fix(evaluation): exclude thought parts from text extraction across criteria

src/google/adk/evaluation/agent_evaluator.py

@@ -60,6 +60,7 @@
 from .eval_sets_manager import EvalSetsManager
 from .evaluator import EvalStatus
 from .in_memory_eval_sets_manager import InMemoryEvalSetsManager
+from .llm_as_judge_utils import get_text_from_content
 from .local_eval_sets_manager import convert_eval_set_to_pydantic_schema
 from .simulation.user_simulator_provider import UserSimulatorProvider
 
@@ -601,10 +602,7 @@ def _print_details(
 
   @staticmethod
   def _convert_content_to_text(content: Optional[genai_types.Content]) -> str:
-    if content and content.parts:
-      return "\n".join([p.text for p in content.parts if p.text])
-
-    return ""
+    return get_text_from_content(content) or ""
 
   @staticmethod
   def _convert_tool_calls_to_text(

src/google/adk/evaluation/final_response_match_v1.py

@@ -32,6 +32,7 @@
 from .evaluator import EvaluationResult
 from .evaluator import Evaluator
 from .evaluator import PerInvocationResult
+from .llm_as_judge_utils import get_text_from_content
 
 
 class RougeEvaluator(Evaluator):
@@ -90,10 +91,7 @@ def evaluate_invocations(
 
 
 def _get_text_from_content(content: Optional[genai_types.Content]) -> str:
-  if content and content.parts:
-    return "\n".join([part.text for part in content.parts if part.text])
-
-  return ""
+  return get_text_from_content(content) or ""
 
 
 def _get_eval_status(score: float, threshold: float) -> EvalStatus:

src/google/adk/evaluation/hallucinations_v1.py

@@ -46,6 +46,7 @@
 from .evaluator import PerInvocationResult
 from .llm_as_judge_utils import get_eval_status
 from .llm_as_judge_utils import get_text_from_content
+from .llm_as_judge_utils import get_text_parts
 from .llm_as_judge_utils import get_tool_declarations_as_json_str
 
 logger = logging.getLogger("google_adk." + __name__)
@@ -461,7 +462,7 @@ def _create_context_for_step(
           for part in event.content.parts
           if part.function_response
       ]
-      nl_responses = [part.text for part in event.content.parts if part.text]
+      nl_responses = get_text_parts(event.content)
 
       if nl_responses:
         context_parts.append("\n".join(nl_responses) + "\n")
@@ -650,11 +651,7 @@ def _get_steps_to_evaluate(self, actual: Invocation) -> list[EvaluationStep]:
 
     if self._criterion.evaluate_intermediate_nl_responses:
       for event in all_events:
-        nl_parts = (
-            [p.text for p in event.content.parts if p.text]
-            if event.content and event.content.parts
-            else []
-        )
+        nl_parts = get_text_parts(event.content)
         if nl_parts:
           context = self._create_context_for_step(
               actual.app_details, actual, events_for_context

src/google/adk/evaluation/llm_as_judge_utils.py

@@ -47,6 +47,13 @@ class Label(enum.Enum):
   NOT_FOUND = "label field not found"
 
 
+def get_text_parts(content: Optional[genai_types.Content]) -> list[str]:
+  """Returns the visible text parts of a `Content`, excluding thoughts."""
+  if not content or not content.parts:
+    return []
+  return [p.text for p in content.parts if p.text and not p.thought]
+
+
 def get_text_from_content(
     content: Optional[Union[genai_types.Content, Invocation]],
     *,
@@ -87,7 +94,7 @@ def get_text_from_content(
     return "\n".join(parts) if parts else None
 
   if content and content.parts:
-    return "\n".join([p.text for p in content.parts if p.text])
+    return "\n".join(get_text_parts(content))
 
   return None
 

src/google/adk/evaluation/rubric_based_multi_turn_trajectory_evaluator.py

@@ -30,6 +30,7 @@
 from .evaluator import _validate_invocation_lengths
 from .evaluator import EvaluationResult
 from .evaluator import PerInvocationResult
+from .llm_as_judge_utils import get_text_parts
 from .rubric_based_evaluator import RubricBasedEvaluator
 
 logger = logging.getLogger("google_adk." + __name__)
@@ -193,7 +194,7 @@ def _assemble_dialogue_history(
     for turn_index, invocation in enumerate(actual_invocations):
       # USER TURN
       if invocation.user_content and invocation.user_content.parts:
-        text_parts = [p.text for p in invocation.user_content.parts if p.text]
+        text_parts = get_text_parts(invocation.user_content)
         if text_parts:
           dialogue_lines.append(
               f"USER TURN {turn_index + 1}: {' '.join(text_parts)}"
@@ -208,7 +209,7 @@ def _assemble_dialogue_history(
               else f"AGENT ({event.author})"
           )
           if event.content and event.content.parts:
-            text_parts = [p.text for p in event.content.parts if p.text]
+            text_parts = get_text_parts(event.content)
             if text_parts:
               dialogue_lines.append(
                   f"{role} TURN {turn_index + 1}: {' '.join(text_parts)}"
@@ -245,7 +246,7 @@ def _assemble_dialogue_history(
         ):
           agent_name = intermediate_data.invocation_events[0].author
         role = f"AGENT ({agent_name})"
-        text_parts = [p.text for p in invocation.final_response.parts if p.text]
+        text_parts = get_text_parts(invocation.final_response)
         if text_parts:
           dialogue_lines.append(
               f"{role} TURN {turn_index + 1}: {' '.join(text_parts)}"

src/google/adk/evaluation/vertex_ai_eval_facade.py

@@ -37,6 +37,7 @@
 from .evaluator import EvaluationResult
 from .evaluator import Evaluator
 from .evaluator import PerInvocationResult
+from .llm_as_judge_utils import get_text_from_content
 
 logger = logging.getLogger("google_adk." + __name__)
 
@@ -113,10 +114,7 @@ def evaluate_invocations(
     """
 
   def _get_text(self, content: Optional[genai_types.Content]) -> str:
-    if content and content.parts:
-      return "\n".join([p.text for p in content.parts if p.text])
-
-    return ""
+    return get_text_from_content(content) or ""
 
   def _get_score(self, eval_result: object) -> Optional[float]:
     summary_metrics: object = getattr(eval_result, "summary_metrics", None)

tests/unittests/evaluation/test_agent_evaluator.py

@@ -488,6 +488,17 @@ def _make_result_with_invocation(
   )
 
 
+def test_convert_content_to_text_excludes_thought_parts():
+  content = genai_types.Content(
+      parts=[
+          genai_types.Part(text="Consider the options.", thought=True),
+          genai_types.Part(text="Paris"),
+      ]
+  )
+
+  assert AgentEvaluator._convert_content_to_text(content) == "Paris"
+
+
 def test_get_results_as_rows_flattens_metrics_and_invocations():
   eval_metric_results = {
       "response_match_score": [

tests/unittests/evaluation/test_final_response_match_v1.py

@@ -59,6 +59,62 @@ def _create_test_invocations(
   )
 
 
+@pytest.mark.parametrize(
+    "actual_parts, expected_parts, score",
+    [
+        (
+            [
+                genai_types.Part(
+                    text="Consider the capital of France.", thought=True
+                ),
+                genai_types.Part(text="Paris"),
+            ],
+            [genai_types.Part(text="Paris")],
+            1.0,
+        ),
+        (
+            [genai_types.Part(text="Paris", thought=False)],
+            [
+                genai_types.Part(
+                    text="Consider the capital of France.", thought=True
+                ),
+                genai_types.Part(text="Paris"),
+            ],
+            1.0,
+        ),
+        (
+            [genai_types.Part(text="Paris", thought=True)],
+            [genai_types.Part(text="Paris")],
+            0.0,
+        ),
+        (
+            [
+                genai_types.Part(text="Paris", thought=True),
+                genai_types.Part(text="London"),
+            ],
+            [genai_types.Part(text="Paris")],
+            0.0,
+        ),
+    ],
+)
+def test_response_match_scores_visible_text_only(
+    actual_parts, expected_parts, score
+):
+  """Thought summaries neither dilute correct answers nor credit wrong ones."""
+  actual, expected = _create_test_invocations("", "")
+  actual.final_response.parts = actual_parts
+  expected.final_response.parts = expected_parts
+  evaluator = _create_test_rouge_evaluator(threshold=0.5)
+
+  result = evaluator.evaluate_invocations([actual], [expected])
+
+  assert result.overall_score == pytest.approx(score)
+  assert result.per_invocation_results[0].score == pytest.approx(score)
+  assert result.overall_eval_status == (
+      EvalStatus.PASSED if score == 1.0 else EvalStatus.FAILED
+  )
+
+
 def test_calculate_rouge_1_scores_empty_candidate_and_reference():
   candidate = ""
   reference = ""

tests/unittests/evaluation/test_final_response_match_v2.py

@@ -268,6 +268,54 @@ def test_format_auto_rater_prompt_includes_intermediate_when_enabled():
   assert "reference intro\nreference final" in prompt
 
 
+def test_format_auto_rater_prompt_excludes_thought_parts():
+  evaluator = _create_test_evaluator_gemini(threshold=0.8)
+  actual_invocation, expected_invocation = _create_test_invocations(
+      "candidate text", "reference text"
+  )
+  actual_invocation.final_response.parts.insert(
+      0, genai_types.Part(text="Considering the response.", thought=True)
+  )
+  expected_invocation.final_response.parts.insert(
+      0, genai_types.Part(text="Considering the reference.", thought=True)
+  )
+
+  prompt = evaluator.format_auto_rater_prompt(
+      actual_invocation, expected_invocation
+  )
+
+  assert "Considering the response." not in prompt
+  assert "Considering the reference." not in prompt
+  assert "candidate text" in prompt
+  assert "reference text" in prompt
+
+
+def test_convert_auto_rater_response_to_score_ignores_judge_thought():
+  """The judge model's own thought text must not break verdict parsing."""
+  evaluator = _create_test_evaluator_gemini(threshold=0.8)
+  auto_rater_response = """```json
+{
+  "is_the_agent_response_valid": "valid",
+  "reasoning": "The response is valid."
+}
+```"""
+  llm_response = LlmResponse(
+      content=genai_types.Content(
+          parts=[
+              genai_types.Part(
+                  text="Let me evaluate this response.", thought=True
+              ),
+              genai_types.Part(text=auto_rater_response),
+          ],
+          role="model",
+      )
+  )
+  auto_rater_score = evaluator.convert_auto_rater_response_to_score(
+      llm_response
+  )
+  assert auto_rater_score == AutoRaterScore(score=1.0)
+
+
 def test_convert_auto_rater_response_to_score_valid():
   evaluator = _create_test_evaluator_gemini(threshold=0.8)
   auto_rater_response = """```json