Ignore empty text parts in GoogleModel

[naveen-corpusant]

We observe when using gemini 2.5 pro, a significant number of errors with the message Please return text or call a tool. On debugging, it seems gemini 2.5 will sometimes return an empty text part after a tool call + tool return node, rather than a message with no parts.

This updates the check to handle this case as well. There's a few threads like this one: https://discuss.ai.google.dev/t/gemini-2-5-pro-with-empty-response-text/81175 that discuss why gemini is returning an empty text response, but from my testing, this happens even when there is a successful tool call and return

Example of empty message

ModelResponse(parts=[TextPart(content='')], usage=RequestUsage(input_tokens=25980, cache_read_tokens=23733, details={'cached_content_tokens': 23733, 'text_prompt_tokens': 25980, 'text_cache_tokens': 23733}), model_name='gemini-2.5-pro', timestamp=datetime.datetime(2025, 11, 6, 20, 59, 48, 480954, tzinfo=TzInfo(UTC)), provider_name='google-vertex', provider_details={'finish_reason': 'STOP'}, provider_response_id='XXX', finish_reason='stop')

[naveen-corpusant]

Here's an example of how ctx.state.message_history looks when we hit this case (last few model response / requests). So in this case, before we even get back the tool return, there's a TextPart with the content, and then we get an empty TextPart afterwards.

ModelResponse(parts=[ThinkingPart(content='**Processing User Query**\n\nI\'ve successfully identified the user\'s request for ...This is the logical next step.\n\n\n', signature='XX', provider_name='google-vertex'), TextPart(content='Sure, going to do this task'), ToolCallPart(tool_name='my_custom_tool', args={xx}, tool_call_id='xx')], usage=RequestUsage(input_tokens=25289, cache_read_tokens=23597, output_tokens=210, details={'cached_content_tokens': 23597, 'thoughts_tokens': 126, 'text_prompt_tokens': 25289, 'text_cache_tokens': 23597, 'text_candidates_tokens': 84}), model_name='gemini-2.5-pro', timestamp=datetime.datetime(2025, 11, 6, 23, 7, 4, 709356, tzinfo=TzInfo(UTC)), provider_name='google-vertex', provider_details={'finish_reason': 'STOP'}, provider_response_id='xx', finish_reason='stop')--

ModelRequest(parts=[ToolReturnPart(tool_name='my_custom_tool', content=None, tool_call_id='XX', timestamp=datetime.datetime(2025, 11, 6, 23, 7, 8, 108794, tzinfo=datetime.timezone.utc))])


-------

ModelResponse(parts=[TextPart(content='')], usage=RequestUsage(input_tokens=25771, cache_read_tokens=23648, details={'cached_content_tokens': 23648, 'text_prompt_tokens': 25771, 'text_cache_tokens': 23648}), model_name='gemini-2.5-pro', timestamp=datetime.datetime(2025, 11, 6, 23, 7, 8, 246945, tzinfo=TzInfo(UTC)), provider_name='google-vertex', provider_details={'finish_reason': 'STOP'}, provider_response_id='XX', finish_reason='stop')

[DouweM]

@naveen-corpusant What do you think about instead not creating TextParts at all when text is empty?

That'd mean changing if part.text is not None: to just if part.text: here:

pydantic-ai/pydantic_ai_slim/pydantic_ai/models/google.py

Lines 825 to 829 in 4b46b1e

	elif part.text is not None:
	if part.thought:
	item = ThinkingPart(content=part.text)
	else:
	item = TextPart(content=part.text)

And here:

pydantic-ai/pydantic_ai_slim/pydantic_ai/models/google.py

Lines 680 to 686 in 4b46b1e

	if part.text is not None:
	if part.thought:
	yield self._parts_manager.handle_thinking_delta(vendor_part_id='thinking', content=part.text)
	else:
	maybe_event = self._parts_manager.handle_text_delta(vendor_part_id='content', content=part.text)
	if maybe_event is not None: # pragma: no branch
	yield maybe_event

[naveen-corpusant]

@DouweM good suggestion- was thinking about this as I was planning to monkey patch for current usage, and realized it'd be cleaner to do this at the google level rather than apply it to all models at the _agent_graph. Let's go with that, I'll update the PR now.

[naveen-corpusant]

updated! thanks for the review