Getting "Unable to submit request because it must include at least one parts field" after gemini model responds with empty text.

[jtattermusch]

Initial Checks

• I confirm that I'm using the latest version of Pydantic AI
• I confirm that I searched for my issue in https://github.com/pydantic/pydantic-ai/issues before opening this issue

Description

Getting pydantic_ai.exceptions.ModelHTTPError: status_code: 400, model_name: gemini-3-pro-preview, body: {'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'}} from gemini3 randomly (at about 10% rate)

What happens:

At some point the gemini generates a response that look as follows. For some reason the model decided to respond with empty text (which is odd, but technically it seems like an ok thing to do - it shouldn't result in the agent crashing).

{
  "content": {
    "role": "model",
    "parts": [
      {
        "text": ""
      }
    ]
  },
  "finishReason": "STOP"
}

Pydantic accepts this response and adds previous response with "parts": [{ "text": ""}] into the next model request, but it seems to leave out the empty text string - the request contains "parts": [] - and that results in getting HTTP 400 from the Gemini backend.

The request indeed contains an empty "parts": "parts": [], so getting back HTTP 400 is to be expected, but the agent fails because of this.

[
  {
    "parts": [
      {
        "text": "\n__REDACTED_PROMPT_HERE__\n"
      }
    ],
    "role": "user"
  },
  {
    "parts": [],
    "role": "model"
  },
  {
    "parts": [
      {
        "text": ""
      }
    ],
    "role": "user"
  }
]

Looks like the bug is either in Pydantic AI or in the underlying libraries (google genai) - but regardless Pydantic AI should have a better way of handling thi

Example Code

Python, Pydantic AI & LLM client version

pydantic AI 1.21.0
Python 3.13.0
google-genai 1.52.0

[pydanticai-bot]

PydanticAI Github Bot Found 1 issues similar to this one:

1. "GoogleModel generates empty text parts that Vertex AI rejects with 400 error #2032" (95% similar)

[jtattermusch]

I also seen multiple occurrences of the same error where the previous response is

{
  "content": {
    "role": "model"
  },
  "finishReason": "MAX_TOKENS"
}

(so it seems that the model has reached the token limit)

But quite unexpectedly, Pydantic AI happily proceeds with constructing new request that contains the "parts: []" - that feels wrong, since after getting MAX_TOKENS, the agent should not continue to send requests.

[jtattermusch]

According to #2032 this should be fixed, but I'm on the lastest PydanticAI version and I'm still getting this error.

[jtattermusch]

I've investigated a bit more and at least some of the instances of the problem look like that Gemini spends all the output tokens (based on max_tokens setting) on thinking - and once it goes over the budget, it returns the response with empty text - and Pydantic AI handles that poorly (it should surface the actual problem that we ran out of output tokens budget).

[jtattermusch]

Possibly related:
googleapis/python-genai#1039
googleapis/python-genai#1011

[DouweM]

@jtattermusch Thanks for filing the issue, looking into it now.

[jtattermusch]

@DouweM I have a hunch that this issue also has something to do with max_tokens vs thinking_budget settings (thinking budget needs to be much smaller than max_tokens) but I'm unable to set thinking_budget via GeminiModelSettings - the max_tokens is being honored, but thinking budget setting is not present when i inspect the outgoing request.

[DouweM]

@jtattermusch Are you using the (deprecated) GeminiModel with GeminiModelSettings, or the newer GoogleModel (which uses google-genai) which needs GoogleModelSettings? GoogleModel it not compatible with GeminiModelSettings, unfortunately.

[jtattermusch]

@jtattermusch Are you using the (deprecated) GeminiModel with GeminiModelSettings, or the newer GoogleModel (which uses google-genai) which needs GoogleModelSettings? GoogleModel it not compatible with GeminiModelSettings, unfortunately.

I'll double check - I wasn't aware of this.

[JoeSham]

Hey, I am getting this error as well, using the GoogleModel and GoogleModelSettings (and newest pydantic-ai version).

Increasing max_tokens helps, but not 100%
(And setting max_tokens to some low value like max_tokens=100, it always results in this error then - just used this to confirm the issue)

        model_settings = GoogleModelSettings(
            max_tokens=65535,
            temperature=0.2 
        )

        model = GoogleModel(
            "gemini-2.5-pro",
            provider=provider
        )