RAI Guardrails Overview

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

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

# Create a safety policy
policy = studio.create_rai_policy(
    name="SafePolicy",
    description="Standard safety guardrails",
    toxicity_threshold=0.4,
    prompt_injection=True,
    secrets_detection=SecretsAction.MASK,
    pii_detection={
        PIIType.CREDIT_CARD: PIIAction.BLOCK,
        PIIType.EMAIL: PIIAction.REDACT,
        PIIType.SSN: PIIAction.BLOCK
    }
)

# Apply to agent
agent = studio.create_agent(
    name="Safe Bot",
    provider="gpt-4o",
    role="Customer support",
    goal="Help users safely",
    rai_policy=policy
)

# Agent now has guardrails active
response = agent.run("Process my order with card 4111-1111-1111-1111")
# Credit card is blocked, agent asks for alternative payment method

policy = studio.create_rai_policy(
    name="NoToxicity",
    description="Block toxic content",
    toxicity_threshold=0.3  # Lower = stricter (0.0-1.0)
)

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

from lyzr.rai import PIIType, PIIAction

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

from lyzr.rai import SecretsAction

policy = studio.create_rai_policy(
    name="SecretsSafe",
    description="Mask secrets",
    secrets_detection=SecretsAction.MASK  # or BLOCK
)

policy = studio.create_rai_policy(
    name="TopicRestricted",
    description="Stay on topic",
    banned_topics=["politics", "religion", "competitors"],
    allowed_topics={"enabled": True, "topics": ["products", "support", "billing"]}
)

policy = studio.create_rai_policy(
    name="FamilyFriendly",
    description="Block NSFW content",
    nsfw_check=True,
    nsfw_threshold=0.8
)

User Input → RAI Pre-check → Agent Processing → RAI Post-check → Response

1. Pre-check: Scan input for toxicity, injection, PII, etc.
2. Processing: Agent generates response (if input passes)
3. Post-check: Scan output for policy violations
4. Response: Clean, safe response delivered to user

support_policy = studio.create_rai_policy(
    name="SupportSafe",
    description="Safe customer support interactions",
    toxicity_threshold=0.4,
    prompt_injection=True,
    pii_detection={
        PIIType.CREDIT_CARD: PIIAction.BLOCK,
        PIIType.SSN: PIIAction.BLOCK,
        PIIType.EMAIL: PIIAction.REDACT
    },
    banned_topics=["competitors", "internal_processes"]
)

health_policy = studio.create_rai_policy(
    name="HealthSafe",
    description="HIPAA-compliant interactions",
    toxicity_threshold=0.3,
    pii_detection={
        PIIType.SSN: PIIAction.BLOCK,
        PIIType.PHONE: PIIAction.REDACT,
        PIIType.PERSON: PIIAction.REDACT
    },
    secrets_detection=SecretsAction.BLOCK
)

kids_policy = studio.create_rai_policy(
    name="KidsSafe",
    description="Safe for children",
    toxicity_threshold=0.2,  # Very strict
    nsfw_check=True,
    nsfw_threshold=0.9,
    banned_topics=["violence", "adult_content", "gambling"]
)