Guardrails — Validate & Secure LLM Outputs

What it is

Guardrails is a Python framework that wraps LLM calls with validation logic. It intercepts inputs and outputs, runs them through configurable validators, and ensures the LLM produces structured, safe, and correct responses. The library ships with pre-built validators for common risks like PII detection, toxic content, JSON schema compliance, and hallucination checks. You define your output schema using Pydantic models, and Guardrails enforces it.

Developers building production LLM applications who need reliable, structured outputs benefit from Guardrails. It is particularly useful for applications where incorrect or unsafe LLM responses carry real consequences.

How it saves time or tokens

Without validation, developers manually inspect and retry LLM outputs when they fail to meet requirements. Guardrails automates this loop: if an output fails validation, it re-prompts the LLM with correction instructions. This automatic retry mechanism reduces manual debugging time and avoids wasted tokens on outputs that would be discarded anyway.

How to use

1. Install Guardrails via pip and define your output schema with Pydantic
2. Wrap your LLM call with a Guardrails guard that applies your chosen validators
3. Call the guard instead of the LLM directly; it returns validated, structured output

Example

from guardrails import Guard
from guardrails.hub import DetectPII, ToxicLanguage
from pydantic import BaseModel

class UserResponse(BaseModel):
    answer: str
    confidence: float

guard = Guard().use_many(
    DetectPII(pii_entities=['EMAIL', 'PHONE']),
    ToxicLanguage(threshold=0.8)
)

result = guard(
    model='gpt-4o',
    messages=[{'role': 'user', 'content': 'Summarize this document.'}],
    output_class=UserResponse
)

print(result.validated_output)

Related on TokRepo

• AI tools for security — Explore security-focused AI tooling
• AI tools for testing — Browse testing and validation frameworks

Common pitfalls

• Stacking too many validators increases latency and token cost per call; validate only what matters for your use case
• Some validators require external models or APIs (e.g., PII detection); check dependencies before deploying
• Automatic retries can loop indefinitely if the LLM consistently fails validation; always set a max retry count