llm_eval 0.4.0

A Dart test harness for LLM evals: assertion checks over model outputs, an optional LLM-as-judge, and cached responses so CI stays deterministic.

llm_eval

A test harness for LLM outputs in Dart. Write eval cases the way you write unit tests: a prompt, a list of checks, and a report you can read in CI.

llm_eval does not talk to any model provider. You hand it one function that takes a prompt and returns the model output, and it handles running cases concurrently, checking outputs, caching responses, and reporting.

Why

Dart and Flutter apps that call LLMs usually have no automated way to catch a prompt regression. Model output is not exact, so plain string equality in a unit test does not work. This package covers the middle ground. It gives you assertion-style checks for the properties that must hold. For the fuzzy parts, there is an optional LLM judge. A response cache keeps CI runs deterministic and free.

Quick start

import 'package:llm_eval/llm_eval.dart';

Future<void> main() async {
  // Bind your model. Any Future<String> Function(String prompt) works:
  // an OpenAI or Anthropic SDK call, an Ollama HTTP request, a fake.
  Future<String> model(String prompt) async {
    // ... call your model client here ...
    return 'The capital of France is Paris.';
  }

  final suite = EvalSuite([
    EvalCase(
      id: 'capital-question',
      prompt: 'What is the capital of France?',
      checks: [
        Check.contains('paris'),
        Check.notContains('berlin'),
      ],
    ),
    EvalCase(
      id: 'structured-output',
      prompt: 'Return a JSON object with a "city" field.',
      checks: [
        Check.isValidJson(where: (v) => v is Map && v.containsKey('city')),
      ],
    ),
  ]);

  final report = await suite.run(model, modelId: 'my-model');
  print(report.toMarkdown());
}

Run the bundled example with dart run example/llm_eval_example.dart.

Checks

Check	Passes when
Check.contains(s)	the output contains s (case-insensitive by default)
Check.notContains(s)	the output does not contain s
Check.matches(regExp)	the pattern matches the output
Check.isValidJson(where: f)	the output parses as JSON (stripping a single wrapping markdown code fence first, if the raw output does not parse) and the optional condition holds
Check.predicate(desc, f)	your function returns true
Check.judge(...)	another model scores the output at or above passAt

Custom checks implement the Check interface: a description for reports and an evaluate method that returns a CheckResult.

Fails and errors are distinct throughout. A fail means the check ran and the output did not satisfy it. An error means no verdict was possible (the judge response could not be parsed, a callback threw, the model call failed). Reports show them separately, so a broken harness is not mistaken for a failing model.

LLM as judge

For properties that plain checks cannot express, ask another model to grade the output against a rubric:

Check.judge(
  judge: judgeModel, // often a stronger model than the one under test
  rubric: 'The answer names Paris and stays under three sentences.',
  passAt: 0.7,
)

The judge receives the rubric and the output in a fixed prompt and must answer with a SCORE: <number> line between 0.0 and 1.0. A response that cannot be parsed becomes an error result, never a silent pass or fail. So does a response with conflicting score lines, which is what a graded output that smuggles in its own SCORE: 1.0 line tends to produce. The graded output is wrapped in delimiters the judge is told to respect; this raises the bar against prompt injection without eliminating it.

One honest caveat: the judge is itself an LLM. Its scores are not calibrated, they drift across judge models and versions, and they can be wrong. Use judges sparingly, pin the judge model, and spot-check its verdicts against your own reading.

The judge is a nested model call that suite.run does not cache: on a warm cache the model under test is skipped but an unwrapped judge fires on every run. Wrap the judge in the same cache so it is cached too:

Check.judge(
  judge: cache.wrap(judgeModel, modelId: 'judge-v1'),
  rubric: 'The answer names Paris and stays under three sentences.',
)

cache.wrap uses the same key scheme as the suite, so the judge's responses sit alongside the model responses in one cache directory. Give each judge its own modelId so pinning or changing a judge re-records only its own responses.

Caching and CI

The core library is pure Dart and runs on every platform, including the web. FileResponseCache needs dart:io, so it lives in a separate library:

import 'package:llm_eval/llm_eval.dart';
import 'package:llm_eval/io.dart' show FileResponseCache;

final cache = FileResponseCache('test/llm_cache');
final report = await suite.run(model, cache: cache, modelId: 'my-model-v1');

The first run calls the model and stores each response in a file named after the SHA-256 of the model id and prompt. Later runs read the cache and never call the model. Commit the cache directory and your CI eval job is deterministic, offline, and free. Delete the directory, or change modelId, to re-record.

The cache covers the model under test. A nested call, such as the judge in a Check.judge, is not cached by suite.run, so a warm cache still calls the judge unless you wrap it with cache.wrap(judgeModel, modelId: ...) (see LLM as judge). The Markdown report's cached column likewise describes the model under test, not any nested judge.

A regression test then looks like any other test:

test('prompt regression suite', () async {
  final report = await suite.run(
    model,
    cache: FileResponseCache('test/llm_cache'),
    modelId: 'my-model-v1',
  );
  expect(report.results, isNotEmpty);
  expect(report.errorCount, 0, reason: report.toMarkdown());
  expect(report.passRate, 1.0, reason: report.toMarkdown());
});

The same shape works as a standalone CI gate:

final report = await suite.run(model, cache: cache, modelId: 'my-model-v1');
stdout.writeln(report.toMarkdown());
if (report.results.isEmpty || report.errorCount > 0 || report.passRate < 1.0) {
  exitCode = 1;
}

Check errorCount separately from the pass rate: errors mean the harness could not produce a verdict (a judge response failed to parse, a callback threw), not that the model answered badly. Note that an empty suite has a pass rate of 1.0, so guard against accidentally building zero cases, as both snippets above do.

Showing results in the CI UI

An exit code tells the build to go red; toJUnitXml tells it which cases went red and why. Write the report where your CI looks for test results and each eval case appears as a test, failing ones expanded to the checks that failed and the model output that failed them:

File('eval-results.xml').writeAsStringSync(report.toJUnitXml());

# GitHub Actions
- run: dart run tool/eval.dart
- uses: dorny/test-reporter@v1
  if: always()
  with:
    path: eval-results.xml
    reporter: java-junit

GitLab, Jenkins, CircleCI and Buildkite read the same format. A case with a model error or an errored check becomes an <error>, any other non-passing case a <failure>, and a flaky case counts as failing with a message saying how many attempts passed. Model output is arbitrary text, so it is escaped and characters XML cannot carry are dropped; a single stray control byte would otherwise make the report unreadable to the CI system.

Repeat and flakiness

final report = await suite.run(model, repeat: 5);
print(report.flakinessRate);

Each case runs five times and the report exposes the fraction of cases whose attempts disagree. Measure flakiness without a cache: a warm cache returns the same response every time.

Reports

EvalReport.toMarkdown() renders a summary table plus details for every non-passing case, suitable for a CI job summary. EvalReport.toJson() returns a JSON-compatible map with the full result tree (outputs, per-check verdicts, scores, latencies, cache hits) for your own tooling.

Planned

Out of scope for 0.1 and planned for later releases:

• side-by-side comparison of several models over one suite
• token and cost accounting
• dataset loaders for existing eval formats
• structured prompts: system messages and multi-turn conversations

License

MIT