fix(memory): match a Latin word embedded in unspaced-script text

src/google/adk/memory/in_memory_memory_service.py

@@ -32,6 +32,7 @@
 
 _UNKNOWN_SESSION_ID = '__unknown_session_id__'
 _MAX_SEARCH_RESULTS = 10
+_SCRIPT_RUN_PATTERN = re.compile(r'[\x00-\x7f]+|[^\x00-\x7f]+')
 
 
 def _user_key(app_name: str, user_id: str) -> tuple[str, str]:
@@ -43,6 +44,25 @@ def _extract_words_lower(text: str) -> set[str]:
   return set(word.lower() for word in re.findall(r'\w+', text))
 
 
+def _extract_searchable_words(text: str) -> set[str]:
+  """Extracts the words an event can be matched on, in lowercase.
+
+  The tokens of _extract_words_lower, plus, for a token that mixes scripts,
+  each of its ASCII and non-ASCII runs. Japanese and Chinese are written
+  without spaces, so 私はPythonを使う is a single \\w+ token and a query for
+  Python matches nothing. Splitting on the boundary between scripts makes the
+  embedded Latin word a token of its own, while still keeping a partial word
+  such as thon from matching.
+
+  Args:
+    text: The text of an event.
+  """
+  words = _extract_words_lower(text)
+  for word in [word for word in words if not word.isascii()]:
+    words.update(_SCRIPT_RUN_PATTERN.findall(word))
+  return words
+
+
 class InMemoryMemoryService(BaseMemoryService):
   """An in-memory memory service for prototyping purpose only.
 
@@ -128,7 +148,7 @@ async def search_memory(
         event_text = ' '.join(
             [part.text for part in event.content.parts if part.text]
         )
-        words_in_event = _extract_words_lower(event_text)
+        words_in_event = _extract_searchable_words(event_text)
         if not words_in_event:
           continue
 

tests/unittests/memory/test_in_memory_memory_service.py

@@ -384,8 +384,15 @@ async def test_search_memory_does_not_collide_on_slash_in_identifiers():
         # Mixed: non-Latin substring + Latin token in same event
         ('太郎 works at ABC Corp', '太郎', 1),
         ('太郎 works at ABC Corp', 'ABC', 1),
+        # Latin word inside an unspaced script (no space to tokenize on)
+        ('私はPythonでADKを使っています', 'Python', 1),
+        ('私はPythonでADKを使っています', 'adk', 1),
+        ('我用Python写代码', 'python', 1),
+        ('私はPythonでADKを使っています', '使って', 1),
+        ('私はPythonでADKを使っています', 'Java', 0),
         # Latin partial-word must NOT match (regression guard)
         ('I like to code in Python.', 'thon', 0),
+        ('私はPythonでADKを使っています', 'thon', 0),
     ],
 )
 async def test_search_memory_non_latin(event_text, query, expected_count):