For business
What it is
A fixed set of real examples with known good answers, run automatically to check quality.
Why it matters
Every change, a new prompt or a new model, is tested before it reaches anyone, like unit tests for AI.
Risk without it
Quality drifts silently, and a “better” model quietly breaks an important case.
For engineers
How it works
Golden datasets built from reviewed real cases; scoring with deterministic checks plus an LLM judge with fixed rubrics; runs in CI on every prompt or model change, with thresholds that block a release.
Typical stack
- custom eval harness
- promptfoo / Langfuse evals
- CI pipeline
Pitfalls & trade-offs
- Calibrate the judge against human ratings.
- Refresh the sets from production feedback.
- Test the edge cases, not just the average.
Example pattern · fictional data · not a client system