Shadow Mode & Canary Deployment for RAG Why Shadow Mode for RAG Offline evaluation diverges from live traffic within weeks because: - Query distribution shifts (product updates, seasonality). - Corpus updates introduce content you never tested against. - Judge LLM scores don't always match user satisfaction. Shadow mode runs a candidate pipeline alongside production without affecting user-facing output. Only after shadow metrics match or beat baseline do you ramp real traffic. Stages | Stage | User Impact | Traffic to Candidate | Purpose | |---|---|---|---| | Shadow (dual-execute) | Zero (can…