Skip to main content

Configuration

LLM Guard provides various configuration options to customize its behavior. This page explains how to configure LLM Guard for your specific needs.

Basic Configuration​

When creating a new LLM Guard instance, you can specify which guards to enable:

import { LLMGuard } from 'llm-guard';

const guard = new LLMGuard({
  pii: true,
  jailbreak: true,
  profanity: true,
  promptInjection: true,
  relevance: true,
  toxicity: true
});

Guard-Specific Options​

Each guard can be configured with specific options:

PII Guard Options​

Not implemented yet

piiOptions is not supported. The PII guard takes no options — enable it with pii: true and it uses its built-in patterns. Tracked in #6.

Profanity Guard Options​

Not implemented yet

profanityOptions is not supported. The profanity guard takes no options — enable it with profanity: true and it uses its built-in word list. Tracked in #6.

Jailbreak Guard Options​

const guard = new LLMGuard({
  jailbreak: true,
  jailbreakOptions: {
    // Additive: these run alongside the built-in patterns, never instead of them.
    customPatterns: [
      /enter developer mode/i,
      /pretend you are unrestricted/i
    ]
  }
});

customPatterns is the only supported field. There is no threshold — the jailbreak guard is pattern-based, not scored.

Prompt Injection Guard Options​

const guard = new LLMGuard({
  promptInjection: true,
  promptInjectionOptions: {
    // Additive: these run alongside the built-in patterns, never instead of them.
    customPatterns: [
      /exfiltrate the vault/i,
      /override system prompt/i
    ]
  }
});

customPatterns is the only supported field. There is no sensitivity — the guard is pattern-based, so widen or narrow it by supplying your own patterns.

Relevance Guard Options​

const guard = new LLMGuard({
  relevance: true,
  relevanceOptions: {
    minLength: 10, // Minimum text length
    maxLength: 5000, // Maximum text length
    minWords: 3, // Minimum word count
    maxWords: 1000 // Maximum word count
  }
});

Toxicity Guard Options​

Not implemented yet

toxicityOptions is not supported. The toxicity guard takes no options — enable it with toxicity: true and it uses its built-in word list. Tracked in #6.

Custom Validation Rules​

You can create custom validation rules to address specific validation needs:

const guard = new LLMGuard({
  customRules: {
    // Check for minimum length
    minLength: (prompt) => {
      const minLength = 10;
      const isValid = prompt.length >= minLength;
      return {
        isValid,
        errors: isValid ? [] : [`Prompt must be at least ${minLength} characters long`]
      };
    },
    
    // Check for specific keywords
    requiredKeywords: (prompt) => {
      const requiredWords = ['security', 'authentication'];
      const missingWords = requiredWords.filter(word => !prompt.toLowerCase().includes(word));
      const isValid = missingWords.length === 0;
      return {
        isValid,
        errors: isValid ? [] : [`Prompt must include: ${missingWords.join(', ')}`]
      };
    }
  }
});

Configuration File​

For CLI usage, you can create a configuration file (e.g., config.json):

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

Then use it with the CLI:

npx llm-guard --config config.json "Your prompt here"

Environment Variables​

LLM Guard also supports configuration through environment variables:

# Enable specific guards
LLM_GUARD_PII=true
LLM_GUARD_JAILBREAK=true
LLM_GUARD_PROFANITY=true

# Configure options
LLM_GUARD_RELEVANCE_MIN_LENGTH=10
LLM_GUARD_RELEVANCE_MAX_LENGTH=5000
LLM_GUARD_TOXICITY_THRESHOLD=0.7

Best Practices​

  1. Enable only the guards you need to improve performance
  2. Adjust sensitivity thresholds based on your use case
  3. Use custom patterns for domain-specific validation
  4. Create custom rules for unique validation requirements
  5. Regularly update your configuration to address new security concerns