthcare Compliance

​Child-Safe Content

​Financial Services

​Best Practices

​Start Strict, Then Relax

​Separate Policies by Use Case

​Document Your Policies

Quick Start

studio.create_rai_policy()

Parameters

Enums

PIIType

PIIAction

SecretsAction

Creating Policies

Basic Policy

Full-Featured Policy

With Advanced Configuration

Getting Policies

Get by ID

List All Policies

Updating Policies

policy.update()

Update via Module

Update Parameters

Deleting Policies

policy.delete()

Delete via Module

Applying to Agents

At Creation

After Creation

RAIPolicy Properties

Examples

Customer Service Policy

Healthcare Compliance

Child-Safe Content

Financial Services

Best Practices

Start Strict, Then Relax

Separate Policies by Use Case

Document Your Policies