Guardrails AI

Guardrails AI is an open-source framework for adding structured validation and safety checks to LLM outputs. Define a Guard with validators — JSON schema enforcement, PII detection, toxicity filtering, length limits, regex matching, and more — and Guardrails automatically re-prompts when outputs fail.

Key features

• Rail spec — declare the expected output structure and validators in YAML or Python
• 100+ validators — built-in checks for PII, toxicity, bias, valid JSON, SQL safety, and more
• Auto-fix and retry — on failure, Guardrails re-prompts with error feedback or programmatically fixes outputs
• Streaming support — validate streaming chunks as they arrive
• Hub — community-contributed validators for domain-specific rules
• Provider-agnostic — works with OpenAI, Anthropic, Cohere, and any LangChain-compatible model

Quick start

npx ai-supply add guardrails-ai-output-validation

# Or install directly
pip install guardrails-ai
guardrails hub install hub://guardrails/valid_json

from guardrails import Guard
from guardrails.hub import ValidJson

guard = Guard().use(ValidJson, on_fail="reask")

result = guard(
    lambda: '{"name": "Alice", "age": 30}',  # Replace with your LLM call
    prompt="Return a JSON object with name and age."
)
print(result.validated_output)  # {"name": "Alice", "age": 30}

Curated mirror of the open-source Guardrails AI project (Apache-2.0). Install upstream from the repository.