Fix CSV quote stripping in custom evaluators

sdk/evaluation/azure-ai-evaluation/azure/ai/evaluation/_evaluate/_utils.py

@@ -1,6 +1,7 @@
 # ---------------------------------------------------------
 # Copyright (c) Microsoft Corporation. All rights reserved.
 # ---------------------------------------------------------
+import csv
 import json
 import logging
 import os
@@ -479,7 +480,11 @@ def __init__(self, filename: Union[os.PathLike, str]):
         self.filename = filename
 
     def load(self) -> pd.DataFrame:
-        return pd.read_csv(self.filename, dtype=str)
+        # Use QUOTE_NONE to preserve quotation marks as literal characters in cell values.
+        # By default, pandas treats quotes as CSV field delimiters and strips them.
+        # This ensures that values like "test" are read as "test" (with quotes), not test.
+        # The escapechar allows escaping special characters like commas within values.
+        return pd.read_csv(self.filename, dtype=str, quoting=csv.QUOTE_NONE, escapechar="\\")
 
 
 class DataLoaderFactory:

sdk/evaluation/azure-ai-evaluation/tests/unittests/data/test_csv_quotes.csv

@@ -0,0 +1,4 @@
+response,ground_truth
+test,"test"
+"quoted",quoted
+start,"""end"

sdk/evaluation/azure-ai-evaluation/tests/unittests/test_evaluate.py

@@ -1652,3 +1652,58 @@ def test_log_metrics_and_instance_results_onedp_no_redundant_tags(self, mock_cli
         call_args = mock_client.start_evaluation_run.call_args
         eval_upload = call_args[1]["evaluation"]
         assert eval_upload.tags == tags
+
+    def test_csv_preserves_quotes_in_values(self):
+        """Test that CSV loading preserves quotes in cell values.
+
+        This test validates the fix for the issue where custom code evaluators
+        were dropping leading and trailing quotation marks from parameter values.
+        The issue occurs when a CSV cell value starts AND ends with quotes.
+
+        Test CSV contains:
+        - Row 0: test,"test" - unquoted vs quoted value
+        - Row 1: "quoted",quoted - quoted vs unquoted value
+        - Row 2: start,"""end" - tests that even multiple quotes are preserved as literals
+
+        With QUOTE_NONE, all quotes are treated as literal characters, not delimiters.
+        """
+        # Get the test CSV file
+        csv_file = _get_file("test_csv_quotes.csv")
+
+        # Define a custom evaluator that checks if quotes are preserved
+        def quote_checker(response: str, ground_truth: str):
+            """Custom evaluator that checks if values match exactly."""
+            return {
+                "match": 1 if response == ground_truth else 0,
+                "response_value": response,
+                "ground_truth_value": ground_truth,
+            }
+
+        # Run evaluation with the custom evaluator
+        result = evaluate(
+            data=csv_file,
+            evaluators={"quote_checker": quote_checker},
+        )
+
+        # Verify the results
+        assert result is not None
+        row_result_df = pd.DataFrame(result["rows"])
+
+        # Check that we have the expected rows
+        assert len(row_result_df) == 3
+
+        # Row 0: response='test', ground_truth='"test"' - should NOT match
+        assert row_result_df["outputs.quote_checker.response_value"][0] == "test"
+        assert row_result_df["outputs.quote_checker.ground_truth_value"][0] == '"test"'
+        assert row_result_df["outputs.quote_checker.match"][0] == 0
+
+        # Row 1: response='"quoted"', ground_truth='quoted' - should NOT match
+        assert row_result_df["outputs.quote_checker.response_value"][1] == '"quoted"'
+        assert row_result_df["outputs.quote_checker.ground_truth_value"][1] == "quoted"
+        assert row_result_df["outputs.quote_checker.match"][1] == 0
+
+        # Row 2: response='start', ground_truth='"""end"' - should NOT match
+        # Note: With QUOTE_NONE, """end" is read as the literal string """end"
+        assert row_result_df["outputs.quote_checker.response_value"][2] == 'start'
+        assert row_result_df["outputs.quote_checker.ground_truth_value"][2] == '"""end"'
+        assert row_result_df["outputs.quote_checker.match"][2] == 0