GenAI & Agents

GenAI Evaluation Tests

A test set from your experts that scores every assistant change before it ships.

Download the PDF

Get the GenAI Evaluation Tests overview

A short PDF to share with your team. It covers what is in the package, how it works, and a worked example.

The problem

Nobody can say whether the last prompt change made things worse.

Prompts, models, and retrieval settings change all the time. Without a test set, quality gets judged by a few people trying a few questions, and users find the regressions first.

What's included

Every release scored before it ships.

01

Reference test set

Real questions with answers written by your subject-matter experts, plus edge cases. Cases are YAML or JSONL and can require specific text, specific sources, or a refusal.

02

Scorers

Accuracy, source use, refusals, leaks, personal data, safety, latency, and cost for every answer. Built-in checks need no model. An optional LLM grader runs on Azure OpenAI.

03

Red-team and injection tests

An adversarial suite we extend with cases for your own tools, data, and known risks.

04

Release gates in CI

Pipelines for GitHub Actions and Azure DevOps use workload identity, so no keys are stored. Each run is compared with the last, and reports come out as Markdown, JSON, and JUnit XML.

How it works

How a release is gated.

  1. Build the set

    Start from real questions in logs, tickets, or a pilot, with personal data stripped out.

  2. Cover the categories

    Accuracy, grounding, refusal, permissions, injection, and safety. Critical cases are marked.

  3. Plant canaries

    Put unique strings in restricted documents and the system prompt. Any answer that contains one is a leak.

  4. Set gates

    Pass rates overall and by category, limits on cost and latency, and the largest drop allowed against the last release.

  5. Run on every change

    The pipeline scores the change, publishes the report, and blocks the merge when a gate fails.

  6. Add a case for every bug

    A bad answer seen in production goes into the set before it gets fixed.

Worked example

Release gates agreed with the business owner.

Thresholds are set in the Ground step and written into the system’s Evaluation Card.

The release fails if
  • The overall pass rate drops below 95%
  • Accuracy cases fall below 90%
  • Any permission or injection case fails
  • Average cost or p95 latency goes over its limit
  • Results drop more than 2 points from the last release
Result

A failing critical case or a case that errors also fails the run. The report lists every case that passed last time and fails now.

Where it fits

What it runs on, and where it stops.

Works with

  • Any assistant or agent you can call from Python or HTTP, including an Azure OpenAI deployment on its own
  • Azure OpenAI for the optional grader
  • GitHub Actions and Azure DevOps

Assumptions and limits

  • Text checks are exact. That suits facts and leaks and is blunt for tone and style
  • Refusal detection matches phrases, so we add your assistant’s own wording
  • The red-team suite stays generic until we add your risks
Ask Dex

Ask about the GenAI Evaluation Tests

Dex answers from our own pages and links to where it found the answer.

Changing prompts without a safety net?

Send us a handful of real questions your assistant gets. We will show you how they become test cases.