Skip to main content

Evaluating AI Quality (Evals)

Advanced

If you ship anything built on AI, evals are how you know it works — and how you know a change made it better, not worse. Without them you're flying blind: a prompt tweak that helps one case can silently break ten others.

The minimum viable eval​

You don't need a framework to start:

  1. Collect a golden set. 20–100 real inputs with the correct or acceptable outputs (or clear criteria). Cover the easy cases, the tricky ones, and the edge cases that bit you.
  2. Define what "good" means per task — exact match, contains key facts, valid JSON schema, no hallucinated numbers, tone, etc.
  3. Run and score your current setup against the set.
  4. Change one thing (prompt, model, retrieval), re-run, compare. Keep the change only if the score improves.

Choosing metrics​

  • Deterministic checks where possible: schema valid? contains the right value? code passes tests? These are cheap and trustworthy.
  • LLM-as-judge for fuzzy quality (helpfulness, tone): have a model grade outputs against a rubric. Useful but calibrate it — judges have biases (length, position). Validate the judge against human ratings on a sample.
  • Human review for the highest-stakes slice.

When to run them​

  • Before/after any prompt or model change.
  • On model migration — a new model can shift behavior (Errors & Migration).
  • In CI for production systems, as a gate.

:::tip Separate the stages For RAG and agents, eval each stage (did retrieval find the right doc? did the tool get called correctly?) — not just the final answer. It localizes failures. :::

Next​