Guardrail Policies

Define safety policies to control agent behavior and protect against harmful outputs.

Quick Start

Define Policies

import { GuardrailPolicy, resolveGuardrailPolicies } from 'praisonai';

const policies: GuardrailPolicy[] = [
  {
    name: 'no-pii',
    action: 'block',
    message: 'PII detected in output'
  },
  {
    name: 'content-filter',
    action: 'warn',
    message: 'Potentially sensitive content'
  }
];

Use Preset Policies

import { resolveGuardrailPolicies, GUARDRAIL_PRESETS } from 'praisonai';

// Use preset by name
const resolved = resolveGuardrailPolicies(['strict', 'no-pii']);

Policy Actions

Action	Behavior
`block`	Stop execution and return error
`warn`	Log warning and continue
`log`	Log event and continue silently
`allow`	Allow without any action

GuardrailPolicy Interface

interface GuardrailPolicy {
  name: string;                    // Policy name
  action: 'block' | 'warn' | 'log' | 'allow';
  conditions?: Record<string, any>; // Optional conditions
  message?: string;                // Message to show/log
}

Preset Policies

Use built-in presets for common scenarios:

import { GUARDRAIL_PRESETS, resolveGuardrailPolicies } from 'praisonai';

// Available presets
GUARDRAIL_PRESETS.strict      // Block harmful content
GUARDRAIL_PRESETS.moderate    // Warn on sensitive content
GUARDRAIL_PRESETS.permissive  // Log only, allow all
GUARDRAIL_PRESETS['no-pii']   // Block PII
GUARDRAIL_PRESETS['no-code']  // Block code execution

// Resolve by name
const policies = resolveGuardrailPolicies(['strict', 'no-pii']);

Preset Definitions

Preset	Action	Description
`strict`	`block`	Block all harmful content
`moderate`	`warn`	Warn on sensitive content
`permissive`	`log`	Log everything, block nothing
`no-pii`	`block`	Block personal information
`no-code`	`block`	Block code execution

Resolve Policies

Convert string presets and policy objects:

import { resolveGuardrailPolicies, GuardrailPolicy } from 'praisonai';

// Mix of presets and custom policies
const policies = resolveGuardrailPolicies([
  'strict',                           // Preset name
  'no-pii',                           // Preset name
  {                                   // Custom policy
    name: 'custom-filter',
    action: 'warn',
    message: 'Custom warning'
  }
]);

// Returns array of GuardrailPolicy objects
console.log(policies);

Common Patterns

Content Moderation
Data Protection
Environment-Based

import { resolveGuardrailPolicies, GuardrailPolicy } from 'praisonai';

const moderationPolicies: GuardrailPolicy[] = [
  {
    name: 'profanity-filter',
    action: 'block',
    message: 'Profanity detected'
  },
  {
    name: 'hate-speech',
    action: 'block',
    message: 'Hate speech detected'
  },
  {
    name: 'spam-detection',
    action: 'warn',
    message: 'Potential spam detected'
  }
];

const resolved = resolveGuardrailPolicies(moderationPolicies);

import { GuardrailPolicy } from 'praisonai';

const dataProtectionPolicies: GuardrailPolicy[] = [
  {
    name: 'pii-detection',
    action: 'block',
    conditions: {
      patterns: ['ssn', 'credit-card', 'phone']
    },
    message: 'Personal information detected'
  },
  {
    name: 'api-key-leak',
    action: 'block',
    conditions: {
      patterns: ['sk-', 'api_key', 'secret']
    },
    message: 'API key detected in output'
  }
];

import { resolveGuardrailPolicies } from 'praisonai';

function getPolicies() {
  if (process.env.NODE_ENV === 'production') {
    return resolveGuardrailPolicies(['strict', 'no-pii']);
  } else if (process.env.NODE_ENV === 'staging') {
    return resolveGuardrailPolicies(['moderate']);
  } else {
    return resolveGuardrailPolicies(['permissive']);
  }
}

Best Practices

Use strict policies in production

Always use strict and no-pii policies in production environments.

Log before blocking

Use warn during development to understand what would be blocked.

Provide clear messages

Include descriptive messages to help users understand why content was blocked.

Layer policies

Combine multiple policies for comprehensive protection.

Guardrails

Full guardrails implementation

LLM Guardrail

LLM-based content filtering

JavaScript

​Quick Start

​Policy Actions

​GuardrailPolicy Interface

​Preset Policies

​Preset Definitions

​Resolve Policies

​Common Patterns

​Best Practices

​Related