No backend team would merge a PR that fails the test suite. Yet many AI teams ship prompt and model changes with no automated quality check at all — they eyeball a few responses and hope. The failure mode is silent: answers get slightly less faithful, citations drift, and nobody notices until a user does.
Atlas (my enterprise RAG copilot project) treats evals exactly like tests: a merge is blocked on a quality or cost regression. Here’s the shape of that pipeline and what I learned building it.
What gates the build
Every PR runs an offline, deterministic, GPU-free eval gate against committed cassettes:
Metric Baseline Floor Policy RAGAS faithfulness 0.706 0.656 blocking, regression band 0.05 Answer relevancy 0.832 0.782 blocking Context recall 0.738 0.668 blocking, band 0.07 Citation correctness 1.000 — tracked Adversarial suite (injection) 100% pass 100% zero-tolerance, blocking Cost per request committed baseline reduction target blocking (separate cost gate)Three design decisions made this workable:
1. Floors + regression bands, not perfection. A hard “faithfulness ≥ 0.9” gate would have blocked every merge forever. Instead: an absolute floor (never ship below this) plus a regression band relative to the committed baseline (never get meaningfully worse). This is the eval equivalent of ratcheting code coverage.
2. Deterministic and GPU-free in CI. Live-model evals in CI are slow, expensive, and flaky. Atlas replays committed cassettes — recorded model interactions — so the gate is fast and reproducible. Cassettes are re-recorded deliberately when behavior is supposed to change, which turns “the model changed” into a reviewed diff instead of an ambient surprise.
3. Cost is a gated metric, not a dashboard. A prompt change that doubles token usage is a regression even if quality holds. Atlas’s cost gate asserts the measured cost-reduction against a committed baseline and fails the build when cost regresses below the band — the same way a p95-latency budget would gate a backend service.
The payoff: decisions you can defend
The gate earned its keep when I fine-tuned a QLoRA adapter for citation formatting. The benchmark (paired bootstrap, McNemar) showed: citation format validity 0.00 → 0.955 (p ≈ 0), serving cost −79% on the same GPU — and a faithfulness trade-off of −0.10, still above the floor. The promotion gate had the data to make that call explicitly, and the pipeline is proven to both promote and block a candidate in CI.
Without evals, that’s a vibes decision. With them, it’s an engineering decision with a paper trail.
Where to start (smaller than you think)
You don’t need RAGAS on day one. Start with:
- 20–30 golden questions with expected source documents — your retrieval hit-rate suite.
- A handful of adversarial prompts (injection attempts, out-of-scope questions) with expected refusals.
- A CI job that fails when either regresses.
That’s an afternoon of work, and it changes the culture immediately: prompts become code, model swaps become reviewed changes, and “did we get worse?” has an answer.
Backend engineering solved this problem twenty years ago — we just called it testing. AI systems don’t need a new philosophy; they need the old one applied.
답글 남기기