wehub-resource-sync
85 lines
3.0 KiB
Python
85 lines
3.0 KiB
Python
"""Component-level evals for Google ADK via ``dataset.evals_iterator``.
|
|
|
|
Mirrors ``tests/test_integrations/test_agentcore/test_evaluate_agent.py``:
|
|
drives a Google ADK agent through the async iterator path, with a
|
|
per-task ``next_agent_span(metrics=[...])`` wrap so the
|
|
``AnswerRelevancyMetric`` lands on the agent span via the
|
|
``stash_pending_metrics`` overlay (carried across OTel transport into
|
|
``ConfidentSpanExporter``). The ``evals_iterator`` itself sets
|
|
``trace_manager.is_evaluating=True``, which:
|
|
|
|
- flips ``ContextAwareSpanProcessor`` to REST routing so the spans
|
|
flow through ``trace_manager`` (instead of OTLP), and
|
|
- gates ``stash_pending_metrics`` so ``BaseMetric`` instances
|
|
actually make it from the interceptor to the exporter.
|
|
|
|
This is the canonical end-to-end shape for Google ADK + component-level
|
|
evals after the OTel POC migration.
|
|
|
|
Skipped without ``GOOGLE_API_KEY`` (Gemini invocation) +
|
|
``OPENAI_API_KEY`` (the metric scorer).
|
|
"""
|
|
|
|
import asyncio
|
|
import os
|
|
|
|
import pytest
|
|
|
|
from deepeval.dataset import EvaluationDataset, Golden
|
|
from deepeval.evaluate.configs import AsyncConfig
|
|
from deepeval.metrics import AnswerRelevancyMetric
|
|
from deepeval.tracing import next_agent_span
|
|
|
|
from tests.test_integrations.test_googleadk.apps.googleadk_eval_app import (
|
|
ainvoke_evals_agent,
|
|
init_evals_googleadk,
|
|
)
|
|
|
|
|
|
pytestmark = pytest.mark.skipif(
|
|
not os.getenv("GOOGLE_API_KEY") or not os.getenv("OPENAI_API_KEY"),
|
|
reason=(
|
|
"GOOGLE_API_KEY is required for Gemini and OPENAI_API_KEY "
|
|
"for the AnswerRelevancyMetric scorer."
|
|
),
|
|
)
|
|
|
|
|
|
answer_relevancy_metric = AnswerRelevancyMetric()
|
|
|
|
|
|
def test_evaluate_agent():
|
|
"""End-to-end: 1 golden through a Google ADK agent, scored by
|
|
AnswerRelevancyMetric attached via ``next_agent_span(metrics=[...])``.
|
|
"""
|
|
invoke_func = init_evals_googleadk(
|
|
name="googleadk-evaluate-agent",
|
|
tags=["googleadk", "evaluate", "iterator"],
|
|
metadata={"test_type": "evaluate_agent"},
|
|
thread_id="evaluate-agent-thread-001",
|
|
user_id="evaluate-agent-user-001",
|
|
)
|
|
|
|
dataset = EvaluationDataset(
|
|
goldens=[Golden(input="What's 7 multiplied by 8?")]
|
|
)
|
|
|
|
async def run_agent(prompt: str):
|
|
# Span-level metric attached to the agent span via
|
|
# next_agent_span; with ``trace_manager.is_evaluating=True`` set
|
|
# by evals_iterator, the interceptor's ``stash_pending_metrics``
|
|
# call carries the metric across OTel transport so the
|
|
# exporter can re-attach it on the rebuilt AgentSpan.
|
|
with next_agent_span(metrics=[answer_relevancy_metric]):
|
|
return await ainvoke_evals_agent(prompt, invoke_func=invoke_func)
|
|
|
|
for golden in dataset.evals_iterator(
|
|
async_config=AsyncConfig(run_async=True),
|
|
metrics=[answer_relevancy_metric],
|
|
):
|
|
task = asyncio.create_task(run_agent(golden.input))
|
|
dataset.evaluate(task)
|
|
|
|
assert answer_relevancy_metric.score is not None
|
|
assert answer_relevancy_metric.score > 0.0
|