output filtering

Output filtering post-processes LLM responses to remove harmful, sensitive, or policy-violating content before delivery. **What to filter**: Toxic/harmful content, PII leakage, confidential information, off-brand responses, hallucinated claims, competitor mentions, unsafe instructions. **Approaches**: **Classifier-based**: Train models to detect violation categories, block or flag violations. **Regex/rules**: Catch specific patterns (SSN formats, internal URLs, profanity). **LLM-as-judge**: Use another model to evaluate response appropriateness. **Content moderation APIs**: OpenAI moderation, Perspective API, commercial services. **Actions on detection**: Block entire response, redact specific content, regenerate with constraints, escalate for review. **Trade-offs**: False positives frustrate users, latency from additional processing, sophisticated attacks may evade filters. **Layered defense**: Combine with input sanitization, RLHF training, system prompts. **Production considerations**: Log filtered content for analysis, monitor filter rates, tune thresholds per use case. **Best practices**: Defense in depth, graceful degradation, transparency about filtering policies. Critical for customer-facing applications.