LLM Guardrails

pip install guardrails-ai
guardrails hub install hub://guardrails/toxic_language
guardrails hub install hub://guardrails/detect_pii
guardrails hub install hub://guardrails/restrict_to_topic

# script_id: day_064_llm_guardrails/basic_input_output_validation
from guardrails import Guard
from guardrails.hub import ToxicLanguage, DetectPII, RestrictToTopic

# Input validation guard
input_guard = Guard().use(
    ToxicLanguage(on_fail="exception"),
    DetectPII(["EMAIL", "PHONE"], on_fail="fix"),
)

# Use with LLM — the guard wraps the API call (routed via LiteLLM)
user_input = "How do I write a Python function?"
result = input_guard(
    model="gpt-4o-mini",
    messages=[{"role": "user", "content": user_input}],
)

print(result.validated_output)

# script_id: day_064_llm_guardrails/structured_output_validation
from guardrails import Guard
from pydantic import BaseModel, Field
from typing import List

class ProductReview(BaseModel):
    product_name: str = Field(description="Name of the product")
    rating: float = Field(ge=1, le=5, description="Rating from 1 to 5")
    pros: List[str] = Field(min_length=1, description="List of pros")
    cons: List[str] = Field(description="List of cons")
    summary: str = Field(max_length=200, description="Brief summary")

guard = Guard.for_pydantic(ProductReview)

result = guard(
    model="gpt-4o-mini",
    messages=[{
        "role": "user",
        "content": "Review the Sony WH-1000XM5 headphones."
    }],
)

# result.validated_output is a validated dict matching ProductReview
print(result.validated_output["rating"])
print(result.validated_output["pros"])

# script_id: day_064_llm_guardrails/topic_restriction
from guardrails import Guard
from guardrails.hub import RestrictToTopic

guard = Guard().use(
    RestrictToTopic(
        valid_topics=["programming", "technology", "software engineering"],
        invalid_topics=["politics", "religion", "medical advice"],
        on_fail="exception",
    )
)

# This passes
result = guard.validate("How do I implement a binary search tree?")

# This raises an exception
guard.validate("What's your opinion on the election?")  # raises ValidationError

# script_id: day_064_llm_guardrails/custom_validator
from guardrails import Validator, register_validator
from guardrails.validator_base import PassResult, FailResult
from typing import Any, Dict, List

@register_validator("no-competitor-mention", data_type="string")
class NoCompetitorMention(Validator):
    """Block mentions of competitor names."""

    def __init__(self, competitors: List[str], on_fail: str = "fix"):
        super().__init__(on_fail=on_fail)
        self.competitors = [c.lower() for c in competitors]

    def validate(self, value: Any, metadata: Dict):
        # Build the redacted version and check whether anything changed.
        fixed = value
        for competitor in self.competitors:
            fixed = fixed.replace(competitor, "[competitor]")
            fixed = fixed.replace(competitor.title(), "[Competitor]")

        if fixed != value:
            # The framework reads fix_value when on_fail="fix".
            return FailResult(
                error_message="Competitor name mentioned in output",
                fix_value=fixed,
            )
        return PassResult()

# Usage
guard = Guard().use(
    NoCompetitorMention(
        competitors=["microsoft", "google", "amazon"],
        on_fail="fix"
    )
)

result = guard.validate("Our product is better than Microsoft's solution")
print(result.validated_output)
# "Our product is better than [Competitor]'s solution"

# script_id: day_064_llm_guardrails/openai_moderation
from openai import OpenAI

client = OpenAI()

def check_moderation(text: str) -> dict:
    """Check text against OpenAI's moderation categories."""
    response = client.moderations.create(input=text)
    result = response.results[0]

    if result.flagged:
        # Identify which categories were triggered
        triggered = [
            category for category, flagged
            in result.categories.model_dump().items()
            if flagged
        ]
        return {"safe": False, "categories": triggered}

    return {"safe": True, "categories": []}

# Use as a pre-check before LLM calls
user_input = "How do I write a Python function?"
moderation = check_moderation(user_input)

if not moderation["safe"]:
    print(f"Blocked: {moderation['categories']}")
else:
    # Proceed with LLM call
    response = client.chat.completions.create(
        model="gpt-4o-mini",
        messages=[{"role": "user", "content": user_input}]
    )

# script_id: day_064_llm_guardrails/claude_system_guardrails
from anthropic import Anthropic

client = Anthropic()

user_input = "How do I write a Python function?"
response = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=1024,
    system="""You are a helpful programming assistant.

GUARDRAILS:
- Only answer questions about programming, software engineering, and technology.
- If asked about topics outside your scope, politely decline.
- Never include personal information (emails, phone numbers, addresses) in responses.
- Do not generate code that could be used for hacking or exploitation.
- Keep responses concise and under 500 words.""",
    messages=[{"role": "user", "content": user_input}]
)

# script_id: day_064_llm_guardrails/guarded_agent
import openai
from guardrails import Guard
from guardrails.hub import ToxicLanguage, DetectPII

class GuardedAgent:
    """Agent with layered guardrails on input and output."""

    def __init__(self):
        self.client = openai.OpenAI()

        # Input: block toxic content, redact PII
        self.input_guard = Guard().use(
            ToxicLanguage(on_fail="exception"),
            DetectPII(["EMAIL", "PHONE", "SSN"], on_fail="fix"),
        )

        # Output: clean PII leaks, enforce length
        self.output_guard = Guard().use(
            DetectPII(["EMAIL", "PHONE", "SSN"], on_fail="fix"),
            ToxicLanguage(on_fail="fix"),
        )

    def generate(self, user_input: str) -> str:
        """Generate a guarded response."""

        # Validate input
        try:
            input_result = self.input_guard.validate(user_input)
            clean_input = input_result.validated_output
        except Exception as e:
            return f"I can't process that request: {e}"

        # Call LLM
        response = self.client.chat.completions.create(
            model="gpt-4o-mini",
            messages=[{"role": "user", "content": clean_input}]
        )
        raw_output = response.choices[0].message.content

        # Validate output
        output_result = self.output_guard.validate(raw_output)
        return output_result.validated_output

# Usage
agent = GuardedAgent()
print(agent.generate("How do I write a Python function?"))

# script_id: day_064_llm_guardrails/quick_reference
# fragment: illustrative cheat-sheet / not standalone-runnable
# Guardrails AI — input/output validation
from guardrails import Guard
from guardrails.hub import ToxicLanguage, DetectPII

guard = Guard().use(ToxicLanguage(on_fail="exception"), DetectPII(on_fail="fix"))
result = guard(model="gpt-4o-mini", messages=[...])

# Guardrails AI — structured output
guard = Guard.for_pydantic(MyModel)
result = guard(model="gpt-4o-mini", messages=[...])

# OpenAI Moderation API
response = client.moderations.create(input=text)
flagged = response.results[0].flagged

# Claude system prompt guardrails
response = client.messages.create(model="claude-sonnet-4-6", system="GUARDRAILS: ...", ...)