How to Evaluate LLM Output Quality Programmatically

When you ship an LLM-powered feature, "does it work?" is not a binary question. An answer can be...