Building an eval set from your own traffic

1Where real cases actually come from

An eval case is two things: an input, and the outcome you would accept for it. The input half is the part people get stuck on, because it feels like it needs to be written. It does not - it needs to be collected.

Your logs already contain the inputs. So does your support queue, your chat history, and the running list of "wait, what did the user actually type there" moments your team has had in Slack. Pull real requests, read them, and write down next to each one what a good response would have looked like. That second step is the part that turns a log line into a test case, and it is manual - nobody else can decide what "acceptable" means for your product.

Three sources, and they are not interchangeable:

Start with traffic-sourced cases. They are the cheapest to get and the hardest to argue with, because nobody has to guess whether a real user would actually type that - one already did.

That doesn't mean expert-written cases are wasted effort. A domain expert who has spent years on the support desk knows shapes of questions that haven't shown up in your logs yet simply because the feature they'd apply to only launched last week. Use them to cover the near future your traffic hasn't caught up to. What you're avoiding is treating expert-written cases as a substitute for real ones rather than a supplement to them - the two blind spots stack differently, and traffic-sourced cases should always be the larger share of the set.

Do not skip the parts of your traffic that are unpleasant to read. The empty input, the message in a different language, the multi-turn conversation that goes sideways in the third message, the question phrased as a complaint - these are disproportionately where prompt changes break something, precisely because they are the cases people leave out of a five-minute manual check.

2Every incident becomes a permanent case

This is the single habit that keeps an eval set useful instead of stale, and it costs nothing extra because the work is already being done.

When something breaks in production - a wrong answer, a policy violation, a support escalation that traces back to your system - somebody is already investigating it. The only new step is: once you understand what went wrong, write the input and the correct output down as a case, and add it to the set. Permanently. Not "until it's fixed," not "for this sprint" - forever, the same way a regression test in ordinary software stays in the suite long after the bug it caught is gone.

A production incident you don't turn into a test case is a regression you're volunteering to repeat.

This is what separates a real eval set from a folder of examples someone built once during a sprint and never opened again. The folder goes stale the day traffic shifts. A set that absorbs every incident keeps pace with exactly the failures that have already cost you something, which are also the failures a stakeholder will ask about by name the next time they recur.

3Filling real gaps, carefully

Real traffic has one honest limitation: it only shows you what has already happened. If you are about to launch a feature nobody has used yet, or you know a category of adversarial input exists but it has not shown up in your logs, traffic alone will not cover it. That is the legitimate use for synthetic generation - not as a shortcut to avoid reading logs, but as a way to fill a hole real traffic hasn't gotten around to producing.

The rule is the same one you'd apply to an intern's first draft: useful as a starting point, not usable unread. Generate candidates, then have a person read every one before it goes in the set, because an unchecked synthetic case tests your system against a model's guess at what's hard, and any blind spot in that guess becomes an invisible blind spot in your eval set too.

import json
import random

# Pull a sample of recent, real requests for a human to review and label.
# This does not write the eval set - it drafts candidates for the day's review pass.

def sample_candidates(log_path: str, n: int = 20) -> list[dict]:
    with open(log_path) as f:
        requests = [json.loads(line) for line in f]

    # Skew toward the awkward ones instead of a plain random sample:
    # empty, very short, or very long inputs are where prompt changes tend to break.
    def is_awkward(r: dict) -> bool:
        length = len(r["input"])
        return length == 0 or length < 5 or length > 800

    awkward = [r for r in requests if is_awkward(r)]
    ordinary = [r for r in requests if not is_awkward(r)]

    picked = awkward[: n // 2] + random.sample(ordinary, min(n - len(awkward[: n // 2]), len(ordinary)))
    return [{"input": r["input"], "expected": None} for r in picked]  # a human fills in "expected"

candidates = sample_candidates("logs/requests.jsonl", n=20)
with open("evals/needs_review.jsonl", "w") as f:
    for c in candidates:
        f.write(json.dumps(c) + "\n")

Nothing in that script decides what a good answer looks like - it only surfaces candidates worth a human's time, weighted toward the inputs most likely to be revealing. If retrieval is part of your system, note that scoring whether the right passage came back is its own narrower measurement, covered in Measuring retrieval - what you're building here is a set that judges the whole system's output, not one pipeline stage.

None of this needs to be finished before it's useful. A set of thirty real cases, checked by a person and run today, catches more than the four-hundred-case set you keep meaning to get around to next quarter. Start with what you can label in an afternoon, run it on the next change, and let the incident rule and the occasional synthetic gap-filler grow it from there. The set that exists beats the set that's still being planned.

Once you have this set - traffic-sourced, incident-backed, a handful of synthetic cases filling real gaps - you have something worth running before every change. The next lesson, Offline evals and online evals, is about when to run it, and what it still cannot tell you even when every case passes.