fix: read and write eval data files as utf-8

Merge #6298

PiperOrigin-RevId: 963885296

Author: anxkhn

src/google/adk/evaluation/agent_evaluator.py

@@ -94,7 +94,7 @@ def __call__(self) -> Awaitable[tuple[BaseAgent, object]]:
 
 
 def load_json(file_path: str) -> Union[Dict[str, Any], List[Any]]:
-  with open(file_path, "r") as f:
+  with open(file_path, "r", encoding="utf-8") as f:
     return cast(Union[Dict[str, Any], List[Any]], json.load(f))
 
 
@@ -365,7 +365,7 @@ def migrate_eval_data_to_new_schema(
         old_eval_data_file, eval_config, initial_session
     )
 
-    with open(new_eval_data_file, "w") as f:
+    with open(new_eval_data_file, "w", encoding="utf-8") as f:
       f.write(eval_set.model_dump_json(indent=2))
 
   @staticmethod
@@ -424,7 +424,7 @@ def _get_initial_session(
   ) -> dict[str, Any]:
     initial_session: dict[str, Any] = {}
     if initial_session_file:
-      with open(initial_session_file, "r") as f:
+      with open(initial_session_file, "r", encoding="utf-8") as f:
         initial_session = json.loads(f.read())
     return initial_session
 

src/google/adk/evaluation/evaluation_generator.py

@@ -596,7 +596,7 @@ def generate_responses_from_session(
     """
     results = []
 
-    with open(session_path, "r") as f:
+    with open(session_path, "r", encoding="utf-8") as f:
       session_data = Session.model_validate_json(f.read())
       logger.info("Loaded session %s", session_path)
 

tests/unittests/evaluation/test_agent_evaluator.py

@@ -16,6 +16,7 @@
 
 from __future__ import annotations
 
+import builtins
 import json
 import os
 from pathlib import Path
@@ -25,8 +26,10 @@
 from google.adk.agents.base_agent import BaseAgent
 from google.adk.apps.app import App
 from google.adk.artifacts.in_memory_artifact_service import InMemoryArtifactService
+from google.adk.evaluation import agent_evaluator as agent_evaluator_module
 from google.adk.evaluation.agent_evaluator import _EvalMetricResultWithInvocation
 from google.adk.evaluation.agent_evaluator import AgentEvaluator
+from google.adk.evaluation.agent_evaluator import load_json
 from google.adk.evaluation.eval_case import EvalCase
 from google.adk.evaluation.eval_case import Invocation
 from google.adk.evaluation.eval_config import EvalConfig
@@ -42,6 +45,9 @@
 import pandas as pd
 import pytest
 
+_NON_ASCII_TEXT = "😀 你好 café"
+_real_open = builtins.open
+
 
 def _make_eval_set() -> EvalSet:
   return EvalSet(
@@ -979,5 +985,89 @@ async def test_evaluate_keeps_positional_initial_session_file_and_print_flag(
   )
 
 
+def _non_utf8_default_open(file, mode="r", *args, **kwargs):
+  """Emulates a platform whose default text encoding is not UTF-8.
+
+  On such platforms (for example Windows, where the default is cp1252),
+  `open()` calls that omit `encoding=` inherit that non-UTF-8 default. This
+  wrapper reproduces that behaviour on any platform by falling back to ASCII
+  when a text-mode open does not specify an encoding, so a missing
+  `encoding="utf-8"` argument raises instead of silently depending on the
+  host locale.
+  """
+  if "b" not in mode and "encoding" not in kwargs:
+    kwargs["encoding"] = "ascii"
+  return _real_open(file, mode, *args, **kwargs)
+
+
+def test_load_json_reads_non_ascii_with_non_utf8_default(tmp_path, mocker):
+  """`load_json` must decode eval data as UTF-8 regardless of platform locale."""
+  file_path = tmp_path / "eval.json"
+  file_path.write_text(
+      json.dumps([{"query": _NON_ASCII_TEXT}], ensure_ascii=False),
+      encoding="utf-8",
+  )
+
+  mocker.patch.object(
+      agent_evaluator_module, "open", _non_utf8_default_open, create=True
+  )
+
+  assert load_json(str(file_path)) == [{"query": _NON_ASCII_TEXT}]
+
+
+def test_get_initial_session_reads_non_ascii_with_non_utf8_default(
+    tmp_path, mocker
+):
+  """`_get_initial_session` must decode the session file as UTF-8."""
+  session_file = tmp_path / "initial_session.json"
+  session_file.write_text(
+      json.dumps({"state": {"city": _NON_ASCII_TEXT}}, ensure_ascii=False),
+      encoding="utf-8",
+  )
+
+  mocker.patch.object(
+      agent_evaluator_module, "open", _non_utf8_default_open, create=True
+  )
+
+  initial_session = AgentEvaluator._get_initial_session(str(session_file))
+
+  assert initial_session == {"state": {"city": _NON_ASCII_TEXT}}
+
+
+def test_migrate_eval_data_round_trips_non_ascii_with_non_utf8_default(
+    tmp_path, mocker
+):
+  """Migration must read the old file and write the new file as UTF-8.
+
+  This exercises both the read (`load_json`) and the write
+  (`model_dump_json`) of eval data, which must stay UTF-8 consistent so that
+  datasets containing non-ASCII characters survive migration on any platform.
+  """
+  old_eval_data_file = tmp_path / "old_format.test.json"
+  old_eval_data_file.write_text(
+      json.dumps(
+          [{
+              "query": _NON_ASCII_TEXT,
+              "reference": _NON_ASCII_TEXT,
+              "expected_tool_use": [],
+          }],
+          ensure_ascii=False,
+      ),
+      encoding="utf-8",
+  )
+  new_eval_data_file = tmp_path / "new_format.json"
+
+  mocker.patch.object(
+      agent_evaluator_module, "open", _non_utf8_default_open, create=True
+  )
+
+  AgentEvaluator.migrate_eval_data_to_new_schema(
+      str(old_eval_data_file), str(new_eval_data_file)
+  )
+
+  migrated = json.loads(new_eval_data_file.read_text(encoding="utf-8"))
+  assert _NON_ASCII_TEXT in json.dumps(migrated, ensure_ascii=False)
+
+
 if __name__ == "__main__":
   raise SystemExit(pytest.main([__file__, "-v"]))

tests/unittests/evaluation/test_evaluation_generator.py

@@ -15,9 +15,11 @@
 from __future__ import annotations
 
 import asyncio
+import builtins
 
 from google.adk.agents.base_agent import BaseAgent
 from google.adk.apps.app import App
+from google.adk.evaluation import evaluation_generator as evaluation_generator_module
 from google.adk.evaluation.app_details import AgentDetails
 from google.adk.evaluation.app_details import AppDetails
 from google.adk.evaluation.conversation_scenarios import ConversationScenario
@@ -1991,3 +1993,65 @@ def test_generate_responses_from_session_scopes_by_invocation_id(tmp_path):
 
   assert results[0][0]["actual_tool_use"] == []
   assert results[0][0]["response"] == "I rolled a 4."
+
+
+_real_open = builtins.open
+
+
+def _non_utf8_default_open(file, mode="r", *args, **kwargs):
+  """Emulates a platform whose default text encoding is not UTF-8.
+
+  Falls back to ASCII when a text-mode open does not specify an encoding, so a
+  missing `encoding="utf-8"` argument raises instead of silently depending on
+  the host locale (for example cp1252 on Windows).
+  """
+  if "b" not in mode and "encoding" not in kwargs:
+    kwargs["encoding"] = "ascii"
+  return _real_open(file, mode, *args, **kwargs)
+
+
+def test_generate_responses_from_session_reads_non_ascii_with_non_utf8_default(
+    tmp_path, mocker
+):
+  """The session file must be read as UTF-8 regardless of platform locale.
+
+  Session files serialized via `model_dump_json` contain raw (unescaped)
+  non-ASCII characters, so reading them without an explicit UTF-8 encoding
+  fails on platforms whose default encoding is not UTF-8.
+  """
+  non_ascii_text = "😀 你好 café"
+  session = Session(
+      id="s1",
+      app_name="app",
+      user_id="u1",
+      events=[
+          Event(
+              author="user",
+              invocation_id="inv1",
+              content=types.Content(
+                  role="user", parts=[types.Part(text=non_ascii_text)]
+              ),
+          ),
+          Event(
+              author="agent",
+              invocation_id="inv1",
+              content=types.Content(
+                  role="model",
+                  parts=[types.Part(text="response " + non_ascii_text)],
+              ),
+          ),
+      ],
+  )
+  session_path = tmp_path / "session.json"
+  session_path.write_text(session.model_dump_json(), encoding="utf-8")
+
+  mocker.patch.object(
+      evaluation_generator_module, "open", _non_utf8_default_open, create=True
+  )
+
+  results = EvaluationGenerator.generate_responses_from_session(
+      str(session_path), [[{"query": non_ascii_text}]]
+  )
+
+  assert results[0][0]["query"] == non_ascii_text
+  assert results[0][0]["response"] == "response " + non_ascii_text