LLM observability is the practice of monitoring and understanding how large language model applications behave in production: what they output, why, and where they fail. It extends traditional observability (logs, metrics, traces) to signals specific to LLMs, such as prompts, responses, latency, cost, and quality problems like hallucination or drift.
In practice, teams capture each prompt and response, score output quality, and set alerts when behavior changes. For example, a support assistant might be tracked for rising hallucination rates after a model or data update, so the team catches the regression before users do.
LLM observability tells you that something changed. Reproducing the exact data state behind a run that worked, so you can see what changed instead of guessing, is a separate question.