fix: eval dataframe should allow a column to be all null if type does…

…n't match expected (#5495)

src/phoenix/trace/span_evaluations.py

@@ -66,14 +66,15 @@ class NeedsResultColumns(ABC):
     )
 
     @classmethod
-    def is_valid_result_columns(cls, dtypes: "pd.Series[Any]") -> bool:
+    def is_valid_result_columns(cls, df: pd.DataFrame) -> bool:
+        dtypes = df.dtypes
         names = cls.result_column_names.keys()
         intersection = dtypes.index.intersection(names)  # type: ignore
         if not len(intersection):
             return False
         for name in intersection:
             check_type = cls.result_column_names[name]
-            if not check_type(dtypes[name]):
+            if not check_type(dtypes[name]) and not df.loc[:, name].isna().all():
                 return False
         return True
 
@@ -138,7 +139,7 @@ def _clean_dataframe(self, dataframe: pd.DataFrame) -> pd.DataFrame:
             )
 
         # Validate that the dataframe contains result columns of appropriate types.
-        if not self.is_valid_result_columns(dataframe.dtypes):
+        if not self.is_valid_result_columns(dataframe):
             raise ValueError(
                 f"The dataframe must contain one of these columns with appropriate "
                 f"value types: {self.result_column_names.keys()} "

tests/unit/trace/test_span_evaluations.py

@@ -9,6 +9,7 @@
 import pandas as pd
 import pyarrow
 import pytest
+from pandas.core.dtypes.common import is_numeric_dtype
 from pandas.testing import assert_frame_equal
 from pyarrow import parquet
 
@@ -42,6 +43,20 @@ def test_span_evaluations_construction() -> None:
     assert "score" in eval_ds.dataframe.columns
 
 
+def test_span_evaluations_construction_allows_all_null_column() -> None:
+    num_records = 5
+    span_ids = [f"span_{index}" for index in range(num_records)]
+    df = pd.DataFrame(
+        {
+            "context.span_id": span_ids,
+            "label": [str(index) for index in range(num_records)],
+            "score": [None] * num_records,
+        }
+    ).set_index("context.span_id")
+    assert not is_numeric_dtype(df.loc[:, "score"])
+    SpanEvaluations(eval_name="my_eval", dataframe=df)
+
+
 def power_set(s: list[tuple[str, Any]]) -> Iterator[dict[str, Any]]:
     for result in chain.from_iterable(combinations(s, r) for r in range(len(s) + 1)):
         yield dict(result)