Production Checklist

Before deploying Director-AI to production, verify each item:

Scoring

• Set use_nli=True with a production NLI model (FactCG-DeBERTa-v3-Large recommended)
• Configure ground_truth_store with your domain knowledge base
• Set appropriate threshold per domain — see Domain Recommendations
• Enable cache_size > 0 for repeated queries (reduces NLI inference cost)

Observability

• Set up OpenTelemetry with setup_otel() for trace collection
• Enable metrics_enabled=True for Prometheus-compatible metrics
• Configure log_json=True for structured log aggregation

Server

• Use director-ai serve --workers N for multi-worker deployment (N = CPU cores)
• Configure rate limiting: rate_limit_rpm > 0 for public endpoints
• Set api_keys for authentication on exposed endpoints
• Set non-root user in Docker (already done in v2.0.0 Dockerfile)

Benchmarking

• Run director-ai bench to establish baseline metrics before deployment
• Run director-ai bench --dataset e2e to verify catch rate on your domain data
• Review director-ai config --profile <domain> settings match your requirements

Security

• Audit dependencies: pip-audit --strict
• Run SAST: bandit -r src/director_ai/
• Review CORS origins — default * is unsafe for production
• Ensure llm_api_key and api_keys are not logged or exposed in responses