RAI Guardrails Overview

Responsible AI guardrails for safe agent behavior

RAI (Responsible AI) guardrails protect your agents by filtering harmful content, detecting prompt injection, handling PII, and enforcing topic boundaries.

Quick Start

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

---

Key Features

Toxicity Detection

Block or filter toxic, harmful, or offensive content:

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

Prompt Injection Protection

Prevent malicious prompt manipulation:

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

PII Detection & Handling

Detect and handle personally identifiable information:

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
    }
)

Secrets Detection

Prevent API keys, passwords, and other secrets from leaking:

from lyzr.rai import SecretsAction

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

Topic Control

Restrict agent to specific topics:

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

NSFW Detection

Filter inappropriate content:

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

---

How RAI Works

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

---

RAI Components

Component	Purpose	Actions
Toxicity	Detect harmful/offensive content	Block if threshold exceeded
Prompt Injection	Prevent malicious prompts	Block suspicious inputs
PII Detection	Find personal information	Block, Redact, or Allow
Secrets Detection	Find API keys, passwords	Mask or Block
NSFW	Detect inappropriate content	Block if threshold exceeded
Topic Control	Enforce topic boundaries	Ban or allow specific topics
Keywords	Filter specific words	Block messages with keywords

---

PII Types

Type	Description	Example
CREDIT_CARD	Credit card numbers	4111-1111-1111-1111
EMAIL	Email addresses	[e-mail]
PHONE	Phone numbers	[phone]
SSN	US Social Security Numbers	123-45-6789
PERSON	Person names	John Smith
LOCATION	Physical 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

---

Actions

PII Actions

Action	Behavior
BLOCK	Reject the entire message
REDACT	Replace PII with [REDACTED]
DISABLED	No action (allow through)

Secrets Actions

Action	Behavior
MASK	Replace secrets with asterisks
BLOCK	Reject messages containing secrets
DISABLED	No action (allow through)

---

Use Cases

Customer Support

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"]
)

Healthcare

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
)

Children’s Content

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"]
)

---

Next Steps

Creating Policies

Learn how to create and configure RAI policies

RAI Features

Detailed guide to each RAI feature