Skip to content

Human review that scales

Queues, keyboard shortcuts and reason codes: how to keep ground truth fresh without building up a review backlog every week.

Relevant teamProduct

6 min read

On this page
  1. Feed the queue from failures and policy changes
  2. Four actions are enough
  3. Make it fast

Every evaluation is only as good as its gold answers, and gold answers go stale whenever a policy changes. The teams that keep ground truth fresh treat review as a queue with a service level, not as a quarterly project.

Feed the queue from failures and policy changes

Two sources produce most review work: failures where the judge and the gold answer disagree, and knowledge-base updates that invalidate existing gold answers. Route both into one queue with the source article and section attached.

Four actions are enough

Reviewers need to approve, edit, reject or flag. Anything more slows the queue down.

  • Approve: the proposed gold answer is correct and complete.
  • Edit: fix the answer and record a reason code (policy change, factual error, incomplete).
  • Reject: out of scope, ambiguous or duplicate.
  • Flag: needs a policy owner or legal decision.

Make it fast

Keyboard shortcuts (J and K to move, A to approve, E to edit, R to reject) and the source excerpt shown next to the answer bring median review time under three minutes. Show progress against the queue size so reviewers see the backlog shrink.

Share
XLinkedIn
All articles
    • Failure analysis
    • Process

    A failure taxonomy for support assistants

    Six categories, clear detection signals and one owner per category: a failure taxonomy your support team will actually use.

    8 min read

    • Model comparison
    • Evaluation

    Choosing a generator with evidence

    Comparing models on the same retrieval, the same prompt and the same questions, then reading the per-category differences.

    7 min read

    • Evaluation
    • Retrieval

    Why aggregate scores hide retrieval failures

    A single accuracy number can rise while the answers your customers care about get worse. Here is how to break the number apart.

    9 min read

See why your retrieval fails

Run an evaluation on your own dataset and read every failure with its retrieved chunks, flagged claims and suggested fix.

Free plan includes 2,000 evaluated samples a month