Pydantic Evals

Evaluating Non-Deterministic Functions with pydantic-evals

pydantic-evals is a code-first framework for evaluating stochastic functions (LLM calls, agents, pipelines). Define test cases, run them against a task function, and score results with evaluators.

Install: pip install pydantic-evals (or pip install 'pydantic-evals[logfire]' for Logfire integration).

Import Reference

from pydantic_evals import Case, Dataset, set_eval_attribute, increment_eval_metric
from pydantic_evals.evaluators import (
    Evaluator, EvaluatorContext, EvaluatorOutput, EvaluationReason,
    ReportEvaluator, ReportEvaluatorContext,
    LLMJudge, HasMatchingSpan,
)
from pydantic_evals.evaluators.common import Equals, EqualsExpected, Contains, IsInstance, MaxDuration
from pydantic_evals.otel import SpanQuery               # requires logfire extra
from pydantic_evals.generation import generate_dataset  # LLM-based dataset generation

Data Model

Dataset -> Cases -> Evaluators -> EvaluationReport. A Dataset holds Case objects and dataset-wide evaluators. Calling dataset.evaluate(task_fn) runs the task against all cases and returns an EvaluationReport. Both Case and Dataset are generic: Case[InputsT, OutputT, MetadataT].

Case

case = Case(
    name="simple",                           # identifier (optional, but recommended)
    inputs="What is the capital of France?", # any type — passed to the task function
    expected_output="Paris",                 # optional — available via ctx.expected_output
    metadata={"difficulty": "easy"},         # optional — available via ctx.metadata
    evaluators=(MyEvaluator(),),             # optional — case-specific evaluators
)

Dataset

dataset = Dataset(
    cases=[case1, case2],
    evaluators=[GlobalEvaluator()],          # applied to every case
    report_evaluators=[MyReportEvaluator()], # experiment-wide analysis (optional)
)

EvaluatorContext

Every evaluator receives an EvaluatorContext:

Writing Evaluators

Subclass Evaluator and implement evaluate (sync or async). Must use @dataclass decorator.

Return Types

evaluate returns EvaluatorOutput:

• bool — pass/fail (stored in ReportCase.assertions)
• int/float — numeric score (stored in ReportCase.scores)
• str — label (stored in ReportCase.labels)
• EvaluationReason(value, reason) — any of the above with an explanation
• dict[str, ...] — multiple named columns from a single evaluator (see Multi-Score Evaluators)

Single-scalar returns use the evaluator class name as the report column name (override with evaluation_name field).

@dataclass
class ContainsExpected(Evaluator[str, str]):
    def evaluate(self, ctx: EvaluatorContext[str, str]) -> EvaluationReason:
        if ctx.expected_output is None:
            return EvaluationReason(value=False, reason="No expected output provided")
        found = ctx.expected_output.lower() in ctx.output.lower()
        return EvaluationReason(value=found, reason=f"{'found' if found else 'not found'}")

Built-in Evaluators

Multi-Score Evaluators (Dict Returns)

When evaluate returns a dict, each key becomes a separate named column in the report. This lets a single evaluator produce multiple independent scores, assertions, or labels from one pass. Values are categorized by type (bool -> assertions, int/float -> scores, str -> labels, EvaluationReason -> unwrapped by inner .value type).

@dataclass
class QualityEvaluator(Evaluator[QAInput, QAOutput]):
    """Single evaluator that produces multiple report columns."""

    def evaluate(self, ctx: EvaluatorContext[QAInput, QAOutput]) -> dict[str, EvaluationReason | bool | float]:
        output = ctx.output.answer
        return {
            "is_nonempty": len(output.strip()) > 0,                          # -> assertions
            "answer_length": float(len(output)),                             # -> scores
            "contains_expected": EvaluationReason(                           # -> assertions (bool value)
                value=ctx.expected_output is not None
                    and ctx.expected_output.answer.lower() in output.lower(),
                reason=f"Output: {output[:50]}",
            ),
            "verbosity": EvaluationReason(                                   # -> scores (float value)
                value=min(len(output) / 100, 1.0),
                reason="Normalized length score",
            ),
        }

Complete Example

import asyncio
from dataclasses import dataclass

from pydantic_evals import Case, Dataset
from pydantic_evals.evaluators import Evaluator, EvaluatorContext, EvaluationReason


@dataclass
class QAInput:
    question: str


@dataclass
class QAOutput:
    answer: str


@dataclass
class AnswerContainsExpected(Evaluator[QAInput, QAOutput]):
    def evaluate(self, ctx: EvaluatorContext[QAInput, QAOutput]) -> EvaluationReason:
        if ctx.expected_output is None:
            return EvaluationReason(value=False, reason="No expected output")
        found = ctx.expected_output.answer.lower() in ctx.output.answer.lower()
        return EvaluationReason(value=found)


async def my_agent(inputs: QAInput) -> QAOutput:
    # Replace with your actual agent/LLM call
    return QAOutput(answer=f"The answer to '{inputs.question}' is 42.")


async def main():
    dataset = Dataset(
        cases=[
            Case(
                name="capital",
                inputs=QAInput(question="What is the capital of France?"),
                expected_output=QAOutput(answer="Paris"),
            ),
            Case(
                name="color",
                inputs=QAInput(question="What color is the sky?"),
                expected_output=QAOutput(answer="blue"),
            ),
        ],
        evaluators=[AnswerContainsExpected()],
    )

    report = await dataset.evaluate(my_agent)
    report.print(include_input=True, include_output=True)


if __name__ == "__main__":
    asyncio.run(main())

Per-Case Evaluators

Cases can carry their own evaluators via evaluators=(...). Dataset-wide evaluators run on every case; case-specific ones run only on that case. Both appear in the report.

dataset = Dataset(
    cases=[
        Case(
            name="fast_lookup",
            inputs=QAInput(question="What is 2+2?"),
            expected_output=QAOutput(answer="4"),
            evaluators=(MaxDuration(seconds=1.0),),  # only this case must be fast
        ),
        Case(
            name="complex_reasoning",
            inputs=QAInput(question="Explain quantum entanglement simply."),
            expected_output=None,
            evaluators=(
                LLMJudge(rubric="The explanation should be accurate and accessible to a layperson."),
            ),
        ),
    ],
    evaluators=[AnswerContainsExpected()],  # applied to ALL cases
)

# Or add to a specific case after construction:
dataset.add_evaluator(MaxDuration(seconds=2.0), specific_case="fast_lookup")

Report Evaluators

Report evaluators analyze results across all cases after case-level evaluation finishes. They receive a ReportEvaluatorContext with access to ctx.report.cases.

@dataclass
class PassRate(ReportEvaluator[QAInput, QAOutput]):
    threshold: float = 0.8

    def evaluate(self, ctx: ReportEvaluatorContext[QAInput, QAOutput]) -> dict[str, float]:
        total = len(ctx.report.cases)
        passed = sum(1 for c in ctx.report.cases if c.assertions.get("AnswerContainsExpected"))
        rate = passed / total if total else 0.0
        return {"pass_rate": rate, "meets_threshold": float(rate >= self.threshold)}

dataset = Dataset(cases=[...], evaluators=[...], report_evaluators=[PassRate(threshold=0.9)])

Reporting

evaluate / evaluate_sync return an EvaluationReport containing:

• cases: list[ReportCase] — successful results, each with scores (float), labels (str), assertions (bool), metrics, task_duration, total_duration
• ReportCase also includes inputs, output, expected_output, and metadata
• failures: list[ReportCaseFailure] — failed cases with error_message and error_stacktrace
• ReportCaseFailure also includes inputs and expected_output
• analyses: list[ReportAnalysis] — report-level analyses (confusion matrices, precision-recall, etc.)

report.print(include_input=True, include_output=True, include_durations=False)
report.render()          # returns formatted string instead of printing
report.case_groups()     # grouped results when using repeat > 1
report.averages()        # aggregated statistics when using repeat > 1

YAML Datasets

dataset.to_file("my_cases.yaml")

dataset = Dataset[QAInput, QAOutput].from_file(
    "my_cases.yaml",
    custom_evaluator_types=(AnswerContainsExpected,),  # required for custom evaluator deserialization
)

Evaluate Options

report = await dataset.evaluate(
    my_agent,
    max_concurrency=5,      # limit parallel case execution
    repeat=3,               # run each case N times, results grouped by case name
    retry_task=2,           # retry task on failure
    retry_evaluators=1,     # retry evaluators on failure
    metadata={"run": "v2"}, # experiment-level metadata
)

Dataset Generation

dataset = await generate_dataset(
    dataset_type=Dataset[QAInput, QAOutput],
    n_examples=10,
    model="openai:gpt-4o",
    extra_instructions="Focus on geography questions of varying difficulty.",
    path="generated_cases.yaml",  # optionally persist to file
)

Always review generated cases — treat them as a starting point, not ground truth.

Span-Based Evaluation

Assert on internal agent behavior via OpenTelemetry traces (requires logfire extra):

Case(
    name="uses_tool",
    inputs=QAInput(question="What's the weather in Paris?"),
    evaluators=(HasMatchingSpan(query=SpanQuery(name_contains="weather_api")),),
)

The full span_tree is also available in custom evaluators via ctx.span_tree.

Runtime Attributes and Metrics

Record data inside the task function that evaluators can access via ctx.attributes and ctx.metrics:

async def my_agent(inputs: QAInput) -> QAOutput:
    set_eval_attribute("model_used", "gpt-4o")
    increment_eval_metric("llm_calls", 1)
    return QAOutput(answer="...")

Common Pitfalls

Always explicitly set generic parameters on Dataset. When serializing or deserializing, pydantic-evals needs to know the concrete types. Use Dataset[InputsT, OutputT, MetadataT] instead of bare Dataset.

# WRONG — triggers "Could not determine the generic parameters" warning
dataset = Dataset(cases=[...])

# CORRECT
dataset = Dataset[MyInput, MyOutput, MyMetadata](cases=[...])

Custom evaluators must use @dataclass. Without it, serialization and construction break silently.

# WRONG
class MyEval(Evaluator[str, str]):
    def evaluate(self, ctx: EvaluatorContext[str, str]) -> bool: ...

# CORRECT
@dataclass
class MyEval(Evaluator[str, str]):
    def evaluate(self, ctx: EvaluatorContext[str, str]) -> bool: ...

Always handle expected_output is None. Cases may omit it. Evaluators that access ctx.expected_output without a guard will crash.

The task function receives the full input object, not unpacked fields:

# WRONG
async def my_agent(query: str, context: str) -> str: ...

# CORRECT — receives the dataclass
async def my_agent(inputs: AgentInput) -> str:
    return f"Answer to {inputs.query} given {inputs.context}"