comet-ml--opik/sdks/python/tests/e2e/evaluation/test_threads_evaluate.py

import time
import uuid
import pytest

import opik
from opik import synchronization
from opik.evaluation import metrics
from opik.evaluation.models import LiteLLMChatModel
from opik.evaluation.threads import evaluator
from opik.types import FeedbackScoreDict

from ...testlib import environment
from .. import verifiers


@pytest.fixture
def real_model_conversation():
    return [
        {
            "role": "user",
            "content": "I need to book a flight to New York and find a hotel.",
        },
        {
            "role": "assistant",
            "content": "I can help you with that. For flights to New York, what dates are you looking to travel?",
        },
    ]


@pytest.fixture
def active_thread_and_project_name(
    opik_client, real_model_conversation, temporary_project_name
):
    thread_id = str(uuid.uuid4())[-6:]

    # create conversation traces
    i = 0
    while i < len(real_model_conversation) - 1:
        opik_client.trace(
            name=f"trace-name-{i}:{thread_id}",
            input={"input": f"{real_model_conversation[i]['content']}"},
            output={"output": f"{real_model_conversation[i + 1]['content']}"},
            project_name=temporary_project_name,
            thread_id=thread_id,
        )
        time.sleep(0.1)
        i += 2

    opik_client.flush()

    return thread_id, temporary_project_name


def _one_thread_is_active(project_name: str, opik_client: opik.Opik) -> bool:
    threads = opik_client.search_threads(
        project_name=project_name, filter_string='status = "active"'
    )
    return len(threads) == 1


@pytest.fixture
def eval_project_name(temporary_project_name: str) -> str:
    return temporary_project_name


@pytest.mark.skipif(
    not environment.has_openai_api_key(), reason="OPENAI_API_KEY is not set"
)
def test_evaluate_threads__happy_path(
    opik_client, active_thread_and_project_name, eval_project_name
):
    active_thread, project_name = active_thread_and_project_name
    # wait for active threads to propagate
    if not synchronization.until(
        lambda: _one_thread_is_active(project_name, opik_client), max_try_seconds=30
    ):
        raise AssertionError(f"Failed to create threads in project '{project_name}'")

    judge_model = LiteLLMChatModel(reasoning_effort="minimal")

    metrics_ = [
        metrics.ConversationalCoherenceMetric(model=judge_model),
    ]

    result = evaluator.evaluate_threads(
        project_name=project_name,
        filter_string=f'id = "{active_thread}"',
        metrics=metrics_,
        eval_project_name=eval_project_name,
        trace_input_transform=lambda x: x["input"],
        trace_output_transform=lambda x: x["output"],
        verbose=1,
    )

    assert result is not None
    assert len(result.results) == 1  # we have only one thread

    thread_result = result.results[0]
    assert thread_result.thread_id == active_thread
    assert len(thread_result.scores) == len(metrics_)

    feedback_scores = [
        FeedbackScoreDict(
            id=active_thread,
            name=score.name,
            value=score.value,
            reason=score.reason.strip(),
            category_name=None,
        )
        for score in thread_result.scores
        if not score.scoring_failed
    ]

    verifiers.verify_thread(
        opik_client=opik_client,
        thread_id=active_thread,
        project_name=project_name,
        feedback_scores=feedback_scores,
    )


def test_evaluate_threads__no_truncation_for_long_traces(
    opik_client, temporary_project_name
):
    """E2E test verifying that long trace content is not truncated during evaluation.

    The test creates a trace with output exceeding 15,000 characters with a unique marker
    at the end, then verifies the marker is present in the transform function, proving
    that truncation did not occur.
    """
    thread_id = str(uuid.uuid4())[-6:]

    # Create a long output that exceeds the truncation threshold (~9935 chars)
    # with a unique marker at the end that would be lost if truncated
    marker = "UNIQUE_END_MARKER_XYZ123"
    long_content = "a" * 15000 + marker

    # Create a trace with very long output
    opik_client.trace(
        name=f"long-trace:{thread_id}",
        input={"input": "test input"},
        output={"output": long_content},
        project_name=temporary_project_name,
        thread_id=thread_id,
    )

    opik_client.flush()

    # Wait for thread to be created
    if not synchronization.until(
        lambda: _one_thread_is_active(temporary_project_name, opik_client),
        max_try_seconds=30,
    ):
        raise AssertionError(
            f"Failed to create thread in project '{temporary_project_name}'"
        )

    # Track what the transform receives
    received_outputs = []

    def input_transform(x):
        return x.get("input", "")

    def output_transform(x):
        """Transform that captures the output to verify it's not truncated."""
        # When truncated, x might be a string (malformed JSON) instead of dict
        if isinstance(x, str):
            # This is the bug! Truncation causes malformed JSON string
            received_outputs.append(f"TRUNCATED_STRING:{x[:100]}...")
            return "TRUNCATED"
        output = x.get("output", "")
        received_outputs.append(output)
        return output

    # Create a simple metric that just checks the output
    class ContentVerificationMetric(metrics.base_metric.BaseMetric):
        def __init__(self):
            super().__init__(
                name="content_verification",
                track=False,
            )

        def score(self, conversation, **ignored_kwargs):
            # Just return a dummy score - we're really testing the transform
            return metrics.score_result.ScoreResult(
                name=self.name,
                value=1.0,
                reason="Content verification",
            )

    # Run evaluation
    result = evaluator.evaluate_threads(
        project_name=temporary_project_name,
        filter_string=f'id = "{thread_id}"',
        metrics=[ContentVerificationMetric()],
        eval_project_name=temporary_project_name,
        trace_input_transform=input_transform,
        trace_output_transform=output_transform,
        verbose=0,
    )

    assert result is not None
    assert len(result.results) == 1

    # Verify that the transform received the full content with the marker
    assert len(received_outputs) > 0, "Transform should have been called"
    transformed_content = received_outputs[0]

    # This is the critical assertion: if truncation occurred, the marker would be missing
    assert marker in transformed_content, (
        f"Content was truncated! Expected marker '{marker}' not found. "
        f"Content length: {len(transformed_content)}, expected: {len(long_content)}. "
        f"Last 100 chars: {transformed_content[-100:]}"
    )

    # Also verify the full length is preserved
    assert len(transformed_content) == len(long_content), (
        f"Content length mismatch: got {len(transformed_content)}, "
        f"expected {len(long_content)}"
    )