Guardrails

Content validation and safety

Guardrails

Guardrails provide content validation and safety checks for agent outputs.

Quick Start

import { createLLMGuardrail } from 'praisonai';

const guardrail = createLLMGuardrail({
  name: 'safety-check',
  criteria: 'Content should be safe, appropriate, and helpful',
  llm: 'openai/gpt-4o-mini'
});

const result = await guardrail.check('Hello world');
console.log(result.status); // 'passed', 'failed', or 'warning'
console.log(result.score);

Configuration

interface LLMGuardrailConfig {
  name: string;
  criteria: string;
  llm?: string;
  threshold?: number;  // Default: 0.7
  verbose?: boolean;
}

Response Format

interface LLMGuardrailResult {
  status: 'passed' | 'failed' | 'warning';
  score: number;  // 0 to 1
  message?: string;
  reasoning?: string;
}

Example

import { createLLMGuardrail, Agent } from 'praisonai';

// Create guardrail
const guardrail = createLLMGuardrail({
  name: 'professional-check',
  criteria: 'Content must be professional and appropriate for business use',
  threshold: 0.8
});

// Check content
const result = await guardrail.check('This is a professional message.');

if (result.status === 'passed') {
  console.log('Content approved');
} else {
  console.log('Content rejected:', result.reasoning);
}

CLI Usage

praisonai-ts guardrail check "Content to validate"
praisonai-ts guardrail check "Text" --criteria "Must be professional" --json