Revert "fix(eval): include intermediate text in final response match" (#5887)

Author: DeanChensj

src/google/adk/evaluation/final_response_match_v2.py

@@ -159,22 +159,13 @@ def format_auto_rater_prompt(
     if expected_invocation is None:
       raise ValueError("expected_invocation is required for this metric.")
 
-    include_intermediate = (
-        self._criterion.include_intermediate_responses_in_final
-    )
-    reference = get_text_from_content(
-        expected_invocation,
-        include_intermediate_responses_in_final=include_intermediate,
-    )
-    response = get_text_from_content(
-        actual_invocation,
-        include_intermediate_responses_in_final=include_intermediate,
-    )
+    reference = get_text_from_content(expected_invocation.final_response)
+    response = get_text_from_content(actual_invocation.final_response)
     user_prompt = get_text_from_content(expected_invocation.user_content)
     return self._auto_rater_prompt_template.format(
         prompt=user_prompt,
-        response=response or "",
-        golden_response=reference or "",
+        response=response,
+        golden_response=reference,
     )
 
   @override

src/google/adk/evaluation/llm_as_judge_utils.py

@@ -25,7 +25,6 @@
 from .app_details import AppDetails
 from .common import EvalBaseModel
 from .eval_case import get_all_tool_calls_with_responses
-from .eval_case import IntermediateData
 from .eval_case import IntermediateDataType
 from .eval_case import Invocation
 from .eval_case import InvocationEvents
@@ -72,12 +71,6 @@ def get_text_from_content(
         text = get_text_from_content(event.content)
         if text:
           parts.append(text)
-    elif isinstance(content.intermediate_data, IntermediateData):
-      for _, response_parts in content.intermediate_data.intermediate_responses:
-        text = get_text_from_content(genai_types.Content(parts=response_parts))
-        if text:
-          parts.append(text)
-
     # Then fetch the final response text and append it to the end.
     final_text = get_text_from_content(content.final_response)
     if final_text:

tests/unittests/evaluation/test_final_response_match_v2.py

@@ -15,8 +15,6 @@
 from __future__ import annotations
 
 from google.adk.evaluation.eval_case import Invocation
-from google.adk.evaluation.eval_case import InvocationEvent
-from google.adk.evaluation.eval_case import InvocationEvents
 from google.adk.evaluation.eval_metrics import BaseCriterion
 from google.adk.evaluation.eval_metrics import EvalMetric
 from google.adk.evaluation.eval_metrics import EvalStatus
@@ -129,18 +127,13 @@ def create_test_template() -> str:
 
 def _create_test_evaluator_gemini(
     threshold: float,
-    *,
-    include_intermediate_responses_in_final: bool = False,
 ) -> FinalResponseMatchV2Evaluator:
   evaluator = FinalResponseMatchV2Evaluator(
       EvalMetric(
           metric_name="final_response_match_v2",
           threshold=threshold,
           criterion=BaseCriterion(
               threshold=0.5,
-              include_intermediate_responses_in_final=(
-                  include_intermediate_responses_in_final
-              ),
           ),
       ),
   )
@@ -175,21 +168,6 @@ def _create_test_invocations(
   return actual_invocation, expected_invocation
 
 
-def _add_intermediate_text(invocation: Invocation, text: str) -> Invocation:
-  invocation.intermediate_data = InvocationEvents(
-      invocation_events=[
-          InvocationEvent(
-              author="agent",
-              content=genai_types.Content(
-                  parts=[genai_types.Part(text=text)],
-                  role="model",
-              ),
-          ),
-      ]
-  )
-  return invocation
-
-
 def test_format_auto_rater_prompt():
   evaluator = _create_test_evaluator_gemini(threshold=0.8)
   actual_invocation, expected_invocation = _create_test_invocations(
@@ -215,59 +193,6 @@ def test_format_auto_rater_prompt():
 """
 
 
-def test_format_auto_rater_prompt_uses_empty_text_for_missing_final_response():
-  evaluator = _create_test_evaluator_gemini(threshold=0.8)
-  actual_invocation, expected_invocation = _create_test_invocations(
-      "candidate text", "reference text"
-  )
-  actual_invocation.final_response = None
-  expected_invocation.final_response = None
-
-  prompt = evaluator.format_auto_rater_prompt(
-      actual_invocation, expected_invocation
-  )
-
-  assert "None" not in prompt
-  assert '"Agent response": ,' in prompt
-  assert '"Reference response": ,' in prompt
-
-
-def test_format_auto_rater_prompt_ignores_intermediate_by_default():
-  evaluator = _create_test_evaluator_gemini(threshold=0.8)
-  actual_invocation, expected_invocation = _create_test_invocations(
-      "candidate final", "reference final"
-  )
-  _add_intermediate_text(actual_invocation, "candidate intro")
-  _add_intermediate_text(expected_invocation, "reference intro")
-
-  prompt = evaluator.format_auto_rater_prompt(
-      actual_invocation, expected_invocation
-  )
-
-  assert "candidate final" in prompt
-  assert "reference final" in prompt
-  assert "candidate intro" not in prompt
-  assert "reference intro" not in prompt
-
-
-def test_format_auto_rater_prompt_includes_intermediate_when_enabled():
-  evaluator = _create_test_evaluator_gemini(
-      threshold=0.8, include_intermediate_responses_in_final=True
-  )
-  actual_invocation, expected_invocation = _create_test_invocations(
-      "candidate final", "reference final"
-  )
-  _add_intermediate_text(actual_invocation, "candidate intro")
-  _add_intermediate_text(expected_invocation, "reference intro")
-
-  prompt = evaluator.format_auto_rater_prompt(
-      actual_invocation, expected_invocation
-  )
-
-  assert "candidate intro\ncandidate final" in prompt
-  assert "reference intro\nreference final" in prompt
-
-
 def test_convert_auto_rater_response_to_score_valid():
   evaluator = _create_test_evaluator_gemini(threshold=0.8)
   auto_rater_response = """```json

tests/unittests/evaluation/test_llm_as_judge_utils.py

@@ -132,36 +132,6 @@ def test_get_text_from_content_with_invocation_include_intermediate_responses_in
   )
 
 
-def test_get_text_from_content_with_intermediate_data_full_response():
-  invocation = Invocation(
-      user_content=genai_types.Content(parts=[genai_types.Part(text="user")]),
-      intermediate_data=IntermediateData(
-          intermediate_responses=[
-              ("agent", [genai_types.Part(text="legacy intro")]),
-              (
-                  "tool",
-                  [
-                      genai_types.Part(
-                          function_call=genai_types.FunctionCall(name="lookup")
-                      )
-                  ],
-              ),
-          ]
-      ),
-      final_response=genai_types.Content(
-          parts=[genai_types.Part(text="final answer")]
-      ),
-  )
-
-  assert get_text_from_content(invocation) == "final answer"
-  assert (
-      get_text_from_content(
-          invocation, include_intermediate_responses_in_final=True
-      )
-      == "legacy intro\nfinal answer"
-  )
-
-
 def test_get_eval_status_with_none_score():
   """Tests get_eval_status returns NOT_EVALUATED for a None score."""
   assert get_eval_status(score=None, threshold=0.5) == EvalStatus.NOT_EVALUATED