pydantic-evals 1.81.0

Framework for evaluating stochastic code execution, especially code making use of LLMs

Pydantic Evals

This is a library for evaluating non-deterministic (or "stochastic") functions in Python. It provides a simple, Pythonic interface for defining and running stochastic functions, and analyzing the results of running those functions.

While this library is developed as part of Pydantic AI, it only uses Pydantic AI for a small subset of generative functionality internally, and it is designed to be used with arbitrary "stochastic function" implementations. In particular, it can be used with other (non-Pydantic AI) AI libraries, agent frameworks, etc.

As with Pydantic AI, this library prioritizes type safety and use of common Python syntax over esoteric, domain-specific use of Python syntax.

Full documentation is available at ai.pydantic.dev/evals.

Example

While you'd typically use Pydantic Evals with more complex functions (such as Pydantic AI agents or graphs), here's a quick example that evaluates a simple function against a test case using both custom and built-in evaluators:

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

# Define a test case with inputs and expected output
case = Case(
    name='capital_question',
    inputs='What is the capital of France?',
    expected_output='Paris',
)

# Define a custom evaluator
class MatchAnswer(Evaluator[str, str]):
    def evaluate(self, ctx: EvaluatorContext[str, str]) -> float:
        if ctx.output == ctx.expected_output:
            return 1.0
        elif isinstance(ctx.output, str) and ctx.expected_output.lower() in ctx.output.lower():
            return 0.8
        return 0.0

# Create a dataset with the test case and evaluators
dataset = Dataset(
    name='capital_eval',
    cases=[case],
    evaluators=[IsInstance(type_name='str'), MatchAnswer()],
)

# Define the function to evaluate
async def answer_question(question: str) -> str:
    return 'Paris'

# Run the evaluation
report = dataset.evaluate_sync(answer_question)
report.print(include_input=True, include_output=True)
"""
                                    Evaluation Summary: answer_question
┏━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━┓
┃ Case ID          ┃ Inputs                         ┃ Outputs ┃ Scores            ┃ Assertions ┃ Duration ┃
┡━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━┩
│ capital_question │ What is the capital of France? │ Paris   │ MatchAnswer: 1.00 │ ✔          │     10ms │
├──────────────────┼────────────────────────────────┼─────────┼───────────────────┼────────────┼──────────┤
│ Averages         │                                │         │ MatchAnswer: 1.00 │ 100.0% ✔   │     10ms │
└──────────────────┴────────────────────────────────┴─────────┴───────────────────┴────────────┴──────────┘
"""

Using the library with more complex functions, such as Pydantic AI agents, is similar — all you need to do is define a task function wrapping the function you want to evaluate, with a signature that matches the inputs and outputs of your test cases.

Logfire Integration

Pydantic Evals uses OpenTelemetry to record traces for each case in your evaluations.

You can send these traces to any OpenTelemetry-compatible backend. For the best experience, we recommend Pydantic Logfire, which includes custom views for evals:

You'll see full details about the inputs, outputs, token usage, execution durations, etc. And you'll have access to the full trace for each case — ideal for debugging, writing path-aware evaluators, or running the similar evaluations against production traces.

Basic setup:

import logfire

logfire.configure(
    send_to_logfire='if-token-present',
    environment='development',
    service_name='evals',
)

...

my_dataset.evaluate_sync(my_task)

Read more about the Logfire integration here.