fix

Author: Lin-Nikaido

src/google/adk/models/google_llm.py

@@ -57,6 +57,50 @@
 """
 
 
+_SUPPORTED_FILE_CONTENT_MIME_TYPES = frozenset({
+    # Images
+    "image/png",
+    "image/jpeg",
+    "image/webp",
+    "image/heic",
+    "image/heif",
+
+    # Documents & Text
+    "application/pdf",
+    "text/plain",
+    "text/csv",
+    "text/html",
+    "text/md",
+    "text/x-python",
+    "text/javascript",
+
+    # Audio
+    "audio/wav",
+    "audio/mp3",
+    "audio/aiff",
+    "audio/aac",
+    "audio/ogg",
+    "audio/flac",
+    "audio/mpeg",
+    "audio/mpga",
+    "audio/m4a",
+    "audio/pcm",
+    "audio/webm",
+
+    # Video
+    "video/mp4",
+    "video/mpeg",
+    "video/mov",
+    "video/quicktime",
+    "video/avi",
+    "video/x-flv",
+    "video/mpg",
+    "video/webm",
+    "video/wmv",
+    "video/3gpp"
+})
+
+
 class _ResourceExhaustedError(ClientError):
   """Represents an resources exhausted error received from the Model."""
 
@@ -440,9 +484,20 @@ async def _preprocess_request(self, llm_request: LlmRequest) -> None:
           for part in content.parts:
             # Create copies to avoid mutating the original objects
             if part.inline_data:
+              mime_type = (part.inline_data.mime_type or "").lower()
+              if mime_type not in _SUPPORTED_FILE_CONTENT_MIME_TYPES:
+                identifier = part.inline_data.display_name or "inline_file"
+                part.text = (part.text or "") + f'\n[File reference: "{identifier}"]'
+                part.inline_data = None
               part.inline_data = copy.copy(part.inline_data)
               _remove_display_name_if_present(part.inline_data)
+
             if part.file_data:
+              mime_type = (part.file_data.mime_type or "").lower()
+              identifier = part.file_data.display_name or part.file_data.file_uri
+              if mime_type not in _SUPPORTED_FILE_CONTENT_MIME_TYPES:
+                part.text = (part.text or "") + f'\n[File reference: "{identifier}"]'
+                part.file_data = None
               part.file_data = copy.copy(part.file_data)
               _remove_display_name_if_present(part.file_data)
 

tests/unittests/models/test_google_llm.py

@@ -2139,3 +2139,50 @@ async def __aexit__(self, *args):
       # Verify the final speech_config is still None
       assert config_arg.speech_config is None
       assert isinstance(connection, GeminiLlmConnection)
+
+
+@pytest.mark.asyncio
+async def test_preprocess_request_unsupported_mime_type(gemini_llm):
+    """Verifies that MS Office files are escaped to a text reference."""
+    unsupported_part = types.Part(
+        file_data=types.FileData(
+            mime_type="application/vnd.ms-excel",
+            file_uri="gs://bucket/data.xls",
+            display_name="data.xls"
+        )
+    )
+    req = LlmRequest(
+        model="gemini-2.0-flash",
+        contents=[types.Content(parts=[unsupported_part])]
+    )
+
+    await gemini_llm._preprocess_request(req)
+
+    processed_part = req.contents[0].parts[0]
+    # File_data should be stripped to avoid the 400 error
+    assert processed_part.file_data is None
+    # Text fallback should be present
+    assert '[File reference: "data.xls"]' in processed_part.text
+
+
+@pytest.mark.asyncio
+async def test_preprocess_request_supported_mime_type(gemini_llm):
+    """Verifies that PDF files are passed through without modification."""
+    supported_part = types.Part(
+        file_data=types.FileData(
+            mime_type="application/pdf",
+            file_uri="gs://bucket/doc.pdf",
+            display_name="doc.pdf"
+        )
+    )
+    req = LlmRequest(
+        model="gemini-2.0-flash",
+        contents=[types.Content(parts=[supported_part])]
+    )
+
+    await gemini_llm._preprocess_request(req)
+
+    processed_part = req.contents[0].parts[0]
+    # file_data should still be intact
+    assert processed_part.file_data is not None
+    assert processed_part.file_data.mime_type == "application/pdf"