pydantic-evals（迂腐的评价）

Pydantic Evals

Overview

Pydantic Evals provides rigorous testing and evaluation for AI agents and LLM outputs using a code-first approach with Pydantic models. It enables "Evaluation-Driven Development" (EDD) where evaluation suites live alongside application code, subject to version control and CI/CD.

Core Concepts

Understand these key primitives:

Case

A single test scenario with inputs, optional expected output, and metadata.

from pydantic_evals import Case

case = Case(
    name="refund_request",
    inputs="What is your refund policy?",
    expected_output="30 days full refund",
    metadata={"category": "policy"}
)

Dataset

Collection of Cases with default evaluators. Generic over input/output types.

from pydantic_evals import Dataset

dataset = Dataset(
    cases=[case1, case2, case3],
    evaluators=[evaluator1, evaluator2]
)

Evaluator

Logic engine that assesses outputs. Returns bool (Pass/Fail), float/int (score), or str (label).

Experiment

Point-in-time performance capture when Dataset runs against a Task.

For detailed explanations, see references/core-concepts.md

Quick Start

Create and run a simple evaluation:

from pydantic_evals import Case, Dataset
from pydantic_evals.evaluators import Contains, LLMJudge

# Define cases
cases = [
    Case(
        name="greeting",
        inputs="Hello, who are you?",
        expected_output="I am an AI assistant."
    )
]

# Define evaluators
evaluators = [
    Contains(value="AI assistant"),
    LLMJudge(rubric="Is this response polite? Answer PASS or FAIL.")
]

# Create dataset
dataset = Dataset(cases=cases, evaluators=evaluators)

# Run evaluation
async def my_agent(query: str) -> str:
    # Your agent logic here
    return "I am an AI assistant."

report = dataset.evaluate_sync(my_agent)
report.print()

Evaluator Types

Pydantic Evals supports a "Pyramid of Evaluation" from fast/cheap to slow/expensive:

1. Deterministic Evaluators

Fast, free, code-based checks. Use as first line of defense.

• Equals: Exact equality check
• EqualsExpected: Compare to Case.expected_output
• Contains: Substring/item presence
• IsInstance: Type validation
• MaxDuration: Latency SLA enforcement

Strategy: Always run deterministic checks before expensive LLM judges.

2. LLM-as-a-Judge

Use secondary LLM to score outputs based on natural language rubrics.

from pydantic_evals.evaluators import LLMJudge

judge = LLMJudge(
    rubric="Response must: 1) Answer the question, 2) Cite context, 3) Be professional",
    include_input=True,
    include_expected_output=True,
    model='openai:gpt-4o'
)

Using OpenRouter for LLMJudge:

from pydantic_evals.evaluators.llm_as_a_judge import set_default_judge_model
from pydantic_ai.models.openai import OpenAIChatModel
from pydantic_ai.providers.openai import OpenAIProvider

# Configure OpenRouter as judge model
provider = OpenAIProvider(
    api_key=os.getenv('OPENROUTER_API_KEY'),
    base_url='https://openrouter.ai/api/v1'
)
model = OpenAIChatModel(model_name='gpt-4o-mini', provider=provider)
set_default_judge_model(model)

# Or pass model directly to LLMJudge
judge = LLMJudge(rubric="Is this polite?", model=model)

Rubric best practices: Be specific and actionable, not vague.

3. Custom Evaluators

Implement arbitrary logic by inheriting from Evaluator.

from dataclasses import dataclass
from pydantic_evals.evaluators import Evaluator, EvaluatorContext

@dataclass
class ValidSQL(Evaluator):
    def evaluate(self, ctx: EvaluatorContext) -> bool:
        import sqlparse
        try:
            parsed = sqlparse.parse(ctx.output)
            return len(parsed) > 0
        except:
            return False

Custom Evaluators for Structured Output (Pydantic Models)

Important: Built-in evaluators like Contains, Equals work with strings/lists/dicts. They do NOT work with Pydantic model outputs. For agents with output_type=MyModel, create custom evaluators:

from dataclasses import dataclass
from pydantic_evals.evaluators import Evaluator, EvaluatorContext
from pydantic import BaseModel

class MyAgentResponse(BaseModel):
    message: str
    status: str
    complete: bool

@dataclass
class HasNonEmptyMessage(Evaluator[MyAgentResponse, None]):
    """Check that response has a non-empty message field."""
    min_length: int = 1

    def evaluate(self, ctx: EvaluatorContext[MyAgentResponse, None]) -> bool:
        if not isinstance(ctx.output, MyAgentResponse):
            return False
        return len(ctx.output.message) >= self.min_length

@dataclass
class StatusIsValid(Evaluator[MyAgentResponse, None]):
    """Check that status is one of allowed values."""
    allowed_values: tuple = ("pending", "complete", "error")

    def evaluate(self, ctx: EvaluatorContext[MyAgentResponse, None]) -> bool:
        return ctx.output.status in self.allowed_values

# Usage
evaluators = [
    IsInstance(type_name="MyAgentResponse"),  # Check type first
    HasNonEmptyMessage(min_length=10),
    StatusIsValid(),
]

4. Span-Based Evaluation

Inspect execution traces to verify internal agent behavior (tool calls, retrieval steps).

from pydantic_evals.evaluators import HasMatchingSpan
from pydantic_evals.otel import SpanQuery

# Verify agent called a specific tool
# NOTE: HasMatchingSpan takes a query parameter with SpanQuery
tool_check = HasMatchingSpan(
    query=SpanQuery(
        name_equals='running tool',
        has_attributes={'gen_ai.tool.name': 'calculator'}
    )
)

For detailed guide, see references/evaluator-types.md

Integration with Pydantic AI

Define Agent as Task

Wrap agent execution in a task function:

from pydantic_ai import Agent

agent = Agent('openai:gpt-4o-mini', system_prompt="You are helpful.")

async def run_agent(query: str) -> str:
    result = await agent.run(query)
    return result.output  # Use result.output, NOT result.data

Handle Dependencies

Use dependency injection for deterministic testing:

from dataclasses import dataclass

@dataclass
class Deps:
    api_key: str

# During testing, override with mocks
test_deps = Deps(api_key="test_key")

For integration guide, see references/integration.md

Logfire Observability

Enable automatic tracing for debugging:

import logfire

logfire.configure(send_to_logfire='if-token-present')
logfire.instrument_pydantic_ai()

# Evaluations now create rich traces viewable in Logfire dashboard

Benefits:

• Trace every evaluation run
• Visualize agent internal execution
• Compare experiments side-by-side
• Debug failures with full context

Dataset Management

Save/Load Datasets

# Save to YAML with schema
dataset.to_file('evals.yaml', fmt='yaml')

# Load from file
dataset = Dataset.from_file('evals.yaml')

Important: Use typed Dataset for proper serialization:

# Define typed dataset to avoid serialization warnings
dataset: Dataset[str, str, None] = Dataset(...)

# Or when loading from file with custom evaluators
from types import NoneType
dataset = Dataset[MyInputType, MyOutputType, NoneType].from_file(
    'evals.yaml',
    custom_evaluator_types=(MyCustomEvaluator,)
)

Generate Datasets with LLM

from pydantic_evals.generation import generate_dataset

dataset = await generate_dataset(
    dataset_type=Dataset[str, str, None],
    model='openai:o1',
    n_examples=10,
    extra_instructions="Generate diverse test cases for customer support agent"
)

Best Practices

1. Fail-fast: Run deterministic evaluators before LLM judges
2. Cost-latency trade-off:

- Commit hooks: Deterministic only - PR merges: Small LLM judges on critical cases - Nightly builds: Full LLM judge suite

1. Concurrency: Use max_concurrency parameter to avoid rate limits
2. Versioning: Store datasets in Git alongside code
3. Regression testing: Compare experiments to detect degradation

Common Workflows

Workflow 1: Create Evaluation Suite

1. Define Cases with inputs and expected outputs
2. Choose evaluators based on requirements
3. Create Dataset with cases and evaluators
4. Save to YAML for version control

Workflow 2: Run Evaluations

1. Load Dataset from file
2. Define task function (agent wrapper)
3. Run dataset.evaluate_sync(task) or dataset.evaluate(task)
4. Analyze report with report.print() or Logfire

Accessing Results:

report = dataset.evaluate_sync(my_task)
report.print()

# Access individual case results
for case in report.cases:  # NOTE: Use .cases, NOT .case_results
    print(f"Case: {case.name}")
    print(f"Output: {case.output}")
    print(f"Passed: {case.passed}")

Workflow 3: Compare Models

1. Run same dataset against different models
2. Generate Experiments for each run
3. Compare metrics (pass rates, latency, scores)
4. Use Logfire comparison view

Examples

Complete example files demonstrating patterns:

• references/examples/generate_dataset.py: Generate test cases with LLM
• references/examples/custom_evaluators.py: Implement custom evaluation logic
• references/examples/unit_testing.py: Run evaluations in CI/CD
• references/examples/compare_models.py: Benchmark different models

Resources

references/

• core-concepts.md: Detailed explanation of Case, Dataset, Evaluator, Experiment
• evaluator-types.md: Deep dive into all evaluator types
• integration.md: Pydantic AI and Logfire integration guide
• examples/: Complete working examples