This product is not supported for your selected
Datadog site. (
).
Overview
Pydantic is an open source framework that provides ready-to-use evaluations and allows for customizable LLM evaluations. For more information, see Pydantic’s documentation.
You can use LLM Observability to run Pydantic evaluations and scalar Pydantic report evaluations in Experiments. Pydantic evaluation results appear as evaluator results tied to each instance in an LLM Observability dataset. Pydantic report evaluations run on an entire LLM Observability dataset and report one scalar result for the dataset.
Setup
- Set up an LLM Observability Experiment and an LLM Observability Dataset.
- Provide a Pydantic evaluator to the
evaluators parameter in an LLMObs Experiment as demonstrated in the following code sample. (Optional) Provide a Pydantic report evaluator to the summary_evaluators parameter in an LLMObs Experiment. Note: Only Pydantic report evaluators that return a ScalarResult are supported.
from pydantic_evals.evaluators import (
EqualsExpected,
EvaluationReason,
Evaluator,
EvaluatorContext,
EvaluatorOutput,
LLMJudge,
ReportEvaluator,
ReportEvaluatorContext,
)
from pydantic_evals.reporting.analyses import ScalarResult
from ddtrace.llmobs import LLMObs
LLMObs.enable(
api_key="<YOUR_API_KEY>", # defaults to DD_API_KEY environment variable
app_key="<YOUR_APP_KEY>", # defaults to DD_APP_KEY environment variable
site="datadoghq.com", # defaults to DD_SITE environment variable
project_name="<YOUR_PROJECT>" # defaults to DD_LLMOBS_PROJECT_NAME environment variable, or "default-project" if the environment variable is not set
)
# this can be any Pydantic evaluator
@dataclass
class ComprehensiveCheck(Evaluator):
def evaluate(self, ctx: EvaluatorContext) -> EvaluatorOutput:
format_valid = self._check_format(ctx.output)
to_return = {
'valid_format': EvaluationReason(
value=format_valid,
reason='Valid JSON format' if format_valid else 'Invalid JSON format',
),
'quality_score': self._score_quality(ctx.output),
'category': self._classify(ctx.output),
}
return to_return
def _check_format(self, output: str) -> bool:
return output.startswith('{') and output.endswith('}')
def _score_quality(self, output: str) -> float:
return len(output) / 100.0
def _classify(self, output: str) -> str:
return 'short' if len(output) < 50 else 'long'
# This can be any Pydantic ReportEvaluator that returns ScalarResult
class TotalCasesEvaluator(ReportEvaluator):
def evaluate(self, ctx: ReportEvaluatorContext) -> ScalarResult:
return ScalarResult(
title='Total',
value=len(ctx.report.cases),
unit='cases',
)
dataset = LLMObs.create_dataset(
dataset_name="capitals-of-the-world",
project_name="capitals-project", # optional, defaults to project_name used in LLMObs.enable
description="Questions about world capitals",
records=[
{
"input_data": {
"question": "What is the capital of China?"
}, # required, JSON or string
"expected_output": "Beijing", # optional, JSON or string
"metadata": {"difficulty": "easy"}, # optional, JSON
},
{
"input_data": {
"question": "Which city serves as the capital of South Africa?"
},
"expected_output": "Pretoria",
"metadata": {"difficulty": "medium"},
},
],
)
def task(input_data: Dict[str, Any], config: Optional[Dict[str, Any]] = None, metadata: Optional[Dict[str, Any]] = None) -> str:
question = input_data['question']
# Your LLM or processing logic here
return "Beijing" if "China" in question else "Unknown"
llm_judge = LLMJudge(
rubric='Response provides the same answer as expected, possibly with explanation',
include_input=True,
include_expected_output=True,
)
experiment = LLMObs.experiment(
name="<EXPERIMENT_NAME>",
task=my_task,
dataset=dataset,
evaluators=[EqualsExpected(), ComprehensiveCheck(), llm_judge],
summary_evaluators=[TotalCasesEvaluator()],
description="<EXPERIMENT_DESCRIPTION>",
)
results = experiment.run(jobs=4, raise_errors=True)
print(f"View experiment: {experiment.url}")
For a working example, see Datadog’s Pydantic demo in GitHub.
Usage
After you run an experiment with a Pydantic evaluation, you can view the Pydantic evaluation results per instance in the corresponding experiment run in Datadog. In the following experiment, two Pydantic evaluations (a custom Pydantic evaluator with the name “ComprehensiveCheck” and a built-in evaluator with the name “EqualsExpected”) and one Pydantic report evaluator (a custom Pydantic report evaluator with the name “TotalCasesEvaluator”) were run:
Further reading
Additional helpful documentation, links, and articles:
