guardrails-ai-setup

guardrails-ai-setup

Configure Guardrails AI validation framework to ensure LLM outputs meet quality, safety, and structural requirements. Implement validators for input sanitization, output format enforcement, and safety constraints.

Overview

Guardrails AI provides:

• Input validation before LLM calls
• Output validation after LLM responses
• Structured output enforcement (JSON, XML, etc.)
• Pre-built validators from Guardrails Hub
• Custom validator creation
• Automatic retry and correction mechanisms

Capabilities

Input Validation

• Sanitize user inputs
• Detect prompt injection attempts
• Validate input formats and lengths
• Check for PII before processing

Output Validation

• Enforce structured output schemas
• Validate content accuracy
• Check for harmful content
• Verify factual consistency

Safety Constraints

• Content moderation
• Toxicity detection
• Bias checking
• Hallucination detection

Integration Features

• LangChain integration
• Streaming support
• Automatic retries
• Correction strategies

Usage

Basic Setup

from guardrails import Guard
from guardrails.hub import ValidJson, ToxicLanguage, DetectPII

# Create guard with validators
guard = Guard().use_many(
    ValidJson(),
    ToxicLanguage(on_fail="fix"),
    DetectPII(on_fail="fix")
)

# Use with LLM
from langchain_openai import ChatOpenAI

llm = ChatOpenAI(model="gpt-4")

result = guard(
    llm,
    prompt="Generate a product description for a laptop",
    max_tokens=500
)

print(result.validated_output)

Schema-Based Validation

from guardrails import Guard
from pydantic import BaseModel, Field
from typing import List

class ProductReview(BaseModel):
    """Schema for product review output."""
    rating: int = Field(ge=1, le=5, description="Rating from 1-5")
    summary: str = Field(max_length=200, description="Brief summary")
    pros: List[str] = Field(min_items=1, max_items=5)
    cons: List[str] = Field(min_items=1, max_items=5)
    recommendation: bool

# Create guard from schema
guard = Guard.from_pydantic(ProductReview)

result = guard(
    llm,
    prompt="""Analyze this product and provide a structured review:
    Product: Wireless Noise-Canceling Headphones
    Price: $299
    Features: 30hr battery, ANC, Bluetooth 5.3
    """,
)

# Result is a validated ProductReview instance
review = result.validated_output
print(f"Rating: {review.rating}")
print(f"Summary: {review.summary}")