Content Guards

toxicity

Profanity, hate speech, threats, and harassment detection.

toxicity({
  action: 'block' | 'warn',
  categories?: ('profanity' | 'hate' | 'threat' | 'harassment')[],
  threshold?: number,  // 0-1, default: 0.5
})

bias

Gender, racial, religious, and age bias detection.

bias({ action: 'block' | 'warn' })

language

Restrict text to specific languages.

language({
  allowed: string[],  // e.g. ['en', 'ko']
  action: 'block' | 'warn',
})

sentiment

Emotional tone control.

sentiment({ action: 'warn' })

copyright

Detect copyright notices, trademarks, and verbatim reproduction.

copyright({ action: 'block' | 'warn' })

urlGuard

URL validation and filtering.

urlGuard({ action: 'block' | 'warn' })

repetitionDetect

Detect repetitive patterns in LLM output (stuck loops, degenerate text).

repetitionDetect({ action: 'warn' })

responseQuality

Check LLM response quality — too short, highly repetitive, or refusal patterns.

responseQuality({
  action: 'block' | 'warn',
  minLength?: number,           // min chars, default: 10
  maxRepetitionRatio?: number,  // 0-1, default: 0.5
  detectRefusal?: boolean,      // detect "As an AI...", default: true
})

Detects:

• Empty or too-short responses
• Highly repetitive sentences (>50% duplicated)
• Common refusal patterns ("I'm sorry, but I cannot...", "As an AI language model...")

languageConsistency

Verify that the LLM response language matches the expected language.

languageConsistency({
  action: 'block' | 'warn',
  expected?: string[],  // e.g. ['ko', 'en']
})

Supports: ko, en, ja, zh, es, fr, de

If expected is not set, reads from ctx.metadata.inputLanguage. Useful for ensuring Korean users get Korean responses.