Skip to main content

API Reference

This page provides detailed information about the LLM Guard API.

LLMGuard​

The main class for validating and securing LLM prompts.

Constructor​

new LLMGuard(options?: LLMGuardOptions)

Parameters​

  • options (optional): Configuration options for the LLM Guard instance.
interface LLMGuardOptions {
  // Enable specific guards
  pii?: boolean;
  jailbreak?: boolean;
  profanity?: boolean;
  promptInjection?: boolean;
  relevance?: boolean;
  toxicity?: boolean;
  
  // Custom validation rules
  customRules?: Record<string, (prompt: string) => ValidationResult>;
  
  // Guard-specific options
  relevanceOptions?: RelevanceOptions;
  piiOptions?: PIIOptions;
  // ... other guard-specific options
}

Methods​

validate​

Validates a single prompt.

async validate(prompt: string): Promise<ValidationResult>
Parameters​
  • prompt: The prompt to validate.
Returns​

A ValidationResult object containing the validation status and any errors.

interface ValidationResult {
  isValid: boolean;
  prompt: string;
  errors: string[];
}

validateBatch​

Validates multiple prompts at once.

async validateBatch(prompts: string[]): Promise<ValidationResult[]>
Parameters​
  • prompts: An array of prompts to validate.
Returns​

An array of ValidationResult objects, one for each prompt.

Guards​

PII Guard​

Detects personally identifiable information.

interface PIIOptions {
  sensitivity?: 'low' | 'medium' | 'high';
  customPatterns?: RegExp[];
}

Profanity Guard​

Filters inappropriate language.

interface ProfanityOptions {
  language?: string;
  customWords?: string[];
}

Jailbreak Guard​

Detects attempts to bypass AI safety measures.

interface JailbreakOptions {
  threshold?: number;
  customPatterns?: RegExp[];
}

Prompt Injection Guard​

Identifies attempts to inject malicious instructions.

interface PromptInjectionOptions {
  sensitivity?: 'low' | 'medium' | 'high';
  customPatterns?: RegExp[];
}

Relevance Guard​

Evaluates the quality and relevance of prompts.

interface RelevanceOptions {
  minLength?: number;
  maxLength?: number;
  minWords?: number;
  maxWords?: number;
  fillerWords?: string[];
}

Toxicity Guard​

Detects harmful or aggressive content.

interface ToxicityOptions {
  threshold?: number;
  language?: string;
}

CLI​

LLM Guard provides a command-line interface for quick validation.

Usage​

npx llm-guard [options] <prompt>

Options​

  • --pii: Enable PII detection
  • --jailbreak: Enable jailbreak detection
  • --profanity: Enable profanity filtering
  • --prompt-injection: Enable prompt injection detection
  • --relevance: Enable relevance checking
  • --toxicity: Enable toxicity detection
  • --config <file>: Use a configuration file
  • --batch: Run in batch mode
  • --batch-file <file>: Process prompts from a file
  • --help: Show help information

Configuration File​

You can use a JSON configuration file to specify options:

{
  "pii": true,
  "jailbreak": true,
  "profanity": true,
  "promptInjection": true,
  "relevance": true,
  "toxicity": true,
  "relevanceOptions": {
    "minLength": 10,
    "maxLength": 5000
  }
}

Types​

ValidationResult​

interface ValidationResult {
  isValid: boolean;
  prompt: string;
  errors: string[];
}

LLMGuardOptions​

interface GuardConfig {
  pii?: boolean;
  jailbreak?: boolean;
  profanity?: boolean;
  promptInjection?: boolean;
  relevance?: boolean;
  toxicity?: boolean;
  customRules?: Record<string, any>;
  relevanceOptions?: RelevanceOptions;
  promptInjectionOptions?: PromptInjectionOptions;
  jailbreakOptions?: JailbreakOptions;
}

interface PromptInjectionOptions {
  /** Additive — these run alongside the built-in patterns, never instead of them. */
  customPatterns?: RegExp[];
}

interface JailbreakOptions {
  /** Additive — these run alongside the built-in patterns, never instead of them. */
  customPatterns?: RegExp[];
}

interface RelevanceOptions {
  minLength?: number;
  maxLength?: number;
  minWords?: number;
  maxWords?: number;
}

This block previously listed piiOptions, profanityOptions and toxicityOptions, and named the config type LLMGuardOptions. None of those existed. The exported type is GuardConfig, and the three guards above take no options — see #6.

Error Handling​

LLM Guard throws specific error types for different scenarios:

class ValidationError extends Error {
  constructor(message: string, public result: ValidationResult) {
    super(message);
    this.name = 'ValidationError';
  }
}

class ConfigurationError extends Error {
  constructor(message: string) {
    super(message);
    this.name = 'ConfigurationError';
  }
}

Examples​

Basic Usage​

import { LLMGuard } from 'llm-guard';

const guard = new LLMGuard({
  pii: true,
  jailbreak: true
});

const result = await guard.validate('My email is user@example.com');
console.log(result);

Custom Validation Rules​

import { LLMGuard } from 'llm-guard';

const guard = new LLMGuard({
  customRules: {
    minLength: (prompt) => {
      const minLength = 10;
      const isValid = prompt.length >= minLength;
      return {
        isValid,
        errors: isValid ? [] : [`Prompt must be at least ${minLength} characters long`]
      };
    }
  }
});

const result = await guard.validate('Short');
console.log(result);

Batch Validation​

import { LLMGuard } from 'llm-guard';

const guard = new LLMGuard();
const prompts = ['First prompt', 'Second prompt'];
const results = await guard.validateBatch(prompts);
console.log(results);