How do you build observability for an LLM application, and how does it differ from traditional monitoring?

You cannot improve an LLM app you cannot see into, and LLM observability is not service monitoring. What counts is tracing multi-step chains, capturing inputs/outputs/tokens/cost, and online quality signals, not just latency and errors.

Updated Sep 2026 · Grounded in real GenAI, LLM, and AI/ML engineering interview loops and written to a senior-engineer editorial bar.