[evaluation] Consolidate message-preprocessing helpers into a single source (_common.utils) (#48013)

The helpers _is_intermediate_response, _drop_mcp_approval_messages, _normalize_function_call_types, and _preprocess_messages had divergent duplicate copies in both _evaluators/_common/_base_prompty_eval.py and _common/utils.py. The utils.py copy of _normalize_function_call_types was missing the function_call payload key-rename to tool_call. This makes _common.utils the single source of truth (adding the missing rename so it is a strict superset) and re-exports the helpers from _base_prompty_eval for backward compatibility. Internal refactor only; no public API or behavior change.

Author: m7md7sien

sdk/evaluation/azure-ai-evaluation/azure/ai/evaluation/_common/utils.py

@@ -1626,6 +1626,8 @@ def _normalize_function_call_types(messages):
             t = item.get("type")
             if t == "function_call":
                 item["type"] = "tool_call"
+                if "function_call" in item:
+                    item["tool_call"] = item.pop("function_call")
             elif t == "function_call_output":
                 item["type"] = "tool_result"
                 if "function_call_output" in item:

sdk/evaluation/azure-ai-evaluation/azure/ai/evaluation/_evaluators/_common/_base_prompty_eval.py

@@ -19,6 +19,17 @@
 from azure.ai.evaluation._constants import EVALUATION_PASS_FAIL_MAPPING
 from azure.ai.evaluation._exceptions import EvaluationException, ErrorBlame, ErrorCategory, ErrorTarget
 from ..._common.utils import construct_prompty_model_config, validate_model_config, parse_quality_evaluator_reason_score
+
+# The message-preprocessing helpers now live in ``azure.ai.evaluation._common.utils`` (single source of
+# truth). They are imported here so this module can use them and so callers that historically imported
+# them from this module keep working. ``_drop_mcp_approval_messages`` / ``_normalize_function_call_types``
+# are re-exported only (used indirectly via ``_preprocess_messages``).
+from ..._common.utils import (  # pylint: disable=unused-import
+    _is_intermediate_response,
+    _drop_mcp_approval_messages,
+    _normalize_function_call_types,
+    _preprocess_messages,
+)
 from . import EvaluatorBase
 
 try:
@@ -34,80 +45,6 @@ def value(self) -> str:
 T = TypeVar("T")
 
 
-def _is_intermediate_response(response):
-    """Check if response is intermediate (last content item is function_call or mcp_approval_request)."""
-    if isinstance(response, list) and len(response) > 0:
-        last_msg = response[-1]
-        if isinstance(last_msg, dict) and last_msg.get("role") == "assistant":
-            content = last_msg.get("content", [])
-            if isinstance(content, list) and len(content) > 0:
-                last_content = content[-1]
-                if isinstance(last_content, dict) and last_content.get("type") in (
-                    "function_call",
-                    "mcp_approval_request",
-                ):
-                    return True
-    return False
-
-
-def _drop_mcp_approval_messages(messages):
-    """Remove MCP approval request/response messages."""
-    if not isinstance(messages, list):
-        return messages
-    return [
-        msg
-        for msg in messages
-        if not (
-            isinstance(msg, dict)
-            and isinstance(msg.get("content"), list)
-            and (
-                (
-                    msg.get("role") == "assistant"
-                    and any(isinstance(c, dict) and c.get("type") == "mcp_approval_request" for c in msg["content"])
-                )
-                or (
-                    msg.get("role") == "tool"
-                    and any(isinstance(c, dict) and c.get("type") == "mcp_approval_response" for c in msg["content"])
-                )
-            )
-        )
-    ]
-
-
-def _normalize_function_call_types(messages):
-    """Normalize function_call/function_call_output/openapi_call/openapi_call_output types to tool_call/tool_result."""
-    if not isinstance(messages, list):
-        return messages
-    for msg in messages:
-        if isinstance(msg, dict) and isinstance(msg.get("content"), list):
-            for item in msg["content"]:
-                if not isinstance(item, dict):
-                    continue
-                item_type = item.get("type")
-                if item_type == "function_call":
-                    item["type"] = "tool_call"
-                    if "function_call" in item:
-                        item["tool_call"] = item.pop("function_call")
-                elif item_type == "function_call_output":
-                    item["type"] = "tool_result"
-                    if "function_call_output" in item:
-                        item["tool_result"] = item.pop("function_call_output")
-                elif item_type == "openapi_call":
-                    item["type"] = "tool_call"
-                elif item_type == "openapi_call_output":
-                    item["type"] = "tool_result"
-                    if "openapi_call_output" in item:
-                        item["tool_result"] = item.pop("openapi_call_output")
-    return messages
-
-
-def _preprocess_messages(messages):
-    """Drop MCP approval messages and normalize function call types."""
-    messages = _drop_mcp_approval_messages(messages)
-    messages = _normalize_function_call_types(messages)
-    return messages
-
-
 class PromptyEvaluatorBase(EvaluatorBase[T]):
     """Base class for all evaluators that make use of context as an input. It's also assumed that such evaluators
     make use of a prompty file, and return their results as a dictionary, with a single key-value pair

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

@@ -746,6 +746,20 @@ def test_function_call_output_to_tool_result(self):
         assert result[0]["content"][0]["tool_result"] == "result data"
         assert "function_call_output" not in result[0]["content"][0]
 
+    def test_function_call_nested_payload_key_renamed(self):
+        # A nested ``function_call`` payload key is renamed to ``tool_call`` so downstream code that
+        # reads ``content["tool_call"]`` finds it (canonical single-source behavior).
+        messages = [
+            {
+                "role": "assistant",
+                "content": [{"type": "function_call", "function_call": {"name": "f", "arguments": {}}}],
+            }
+        ]
+        result = _normalize_function_call_types(messages)
+        assert result[0]["content"][0]["type"] == "tool_call"
+        assert result[0]["content"][0]["tool_call"] == {"name": "f", "arguments": {}}
+        assert "function_call" not in result[0]["content"][0]
+
     def test_openapi_call_to_tool_call(self):
         messages = [
             {
@@ -779,6 +793,21 @@ def test_non_list_passthrough(self):
         assert _normalize_function_call_types("not a list") == "not a list"
 
 
+@pytest.mark.unittest
+class TestPreprocessHelpersSingleSource:
+    """The message-preprocessing helpers have one definition in ``_common.utils`` and are re-exported
+    from ``_base_prompty_eval`` for backward compatibility."""
+
+    def test_base_prompty_eval_reexports_the_same_utils_helpers(self):
+        from azure.ai.evaluation._common import utils as u
+        from azure.ai.evaluation._evaluators._common import _base_prompty_eval as b
+
+        assert b._is_intermediate_response is u._is_intermediate_response
+        assert b._drop_mcp_approval_messages is u._drop_mcp_approval_messages
+        assert b._normalize_function_call_types is u._normalize_function_call_types
+        assert b._preprocess_messages is u._preprocess_messages
+
+
 @pytest.mark.unittest
 class TestPreprocessMessages:
     def test_drops_mcp_and_normalizes(self):