Guide: Evaluating LLM Outputs in Production with Rubrics and Guardrails

A practical guide details how to evaluate LLM outputs in production using system prompts, scored rubrics, and runtime guardrails. The approach combines LLM-as-a-judge for dimensions like correctness and relevance with filters to catch unsafe outputs. It highlights the Air Canada chatbot case where a hallucinated policy caused legal trouble despite a 200 status code.