Handle function calls without text #2557

ethanabrooks · 2025-08-14T18:36:34Z

This addresses the following error:

Traceback (most recent call last):
  File ".venv/lib/python3.13/site-packages/opentelemetry/trace/__init__.py", line 589, in use_span
    yield span
  File ".venv/lib/python3.13/site-packages/opentelemetry/sdk/trace/__init__.py", line 1105, in start_as_current_span
    yield span
  File ".venv/lib/python3.13/site-packages/temporalio/contrib/opentelemetry.py", line 295, in execute_activity
    return await super().execute_activity(input)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File ".venv/lib/python3.13/site-packages/temporalio/worker/_interceptor.py", line 123, in execute_activity
    return await self.next.execute_activity(input)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File ".venv/lib/python3.13/site-packages/temporalio/worker/_activity.py", line 784, in execute_activity
    return await input.fn(*input.args)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "app/temporal/heartbeater.py", line 26, in wrapper
    return await fn(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^
  File "app/temporal/activities.py", line 160, in wrapper
    raise ApplicationError(
    ...<3 lines>...
    ) from e
temporalio.exceptions.ApplicationError: ClientError: ClientError: 400 INVALID_ARGUMENT. {'error': {'code': 400, 'message': 'Unable to submit request because it must include at least one parts field, which describes the prompt input. Learn more: https://cloud.google.com/vertex-ai/generative-ai/docs/model-reference/gemini', 'status': 'INVALID_ARGUMENT'}}

DouweM

@ethanabrooks Can you show me the error you'd get without this? It's odd that Google could generate a model response with only tool calls and no text, that it then wouldn't accept when sent back on a subsequent request.

DouweM · 2025-08-14T19:15:28Z

pydantic_ai_slim/pydantic_ai/models/google.py

+                        # If we only have function calls without text, add minimal text to satisfy Google API
+                        if has_function_calls and not has_text_parts:
+                            # Add a minimal text part to make the conversation valid for Google API
+                            parts.append({'text': 'I have completed the function calls above.'})


Does this need text or could be empty?

Could we do this inside _content_model_response based on the ModelResponse's parts rather than parsing the returned ContentDict? We already iterate over all items there, so we can keep track of what we've seen.

Updated the description to include the full error. Yes your suggestion sounds like a good idea.

Does this need text or could be empty?

Gave that a shot, but still got:

google.genai.errors.ClientError: 400 INVALID_ARGUMENT. {'error': {'code': 400, 'message': 'Unable to submit request because it must have a text parameter. Add a text parameter and try again. Learn more: https://cloud.google.com/vertex-ai/generative-ai/docs/model-reference/gemini', 'status': 'INVALID_ARGUMENT'}}

Could do something more minimal like 'Function call complete'.

Could we do this inside _content_model_response based on the ModelResponse's parts rather than parsing the returned ContentDict? We already iterate over all items there, so we can keep track of what we've seen.

Just pushed a new commit that does this. Might need to iterate on the tests.

This is only a Vertex AI issue right? Could we gate it by checking if self.system so that we don't send unnecessary tokens to the non-Vertex Gemini API?

Does it accept ' '? :) I wonder how strict their check is, and I'd rather not waste any tokens

DouweM · 2025-08-14T19:44:18Z

Unable to submit request because it must include at least one parts field, which describes the prompt input. Learn more: https://cloud.google.com/vertex-ai/generative-ai/docs/model-reference/gemini

@ethanabrooks The error is odd, because there is in fact a parts field with the function_call, right?

We already have similar logic for the user role message, but that's only for when parts is actually empty:

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

Lines 455 to 457 in eeeb32d

    
           # Google GenAI requires at least one part in the message. 
        
           if not message_parts: 
        
               message_parts = [{'text': ''}]

This looks like a bug on Google's side, but I agree we should work around it.

DouweM · 2025-08-14T19:57:33Z

@ethanabrooks By the way nice to see you using Pydantic AI with Temporal! Is that already using our brand-new integration? I'd love feedback on Slack!

ethanabrooks · 2025-08-14T21:33:45Z

@ethanabrooks By the way nice to see you using Pydantic AI with Temporal! Is that already using our brand-new integration? I'd love feedback on Slack!

We just had a big planning meeting where talked about upgrading to v0.7 and using the new temporal features. Will keep you posted!

ethanabrooks · 2025-08-14T21:48:34Z

Might need some help with the test-coverage issue. It seems to be referencing lines that I didn't change... Also I get the test-coverage error on main as well.

DouweM · 2025-08-14T22:29:34Z

Might need some help with the test-coverage issue. It seems to be referencing lines that I didn't change... Also I get the test-coverage error on main as well.

How strange, line 584 doesn't even have a pragma: no cover :/ I can look into that once the code here is ready

DouweM · 2025-08-16T00:39:05Z

@ethanabrooks Just a heads-up that I'll be out this coming week and will be back the 25th. Assuming this is not urgent I'll review it then. If it is, please ping Kludex! Appreciate the patience :)

ethanabrooks · 2025-08-18T13:16:08Z

@ethanabrooks Just a heads-up that I'll be out this coming week and will be back the 25th. Assuming this is not urgent I'll review it then. If it is, please ping Kludex! Appreciate the patience :)

Not a problem. I'll check again next week.

# Conflicts: # tests/models/test_google.py

DouweM · 2025-09-05T16:45:16Z

pydantic_ai_slim/pydantic_ai/models/google.py

@@ -472,7 +472,10 @@ async def _map_messages(self, messages: list[ModelMessage]) -> tuple[ContentDict
                    message_parts = [{'text': ''}]
                contents.append({'role': 'user', 'parts': message_parts})
            elif isinstance(m, ModelResponse):
-                contents.append(_content_model_response(m))
+                model_content = _content_model_response(m)
+                # Skip model responses with empty parts (e.g., thinking-only responses)


@ethanabrooks Does this mean this would be a non-issue once we start sending back thinking parts? (#2594) I mean to work on that next week

DouweM · 2025-09-05T16:49:22Z

tests/models/test_google.py

@@ -1828,3 +1834,184 @@ class CityLocation(BaseModel):
    agent = Agent(m, output_type=PromptedOutput(CityLocation), builtin_tools=[UrlContextTool()])
    result = await agent.run('What is the largest city in Mexico?')
    assert result.output == snapshot(CityLocation(city='Mexico City', country='Mexico'))
+
+
+@pytest.mark.parametrize(


It would be great if we had a minimal reproducible example that triggers the error against the actual API, so we can verify what parts of the fix are necessary (e.g. is sending back thinking parts enough? does a one-space string work? is it only Vertex or also GLA?). Right now, we're effectively duplicating the fix in tests, rather than verifying that a previously-failing example now succeeds.

ethanabrooks mentioned this pull request Aug 14, 2025

Handle function calls without text (#3) #2551

Closed

ethanabrooks changed the title ~~add handling for thinking-only requests (currently causes UnexpectedModelBehavior)~~ Handle function calls without text (#3) Aug 14, 2025

ethanabrooks changed the title ~~Handle function calls without text (#3)~~ Handle function calls without text Aug 14, 2025

DouweM requested changes Aug 14, 2025

View reviewed changes

DouweM self-assigned this Aug 14, 2025

DouweM added the awaiting author revision label Aug 14, 2025

Handle function calls without text (pydantic#3)

e436234

ethanabrooks force-pushed the eb/handle-textless-function-calls branch from 9facb04 to e456ad3 Compare August 14, 2025 21:49

move logic inside _content_model_response

8f38d24

ethanabrooks force-pushed the eb/handle-textless-function-calls branch from e456ad3 to 8f38d24 Compare August 14, 2025 21:55

Merge branch 'main' into pr/ethanabrooks/2557

66ec4a1

# Conflicts: # tests/models/test_google.py

DouweM requested changes Sep 5, 2025

View reviewed changes

Handle function calls without text #2557

Are you sure you want to change the base?

Handle function calls without text #2557

Conversation

ethanabrooks commented Aug 14, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

DouweM left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

DouweM commented Aug 14, 2025

Uh oh!

DouweM commented Aug 14, 2025

Uh oh!

ethanabrooks commented Aug 14, 2025

Uh oh!

ethanabrooks commented Aug 14, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

DouweM commented Aug 14, 2025

Uh oh!

DouweM commented Aug 16, 2025

Uh oh!

ethanabrooks commented Aug 18, 2025

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

ethanabrooks commented Aug 14, 2025 •

edited

Loading

ethanabrooks commented Aug 14, 2025 •

edited

Loading