mlflow--mlflow/tests/genai/scorers/test_serialization.py

import json
from typing import Any, ClassVar
from unittest.mock import Mock, patch

import pytest

import mlflow
from mlflow.entities import Feedback
from mlflow.entities.scorer import ScorerVersion
from mlflow.exceptions import MlflowException
from mlflow.genai.scorers import Scorer, scorer
from mlflow.genai.scorers.base import SerializedScorer
from mlflow.genai.scorers.builtin_scorers import Guidelines


@pytest.fixture(autouse=True)
def mock_databricks_runtime():
    with patch("mlflow.genai.scorers.base.is_databricks_uri", return_value=True):
        yield


# ============================================================================
# FORMAT VALIDATION TESTS (Minimal - just check serialization structure)
# ============================================================================


def test_decorator_scorer_serialization_format():
    @scorer(name="test_scorer", aggregations=["mean"])
    def test_scorer(outputs):
        return outputs == "correct"

    serialized = test_scorer.model_dump()

    # Check required fields for decorator scorers
    assert serialized["name"] == "test_scorer"
    assert serialized["aggregations"] == ["mean"]
    assert "call_source" in serialized
    assert "original_func_name" in serialized
    assert serialized["original_func_name"] == "test_scorer"
    assert "call_signature" in serialized

    # Check version metadata
    assert "mlflow_version" in serialized
    assert "serialization_version" in serialized
    assert serialized["serialization_version"] == 1

    # Builtin scorer fields should be None (not populated for decorator scorers)
    assert serialized["builtin_scorer_class"] is None
    assert serialized["builtin_scorer_pydantic_data"] is None


def test_builtin_scorer_serialization_format():
    from mlflow.genai.scorers.builtin_scorers import RelevanceToQuery

    serialized = RelevanceToQuery().model_dump()

    # Check required top-level fields for builtin scorers
    assert serialized["name"] == "relevance_to_query"
    assert "builtin_scorer_class" in serialized
    assert serialized["builtin_scorer_class"] == "RelevanceToQuery"
    assert "builtin_scorer_pydantic_data" in serialized

    # Check fields within builtin_scorer_pydantic_data
    pydantic_data = serialized["builtin_scorer_pydantic_data"]
    assert "required_columns" in pydantic_data

    # Check version metadata
    assert "mlflow_version" in serialized
    assert "serialization_version" in serialized
    assert serialized["serialization_version"] == 1

    # Decorator scorer fields should be None (not populated for builtin scorers)
    assert serialized["call_source"] is None
    assert serialized["call_signature"] is None
    assert serialized["original_func_name"] is None


# ============================================================================
# ROUND-TRIP FUNCTIONALITY TESTS (Comprehensive - test complete cycles)
# ============================================================================


def test_simple_scorer_round_trip():
    @scorer
    def simple_scorer(outputs):
        return outputs == "correct"

    # Test original functionality
    assert simple_scorer(outputs="correct") is True
    assert simple_scorer(outputs="wrong") is False

    # Serialize and deserialize
    serialized = simple_scorer.model_dump()
    deserialized = Scorer.model_validate(serialized)

    # Test deserialized functionality matches original
    assert deserialized.name == "simple_scorer"
    assert deserialized(outputs="correct") is True
    assert deserialized(outputs="wrong") is False


def test_custom_name_and_aggregations_round_trip():
    @scorer(name="length_check", aggregations=["mean", "max"])
    def my_scorer(inputs, outputs):
        return len(outputs) > len(inputs)

    # Test original
    assert my_scorer(inputs="hi", outputs="hello world") is True
    assert my_scorer(inputs="hello", outputs="hi") is False

    # Round-trip
    serialized = my_scorer.model_dump()
    deserialized = Scorer.model_validate(serialized)

    # Test preserved properties and functionality
    assert deserialized.name == "length_check"
    assert deserialized.aggregations == ["mean", "max"]
    assert deserialized(inputs="hi", outputs="hello world") is True
    assert deserialized(inputs="hello", outputs="hi") is False


def test_multiple_parameters_round_trip():
    @scorer
    def multi_param_scorer(inputs, outputs, expectations):
        return outputs.startswith(inputs) and len(outputs) > expectations.get("min_length", 0)

    # Test original
    test_args = {
        "inputs": "Hello",
        "outputs": "Hello world!",
        "expectations": {"min_length": 5},
    }
    assert multi_param_scorer(**test_args) is True

    # Round-trip
    serialized = multi_param_scorer.model_dump()
    deserialized = Scorer.model_validate(serialized)

    # Test functionality preserved
    assert deserialized(**test_args) is True
    assert (
        deserialized(inputs="Hi", outputs="Hello world!", expectations={"min_length": 5}) is False
    )


def test_complex_logic_round_trip():
    @scorer
    def complex_scorer(outputs):
        if not outputs:
            return 0

        words = outputs.split()
        score = 0
        for word in words:
            if word.isupper():
                score += 2
            elif word.islower():
                score += 1

        return score

    # Test original functionality
    test_cases = [
        ("", 0),
        ("hello world", 2),  # 2 lowercase words
        ("HELLO WORLD", 4),  # 2 uppercase words
        ("Hello WORLD", 2),  # mixed case "Hello" (0) + "WORLD" (2)
    ]

    for test_input, expected in test_cases:
        assert complex_scorer(outputs=test_input) == expected

    # Round-trip
    serialized = complex_scorer.model_dump()
    deserialized = Scorer.model_validate(serialized)

    # Test all cases still work
    for test_input, expected in test_cases:
        assert deserialized(outputs=test_input) == expected


def test_imports_and_feedback_round_trip():
    @scorer
    def feedback_scorer(outputs):
        import re  # clint: disable=lazy-import

        pattern = r"\b\w+\b"
        words = re.findall(pattern, outputs)
        return Feedback(value=len(words), rationale=f"Found {len(words)} words")

    # Test original
    result = feedback_scorer(outputs="hello world test")
    assert isinstance(result, Feedback)
    assert result.value == 3
    assert "Found 3 words" in result.rationale

    # Round-trip
    serialized = feedback_scorer.model_dump()
    deserialized = Scorer.model_validate(serialized)

    # Test functionality preserved
    result = deserialized(outputs="hello world test")
    assert isinstance(result, Feedback)
    assert result.value == 3
    assert "Found 3 words" in result.rationale


def test_default_parameters_round_trip():
    @scorer
    def default_scorer(outputs, threshold=5):
        return len(outputs) > threshold

    # Test original with and without default
    assert default_scorer(outputs="short") is False  # len=5, not > 5
    assert default_scorer(outputs="longer") is True  # len=6, > 5
    assert default_scorer(outputs="hi", threshold=1) is True  # len=2, > 1

    # Round-trip
    serialized = default_scorer.model_dump()
    deserialized = Scorer.model_validate(serialized)

    # Test defaults work
    assert deserialized(outputs="short") is False
    assert deserialized(outputs="longer") is True


def test_json_workflow_round_trip():
    @scorer(name="json_test", aggregations=["mean"])
    def json_scorer(outputs):
        return len(outputs.split()) > 3

    # Test original
    assert json_scorer(outputs="one two three four") is True
    assert json_scorer(outputs="one two") is False

    # JSON round-trip
    serialized = json_scorer.model_dump()
    json_str = json.dumps(serialized)
    loaded_dict = json.loads(json_str)
    deserialized = Scorer.model_validate(loaded_dict)

    # Test functionality preserved through JSON
    assert deserialized.name == "json_test"
    assert deserialized.aggregations == ["mean"]
    assert deserialized(outputs="one two three four") is True
    assert deserialized(outputs="one two") is False


def test_end_to_end_complex_round_trip():
    @scorer(name="complete_test", aggregations=["mean", "max"])
    def complete_scorer(inputs, outputs, expectations):
        input_words = len(inputs.split())
        output_words = len(outputs.split())
        expected_ratio = expectations.get("word_ratio", 1.0)

        actual_ratio = output_words / input_words if input_words > 0 else 0
        return actual_ratio >= expected_ratio

    test_args = {
        "inputs": "hello world",
        "outputs": "hello beautiful world today",
        "expectations": {"word_ratio": 1.5},
    }

    # Test original
    original_result = complete_scorer(**test_args)
    assert original_result is True

    # Round-trip
    serialized = complete_scorer.model_dump()
    deserialized = Scorer.model_validate(serialized)

    # Test everything preserved
    assert deserialized.name == "complete_test"
    assert deserialized.aggregations == ["mean", "max"]
    deserialized_result = deserialized(**test_args)
    assert original_result == deserialized_result is True


def test_deserialized_scorer_runs_without_global_context():
    # Create a simple scorer that only uses built-in functions and parameters
    @scorer(name="isolated_test")
    def simple_scorer(outputs):
        # Only use built-in functions and the parameter - no external dependencies
        return len(outputs.split()) > 2

    # Test original works
    assert simple_scorer(outputs="one two three") is True
    assert simple_scorer(outputs="one two") is False

    # Serialize the scorer
    serialized_data = simple_scorer.model_dump()

    # Test deserialized scorer in completely isolated namespace using exec
    test_code = """
# Import required modules in isolated namespace
from mlflow.genai.scorers import Scorer

# Deserialize the scorer (no external context available)
deserialized = Scorer.model_validate(serialized_data)

# Test that it can run successfully in isolation
result1 = deserialized(outputs="one two three")
result2 = deserialized(outputs="one two")
result3 = deserialized(outputs="hello world test case")

# Store results for verification
test_results = {
    "result1": result1,
    "result2": result2,
    "result3": result3,
    "name": deserialized.name,
    "aggregations": deserialized.aggregations
}
"""

    # Execute in isolated namespace with only serialized_data available
    isolated_namespace = {"serialized_data": serialized_data}
    exec(test_code, isolated_namespace)  # noqa: S102

    # Verify results from isolated execution
    results = isolated_namespace["test_results"]
    assert results["result1"] is True  # "one two three" has 3 words > 2
    assert results["result2"] is False  # "one two" has 2 words, not > 2
    assert results["result3"] is True  # "hello world test case" has 4 words > 2
    assert results["name"] == "isolated_test"
    assert results["aggregations"] is None


def test_builtin_scorer_round_trip():
    # from mlflow.genai.scorers import relevance_to_query
    from mlflow.genai.scorers.builtin_scorers import RelevanceToQuery

    # Round-trip serialization
    serialized = RelevanceToQuery().model_dump()
    deserialized = Scorer.model_validate(serialized)

    # Test class type and properties preserved
    assert isinstance(deserialized, RelevanceToQuery)
    assert deserialized.name == "relevance_to_query"
    assert hasattr(deserialized, "required_columns")
    assert deserialized.required_columns == {"inputs", "outputs"}

    # Test execution with mocking
    with patch(
        "mlflow.genai.judges.is_context_relevant",
        return_value=Feedback(name="relevance_to_query", value="yes", metadata={"chunk_index": 0}),
    ) as mock_judge:
        result = deserialized(
            inputs={"question": "What is machine learning?"},
            outputs=(
                "Machine learning is a subset of AI that enables computers to learn without "
                "explicit programming."
            ),
        )

    # Verify execution worked correctly
    mock_judge.assert_called_once_with(
        request="{'question': 'What is machine learning?'}",
        context=(
            "Machine learning is a subset of AI that enables computers to learn without "
            "explicit programming."
        ),
        name="relevance_to_query",
        model=None,
    )

    assert isinstance(result, Feedback)
    assert result.name == "relevance_to_query"
    assert result.value == "yes"
    assert result.metadata == {"chunk_index": 0}  # chunk_index should be preserved


def test_builtin_scorer_with_parameters_round_trip():
    from mlflow.genai.scorers.builtin_scorers import Guidelines

    # Create scorer with custom parameters
    tone = (
        "The response must maintain a courteous, respectful tone throughout. "
        "It must show empathy for customer concerns."
    )
    tone_scorer = Guidelines(name="tone", guidelines=[tone])

    # Verify original properties
    assert tone_scorer.name == "tone"
    assert tone_scorer.guidelines == [tone]
    assert isinstance(tone_scorer, Guidelines)

    # Round-trip serialization
    serialized = tone_scorer.model_dump()

    # Verify serialization format includes all fields
    assert "builtin_scorer_class" in serialized
    assert serialized["builtin_scorer_class"] == "Guidelines"
    assert "builtin_scorer_pydantic_data" in serialized
    pydantic_data = serialized["builtin_scorer_pydantic_data"]
    assert "guidelines" in pydantic_data
    assert pydantic_data["guidelines"] == [tone]
    assert pydantic_data["name"] == "tone"

    # Deserialize
    deserialized = Scorer.model_validate(serialized)

    # Test class type and all properties preserved
    assert isinstance(deserialized, Guidelines)
    assert deserialized.name == "tone"
    assert deserialized.guidelines == [tone]
    assert hasattr(deserialized, "required_columns")
    assert deserialized.required_columns == {"inputs", "outputs"}

    # Test that it can be executed with mocking
    with patch(
        "mlflow.genai.judges.meets_guidelines",
        return_value=Feedback(
            name="tone", value=True, rationale="Response is appropriately courteous"
        ),
    ) as mock_judge:
        result = deserialized(
            inputs={"question": "What is the issue?"},
            outputs=(
                "Thank you for bringing this to my attention. I understand your concern and "
                "will help resolve this issue promptly."
            ),
        )

    # Verify execution worked correctly
    mock_judge.assert_called_once_with(
        guidelines=[tone],
        context={
            "request": "{'question': 'What is the issue?'}",
            "response": (
                "Thank you for bringing this to my attention. I understand your concern and "
                "will help resolve this issue promptly."
            ),
        },
        name="tone",
        model=None,
    )

    assert isinstance(result, Feedback)
    assert result.name == "tone"
    assert result.value is True


def test_direct_subclass_scorer_rejected():
    class DirectSubclassScorer(Scorer):
        """An unsupported direct subclass of Scorer."""

        def __init__(self, **data):
            super().__init__(name="direct_subclass", **data)

        def __call__(self, *, outputs):
            return len(outputs) > 5

    # Create instance - this should work
    direct_scorer = DirectSubclassScorer()

    # Calling it should work
    assert direct_scorer(outputs="hello world") is True
    assert direct_scorer(outputs="hi") is False

    # But serialization should raise an error
    with pytest.raises(MlflowException, match="Unsupported scorer type: DirectSubclassScorer"):
        direct_scorer.model_dump()

    # Verify the error message is informative
    try:
        direct_scorer.model_dump()
    except MlflowException as e:
        error_msg = str(e)
        assert "Builtin scorers" in error_msg
        assert "Decorator-created scorers" in error_msg
        assert "@scorer decorator" in error_msg
        assert "Direct subclassing of Scorer is not supported" in error_msg


def test_builtin_scorer_with_aggregations_round_trip():
    from mlflow.genai.scorers.builtin_scorers import RelevanceToQuery

    scorer_with_aggs = RelevanceToQuery(name="relevance_with_aggs", aggregations=["mean", "max"])

    # Test that aggregations were set
    assert scorer_with_aggs.name == "relevance_with_aggs"
    assert scorer_with_aggs.aggregations == ["mean", "max"]

    # Round-trip serialization
    serialized = scorer_with_aggs.model_dump()
    deserialized = Scorer.model_validate(serialized)

    # Test properties preserved
    assert isinstance(deserialized, RelevanceToQuery)
    assert deserialized.name == "relevance_with_aggs"
    assert deserialized.aggregations == ["mean", "max"]
    assert hasattr(deserialized, "required_columns")
    assert deserialized.required_columns == {"inputs", "outputs"}

    # Test that both can be executed with mocking
    test_args = {
        "inputs": {"question": "What is machine learning?"},
        "outputs": "Machine learning is a subset of AI.",
    }

    with patch(
        "mlflow.genai.judges.is_context_relevant",
        return_value=Feedback(name="relevance_with_aggs", value="yes"),
    ) as mock_judge:
        # Test original scorer
        original_result = scorer_with_aggs(**test_args)

        # Test deserialized scorer
        deserialized_result = deserialized(**test_args)

    # Verify both results are equivalent
    assert original_result.name == deserialized_result.name == "relevance_with_aggs"
    assert original_result.value == deserialized_result.value == "yes"

    # Judge should be called twice (once for each scorer)
    assert mock_judge.call_count == 2


# ============================================================================
# COMPATIBILITY TESTS (Fixed serialized strings for backward compatibility)
# ============================================================================


def test_builtin_scorer_with_custom_name_compatibility():
    # Fixed serialized string for Guidelines scorer with custom name and parameters
    fixed_serialized_data = {
        "name": "custom_guidelines",
        "aggregations": ["mean", "max"],
        "mlflow_version": "3.1.0",
        "serialization_version": 1,
        "builtin_scorer_class": "Guidelines",
        "builtin_scorer_pydantic_data": {
            "name": "custom_guidelines",
            "aggregations": ["mean", "max"],
            "required_columns": ["inputs", "outputs"],
            "guidelines": [
                "Be polite and professional",
                "Provide accurate information",
            ],
        },
        "call_source": None,
        "call_signature": None,
        "original_func_name": None,
    }

    # Test deserialization
    deserialized = Scorer.model_validate(fixed_serialized_data)

    # Verify correct type and properties
    from mlflow.genai.scorers.builtin_scorers import Guidelines

    assert isinstance(deserialized, Guidelines)
    assert deserialized.name == "custom_guidelines"
    assert deserialized.aggregations == ["mean", "max"]
    assert deserialized.guidelines == [
        "Be polite and professional",
        "Provide accurate information",
    ]
    assert deserialized.required_columns == {"inputs", "outputs"}


def test_custom_scorer_compatibility_from_fixed_string():
    # Fixed serialized string representing a simple custom scorer
    fixed_serialized_data = {
        "name": "word_count_scorer",
        "aggregations": ["mean"],
        "mlflow_version": "3.1.0",
        "serialization_version": 1,
        "builtin_scorer_class": None,
        "builtin_scorer_pydantic_data": None,
        "call_source": "return len(outputs.split())",
        "call_signature": "(outputs)",
        "original_func_name": "word_count_scorer",
    }

    # Test deserialization
    deserialized = Scorer.model_validate(fixed_serialized_data)

    # Verify correct properties
    assert deserialized.name == "word_count_scorer"
    assert deserialized.aggregations == ["mean"]

    # Test functionality
    assert deserialized(outputs="hello world test") == 3
    assert deserialized(outputs="single") == 1
    assert deserialized(outputs="") == 0


def test_complex_custom_scorer_compatibility():
    # Fixed serialized string for a more complex custom scorer
    fixed_serialized_data = {
        "name": "length_comparison",
        "aggregations": None,
        "mlflow_version": "2.9.0",
        "serialization_version": 1,
        "builtin_scorer_class": None,
        "builtin_scorer_pydantic_data": None,
        "call_source": (
            "input_len = len(inputs) if inputs else 0\n"
            "output_len = len(outputs) if outputs else 0\n"
            "min_ratio = expectations.get('min_ratio', 1.0) if expectations else 1.0\n"
            "return output_len >= input_len * min_ratio"
        ),
        "call_signature": "(inputs, outputs, expectations)",
        "original_func_name": "length_comparison",
    }

    # Test deserialization
    deserialized = Scorer.model_validate(fixed_serialized_data)

    # Verify properties
    assert deserialized.name == "length_comparison"
    assert deserialized.aggregations is None

    # Test functionality with various inputs
    assert (
        deserialized(inputs="hello", outputs="hello world", expectations={"min_ratio": 1.5}) is True
    )  # 11 >= 5 * 1.5 (7.5)

    assert (
        deserialized(inputs="hello", outputs="hi", expectations={"min_ratio": 1.5}) is False
    )  # 2 < 5 * 1.5 (7.5)

    assert deserialized(inputs="test", outputs="test", expectations={}) is True  # 4 >= 4 * 1.0


def test_decorator_scorer_multiple_serialization_round_trips():
    @scorer
    def multi_round_scorer(outputs):
        return len(outputs) > 5

    # First serialization
    first_dump = multi_round_scorer.model_dump()

    # Deserialize
    recovered = Scorer.model_validate(first_dump)

    # Second serialization - this should work now with caching
    second_dump = recovered.model_dump()

    # Verify the dumps are identical
    assert first_dump == second_dump

    # Third serialization to ensure it's truly reusable
    third_dump = recovered.model_dump()
    assert first_dump == third_dump

    # Verify functionality is preserved
    assert recovered(outputs="hello world") is True
    assert recovered(outputs="hi") is False


def test_builtin_scorer_instructions_preserved_through_serialization():
    scorer = Guidelines(name="test_guidelines", guidelines=["Be helpful"])

    original_instructions = scorer.instructions

    serialized = scorer.model_dump()
    assert "builtin_scorer_pydantic_data" in serialized
    pydantic_data = serialized["builtin_scorer_pydantic_data"]

    assert "instructions" in pydantic_data
    assert pydantic_data["instructions"] == original_instructions

    deserialized = Scorer.model_validate(serialized)

    assert isinstance(deserialized, Guidelines)
    assert deserialized.instructions == original_instructions
    assert deserialized.name == "test_guidelines"
    assert deserialized.guidelines == ["Be helpful"]


# ============================================================================
# THIRD-PARTY SCORER (de)serialization
# ============================================================================


@pytest.mark.parametrize(
    ("third_party_data", "match"),
    [
        pytest.param(
            {
                "module": "os",
                "class": "system",
                "metric_name": "system",
                "model": None,
                "kwargs": {},
            },
            "not in the allow-list",
            id="module_not_allow_listed",
        ),
        pytest.param(
            {
                "module": "mlflow.genai.scorers.ragas",
                "class": "",
                "metric_name": "Faithfulness",
                "model": None,
                "kwargs": {},
            },
            "missing required fields",
            id="missing_class_name",
        ),
        pytest.param(
            {
                "module": "mlflow.genai.scorers.ragas",
                "class": "ExactMatch",
                "metric_name": "",
                "model": None,
                "kwargs": {},
            },
            "missing required fields",
            id="missing_metric_name",
        ),
    ],
)
def test_third_party_scorer_invalid_payload_rejected(third_party_data, match):
    payload = SerializedScorer(name="bad", third_party_scorer_data=third_party_data)
    with pytest.raises(MlflowException, match=match):
        Scorer.model_validate(payload)


def test_third_party_scorer_import_failure():
    payload = SerializedScorer(
        name="x",
        third_party_scorer_data={
            "module": "mlflow.genai.scorers.ragas",
            "class": "ExactMatch",
            "metric_name": "ExactMatch",
            "model": None,
            "kwargs": {},
        },
    )
    with patch(
        "mlflow.genai.scorers.base.importlib.import_module",
        side_effect=ImportError("library not installed"),
    ):
        with pytest.raises(MlflowException, match="could not import"):
            Scorer.model_validate(payload)


def test_third_party_scorer_class_not_found():
    payload = SerializedScorer(
        name="x",
        third_party_scorer_data={
            "module": "mlflow.genai.scorers.ragas",
            "class": "DoesNotExist",
            "metric_name": "X",
            "model": None,
            "kwargs": {},
        },
    )
    with patch(
        "mlflow.genai.scorers.base.importlib.import_module",
        return_value=Mock(spec=[]),
    ):
        with pytest.raises(MlflowException, match="not found in module"):
            Scorer.model_validate(payload)


def test_third_party_scorer_metric_name_mismatch_with_classvar():
    class RenamedScorer:
        metric_name: ClassVar[str] = "NewName"

    fake_module = Mock(RenamedScorer=RenamedScorer)
    payload = SerializedScorer(
        name="stale",
        third_party_scorer_data={
            "module": "mlflow.genai.scorers.ragas",
            "class": "RenamedScorer",
            "metric_name": "OldName",
            "model": None,
            "kwargs": {},
        },
    )
    with patch(
        "mlflow.genai.scorers.base.importlib.import_module",
        return_value=fake_module,
    ):
        with pytest.raises(MlflowException, match="does not match class"):
            Scorer.model_validate(payload)


def test_third_party_scorer_instantiation_failure():
    class BoomScorer:
        def __init__(self, **kwargs):
            raise RuntimeError("boom")

    fake_module = Mock(BoomScorer=BoomScorer)
    payload = SerializedScorer(
        name="x",
        third_party_scorer_data={
            "module": "mlflow.genai.scorers.ragas",
            "class": "BoomScorer",
            "metric_name": "X",
            "model": None,
            "kwargs": {},
        },
    )
    with patch(
        "mlflow.genai.scorers.base.importlib.import_module",
        return_value=fake_module,
    ):
        with pytest.raises(MlflowException, match="failed to instantiate"):
            Scorer.model_validate(payload)


def test_serialized_scorer_rejects_multiple_scorer_field_types():
    with pytest.raises(ValueError, match="cannot have multiple types"):
        SerializedScorer(
            name="oops",
            builtin_scorer_class="Safety",
            third_party_scorer_data={
                "module": "mlflow.genai.scorers.ragas",
                "class": "ExactMatch",
                "metric_name": "ExactMatch",
                "model": None,
                "kwargs": {},
            },
        )


def test_from_dict_round_trips_known_fields():
    @scorer(name="round_trip")
    def my_scorer(outputs):
        return outputs == "ok"

    payload = my_scorer.model_dump()
    restored = SerializedScorer.from_dict(payload)
    assert restored.name == "round_trip"
    assert restored.call_source is not None


def _deserialize_via_scorer_version(payload: dict[str, Any]) -> SerializedScorer:
    return ScorerVersion(
        experiment_id="123",
        scorer_name=payload["name"],
        scorer_version=1,
        serialized_scorer=json.dumps(payload),
        creation_time=0,
    ).serialized_scorer


def _formatted_error_logs(mock_logger) -> str:
    return " ".join(
        call.args[0] % call.args[1:] if len(call.args) > 1 else call.args[0]
        for call in mock_logger.error.call_args_list
    )


@pytest.mark.parametrize(
    "deserialize",
    [
        SerializedScorer.from_dict,
        Scorer.model_validate,
        _deserialize_via_scorer_version,
    ],
    ids=[
        "SerializedScorer.from_dict",
        "Scorer.model_validate",
        "ScorerVersion.serialized_scorer",
    ],
)
def test_unknown_field_is_dropped_with_version_aware_log(deserialize):
    payload = {
        "name": "future_scorer",
        "mlflow_version": "99.0.0",
        "call_source": "return 1",
        "call_signature": "(outputs)",
        "original_func_name": "future_scorer",
        "field_from_the_future": "value",
    }
    with patch("mlflow.genai.scorers.base._logger") as mock_logger:
        result = deserialize(payload)
    assert result.name == "future_scorer"
    assert not hasattr(result, "field_from_the_future")
    log_message = _formatted_error_logs(mock_logger)
    assert "field_from_the_future" in log_message
    assert "future_scorer" in log_message
    assert "99.0.0" in log_message
    assert mlflow.__version__ in log_message


def test_from_dict_unknown_field_falls_back_to_unknown_serialized_version():
    payload = {
        "name": "no_version",
        "call_source": "return 1",
        "call_signature": "(outputs)",
        "original_func_name": "no_version",
        "mystery_field": True,
    }
    with patch("mlflow.genai.scorers.base._logger") as mock_logger:
        result = SerializedScorer.from_dict(payload)
    assert result.name == "no_version"
    log_message = _formatted_error_logs(mock_logger)
    assert "unknown" in log_message
    assert "mystery_field" in log_message