Quack Test

A plugin for pytest to evaluate non-deterministic agent components.

Installation & Setup

Simply pip install it.

pip install quack-test

To use the LLM judge you will need to add the required config to your .env file.

# OpenAI Provider Configuration
# Set to "OpenAI" or "AzureOpenAI"
OPENAI_PROVIDER=OpenAI

# Unified Configuration
OPENAI_API_KEY=your_api_key_here
OPENAI_MODEL_NAME=gpt-4
OPENAI_ENDPOINT=https://your-resource-name.openai.azure.com/

Run tests

Simply run your tests using pytest (quack_tests are just pytests)

pytest

Writing Tests

You can simply specify @nondeterministic for components, which are not deterministic and should be run multiple times. Implement everything in a test_*.py file in your test or tests folder.

import random
from quack_test import nondeterministic_fixture, nondeterministic_test

@nondeterministic_fixture(n=5)
def sample_text():
    return f"{random.randint(1, 10)} apples."

@nondeterministic_test(threshold=0.8)
def test_simple_assertion(sample_text):
    apples = float(sample_text.split(" "))
    assert apples > 1, f"Expected more than 1 apple found {apples}."

In practice choosing n=4 or n=5 is recommended. The threshold should be then 0.75 or 0.8, allowing for a single failure. If you are choosing a bigger n, you are rather running a benchmark than just a test. Here we assume, that if a feature is really broken, it does not only fail once, but more often. On the other side, if a feature works robustly, it will should not fail or at maximum once.

Async Tests

Async test and fixture functions (async def) are supported natively and awaited on a single event loop shared across the whole test session — no manual asyncio.run() wrapping needed, and loop-bound resources (e.g. singleton agent or MCP clients created at module level) keep working across fixtures, runs and tests.

@nondeterministic_fixture(n=5)
async def agent_run():
    return await my_agent.arun("Generate a greeting")

@nondeterministic_test(threshold=0.8)
async def test_agent_greeting(agent_run):
    return judge(agent_run, criterion="Contains 'hello' or 'hi'")

Parallel Execution

Runs are executed sequentially by default. Since LLM calls are I/O-bound and independent, you can fan out with parallel=True. This also works on nondeterministic_fixture for concurrent sample generation.

@nondeterministic_test(threshold=0.8, parallel=True)
def test_judge_with_criterion(sample_text):
    return judge(sample_text, criterion="More than 1 apple")

Early Stopping

With stop_early=True, runs stop as soon as the outcome is provably decided (pass or fail cannot change anymore), saving tokens on expensive evaluations. This assumes scores lie in [0, 1], which is the scale the judge uses. Cannot be combined with parallel=True.

@nondeterministic_test(threshold=0.8, stop_early=True)
def test_judge_with_criterion(sample_text):
    return judge(sample_text, criterion="More than 1 apple")

Samples vs. Lists

Only arguments of type Samples (returned automatically by nondeterministic_fixture) are indexed per run. Plain list arguments (e.g. chat histories) are passed to every run unchanged. To opt into per-run indexing for your own data, wrap it with Samples(...):

from quack_test import Samples

@nondeterministic_test(n=1)  # n from Samples length if -1
def test_messages(history, static_context):
    # history is a Samples parameter: each run gets one conversation
    # static_context is a plain list: passed unchanged to every run
    ...

Filtering With Markers

All nondeterministic_test tests are automatically marked with the nondeterministic marker. To run your plain deterministic test suite without any quack tests:

pytest -m "not nondeterministic"

At the end of a test session, a summary of all quack test outcomes is printed.

Judging via LLM

In many cases it is hard to evaluate with code, if an answer is actually correct. In those cases, use a judge.

from quack_test import judge, nondeterministic_test
# [...]

@nondeterministic_test(threshold=0.8)
def test_judge_with_criterion(sample_text):
    return judge(sample_text, criterion="More than 1 apple")

@nondeterministic_test(threshold=0.8)
def test_judge_with_gt(sample_text):
    return judge(sample_text, gt="N (0-10) apples")

Class Based Tests

Testing in classes can be beneficial in some use-cases. This is not hindered by quack-test.

class TestClassBased:
    @nondeterministic_test(n=3, threshold=0.2)
    def test_assert_only(self):
        # You can also create your own score and explanation
        return random.random(), "The expected value is above 0.2, but was lower."

Negative Tests

Writing negative tests that must fail is important. In quack test, you can simply specify that a test has to fail via should_fail=True.

@nondeterministic_test(threshold=0.8, should_fail=True)
def test_negative(sample_text):
    return judge(sample_text, criterion="Has only coal")

Showcase for Error Messages

The following example code will fail and cause an error message.

@nondeterministic_test(threshold=0.8)
def test_assertion_demo_failure(sample_text):
    assert False, "This message is shown in the pytest summary."
    # -> FAILED test/test_example.py::test_assertion_demo_failure - AssertionError: Test failed to meet success threshold. Score: 0.0 (required: 0.8), Success rate: 0.00% (0/5), run scores: [0.00, 0.00, 0.00, 0.00, 0.00], This message is shown in the pytest summary.

@nondeterministic_test(threshold=0.8)
def test_judge_demo_failure(sample_text):
    return judge(sample_text, criterion="Has only coal")
    # -> FAILED test/test_example.py::test_judge_demo_failure - AssertionError: Test failed to meet success threshold. Score: 0.0 (required: 0.8), Success rate: 0.00% (0/5), run scores: [0.00, 0.00, 0.00, 0.00, 0.00], Text: '3 apples.' Criterion: 'Has only coal'

The failure message (and the session-end quack-test summary) always includes the per-run scores, which helps tune thresholds.

Advanced Judge Setup

If you do not want to setup the judge via the .env file as shown above, you can also set it up via code. In this example we hardcode a local ollama instance. Of course you should NEVER put your api key in the code, but if you manage your secrets differently than in a .env this can come in clutch to pass your secrets to the judge.

def setup_module():
    configure_judge(
        provider="OpenAI",
        api_key="ollama",
        model_name="qwen3:4b",
        endpoint="http://localhost:11434/v1"
    )

Have fun quack testing your agents!