DeepEval integration

agento11y-deepeval supports live target execution with native DeepEval metrics, and publication of completed native evaluation results. Both use the existing Agent Observability experiment lifecycle.

Install

pip install agento11y agento11y-deepeval

Set AGENTO11Y_ENDPOINT, AGENTO11Y_AUTH_TOKEN, and, when required by your stack, AGENTO11Y_AUTH_TENANT_ID. Set AGENTO11Y_GRAFANA_URL for experiment deep links.

Live runner

run_deepeval registers the planned count before invoking the target and exports each case as it completes. Configure your OTel TracerProvider/exporter once; the runner activates suite and trial spans around your instrumented target.

from agento11y_deepeval import run_deepeval
from deepeval.metrics import ExactMatchMetric
from deepeval.test_case import LLMTestCase

run = run_deepeval(
    [LLMTestCase(input="2+2?", actual_output="", expected_output="4")],
    lambda case, trial: instrumented_agent(case.input),
    [ExactMatchMetric()],
    experiment_name="Live regression",
)
print(run.url)

The target receives the trial for conversation binding when needed. Execution is sequential, with a fresh copy of each metric per case. Target and metric exceptions mark the trial errored and retain successful siblings. Target duration excludes grading. record_io=False avoids extra anchor generations by default. Native batch scheduling, caching, and pytest can continue using the completed-result importers below.

For ConversationalTestCase, the same target receives a private copy of the case and must return the completed ordered Sequence[Turn]. The single-turn target(case, trial) -> str contract is unchanged. Returning turns instead of mutating the case keeps caller-owned cases reusable:

from deepeval.metrics import ConversationCompletenessMetric
from deepeval.test_case import ConversationalTestCase, Turn

case = ConversationalTestCase(turns=[Turn(role="user", content="Can I return this?")])

def chat(case, trial):
    return [*case.turns, Turn(role="assistant", content="Yes, within 30 days.")]

run_deepeval(
    [case],
    chat,
    [ConversationCompletenessMetric()],
    experiment_name="Live support conversations",
    record_io=True,
)

With record_io=True, each assistant turn is exported as one generation using all preceding turns as structured input. Assistant generations share one stable conversation id and form a parent chain. A final user turn is retained as a chained input-only generation awaiting a response. If the target already emitted its own telemetry, bind that conversation on trial and leave record_io=False.

Evaluate existing outputs and publish afterward

evaluate_with_agento11y below waits for the whole native batch before creating the Grafana experiment. Use run_deepeval for live target execution/progress.

from deepeval.metrics import AnswerRelevancyMetric, FaithfulnessMetric
from deepeval.test_case import LLMTestCase
from agento11y_deepeval import evaluate_with_agento11y

test_cases = [
    LLMTestCase(
        name="refund-policy",
        input="Can I return an opened item?",
        actual_output=run_agent("Can I return an opened item?"),
        expected_output="Opened items can be returned within 30 days.",
        retrieval_context=["Opened products may be returned within 30 days."],
        metadata={"agento11y.test_case_id": "refund-policy"},
    )
]

run = evaluate_with_agento11y(
    test_cases,
    [AnswerRelevancyMetric(), FaithfulnessMetric()],
    experiment_name="Checkout agent regression",
    primary_metric="Answer Relevancy",
    suite_id="checkout-regression",
    suite_version="v1",
)
print(run.published.url)

Each DeepEval test result becomes one trial. The metric named by primary_metric is the one primary_verdict; every other metric is a diagnostic. If all results contain exactly one metric name, the adapter selects it automatically. With multiple metrics, the explicit selection is required and validated before the experiment is written.

Conversational results preserve their ordered public Turn fields in the test case snapshot. With record_io=True, each assistant response becomes a separate generation; preceding user and assistant turns remain structured messages, not a flattened transcript. Whole-conversation metric scores attach to the typed trial and conversation, not to the final assistant generation.

from deepeval import evaluate
from deepeval.metrics import ConversationCompletenessMetric
from deepeval.test_case import ConversationalTestCase, Turn

conversation = ConversationalTestCase(
    turns=[
        Turn(role="user", content="Can I return an opened item?"),
        Turn(role="assistant", content="Yes, within 30 days."),
    ],
    expected_outcome="The return policy is explained.",
)
result = evaluate(test_cases=[conversation], metrics=[ConversationCompletenessMetric()])
publish_deepeval_results(result, experiment_name="Support conversations")

Only public DeepEval user and assistant turns with non-blank string content are accepted. Malformed turns fail validation before the remote experiment is created. Turn order, user id, retrieval context, tool/MCP call details, audio, latency, interruption state, comments, and custom metadata are retained in the case snapshot and generation metadata. Text roles/content and user id also map to their native Agent Observability fields.

Metric errors make the trial operationally failed and remain unrated; valid sibling results are still published. Native duration can be supplied as metadata={"agento11y.duration_ms": 12345}. Missing duration remains unknown.

Metric scores, pass/fail values, reasons, thresholds, model identity, evaluator cost, and evaluator token counts are preserved. Evaluator cost and tokens remain score metadata rather than candidate trial usage, because they describe the grader rather than the agent under test.

Publish a result you already evaluated

from deepeval import evaluate
from agento11y_deepeval import publish_deepeval_results

result = evaluate(test_cases=test_cases, metrics=metrics)
published = publish_deepeval_results(
    result,
    experiment_name="Existing DeepEval run",
    primary_metric="Answer Relevancy",
)

Generation and conversation correlation

The adapter records each single-turn result as one anchor generation by default; conversational results produce one generation per assistant turn. If the evaluated target already emits Agent Observability telemetry, use that telemetry instead:

published = publish_deepeval_results(
    result,
    experiment_name="Instrumented target",
    primary_metric="Answer Relevancy",
    record_io=False,
    resolve_conversation_id=lambda item: item.metadata["conversation_id"],
    resolve_generation_id=lambda item: item.metadata["generation_id"],
)

Without resolver functions, metadata keys agento11y.conversation_id, agento11y.generation_id, and agento11y.trace_id are recognized. A stable case id can be supplied as agento11y.test_case_id. For conversational results, a resolved conversation id suppresses generation export even when record_io=True; the optional generation id is not attached to whole-conversation scores.

To publish browsable cases, pass test_suites_client and explicitly choose suite_publication_policy="subset" to retain remote-only cases or "replace" to prune them. Trial snapshots use the canonical server-returned cases and record digests when sanitization changed the executed case.

This adapter does not create or invoke Agent Observability stored evaluators. DeepEval runs metrics locally (including model-backed metrics) and the adapter publishes their scores.

See the runnable deterministic example.