Per-Component Evaluation

~40 min read

How to decompose LLM systems and evaluate each component independently with the right metric.

Per-Component Evaluation is a Pro topic

Sign in, then upgrade to Pro or Power to unlock this topic and the full AI Engineering curriculum.

Key points

  • Each component can fail independently; end-to-end eval cannot isolate which stage caused the failure
  • Use automated metrics (Recall@K, MRR) for retrieval; AI-as-Judge for generation quality
  • Inject perfect context to test generation in isolation from retrieval errors