Mishandling of Gemini "RECITATION" Error

[boazk485]

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

When using Gemini, RECITATION errors may occur. Pydantic AI currently does not handle them gracefully.
I do not know how to provide a code snippet to reproduce that specific problem, as it's elusive.
See discussion of the Gemini RECITATION problem here: https://issuetracker.google.com/issues/331677495

Similar issues in pydantic-ai have been rased for other untexpected Finish Reasons:
#509
#631

See traceback:

  File "/my_folder/.venv/lib/python3.13/site-packages/opentelemetry/trace/__init__.py", line 589, in use_span
    yield span
  File "/my_folder/.venv/lib/python3.13/site-packages/pydantic_graph/graph.py", line 246, in iter
    yield GraphRun[StateT, DepsT, RunEndT](
    ...<6 lines>...
    )
  File "/my_folder/.venv/lib/python3.13/site-packages/pydantic_ai/agent/__init__.py", line 674, in iter
    yield agent_run
  File "/my_folder/.venv/lib/python3.13/site-packages/pydantic_ai/agent/abstract.py", line 235, in run
    async for node in agent_run:
    ...<4 lines>...
                await event_stream_handler(_agent_graph.build_run_context(agent_run.ctx), stream)
  File "/my_folder/.venv/lib/python3.13/site-packages/pydantic_ai/run.py", line 150, in __anext__
    next_node = await self._graph_run.__anext__()
                ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/my_folder/.venv/lib/python3.13/site-packages/pydantic_graph/graph.py", line 758, in __anext__
    return await self.next(self._next_node)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/my_folder/.venv/lib/python3.13/site-packages/pydantic_graph/graph.py", line 731, in next
    self._next_node = await node.run(ctx)
                      ^^^^^^^^^^^^^^^^^^^
  File "/my_folder/.venv/lib/python3.13/site-packages/pydantic_ai/_agent_graph.py", line 397, in run
    return await self._make_request(ctx)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/my_folder/.venv/lib/python3.13/site-packages/pydantic_ai/_agent_graph.py", line 439, in _make_request
    model_response = await ctx.deps.model.request(message_history, model_settings, model_request_parameters)
                     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/my_folder/.venv/lib/python3.13/site-packages/pydantic_ai/models/instrumented.py", line 362, in request
    response = await self.wrapped.request(messages, model_settings, model_request_parameters)
               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/my_folder/.venv/lib/python3.13/site-packages/pydantic_ai/models/google.py", line 240, in request
    return self._process_response(response)
           ~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^
  File "/my_folder/.venv/lib/python3.13/site-packages/pydantic_ai/models/google.py", line 463, in _process_response
    raise UnexpectedModelBehavior(
    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
        'Content field missing from Gemini response', str(response)
        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
    )  # pragma: no cover
    ^
pydantic_ai.exceptions.UnexpectedModelBehavior: Content field missing from Gemini response, body:
sdk_http_response=HttpResponse(
  headers=<dict len=11>
) candidates=[Candidate(
  finish_reason=<FinishReason.RECITATION: 'RECITATION'>,
  index=0
)] create_time=None model_version='gemini-2.5-pro' prompt_feedback=None response_id='<my_response_id' usage_metadata=GenerateContentResponseUsageMetadata(
  cache_tokens_details=[
    ModalityTokenCount(
      modality=<MediaModality.TEXT: 'TEXT'>,
      token_count=6800
    ),
  ],
  cached_content_token_count=6800,
  prompt_token_count=9939,
  prompt_tokens_details=[
    ModalityTokenCount(
      modality=<MediaModality.TEXT: 'TEXT'>,
      token_count=9939
    ),
  ],
  thoughts_token_count=468,
  total_token_count=10407
) automatic_function_calling_history=[] parsed=None```

### Example Code

```Python
# This is the basic framework with which I ran my code, I do not know how to produce a fully reproducible code, as the issue appears at random and I'm not able to share the exact Prompt used.

provider = GoogleProvider(api_key=self.api_key)
agent = Agent(
    GoogleModel("gemini-2.5-pro", provider=provider),
    model_settings=ModelSettings(max_tokens=1000000),
    output_type=SomePydanticModel,
    deps_type=AnalysisDeps,
    system_prompt=system_prompt,
    toolsets=mcp_servers,  
)
deps = AnalysisDeps(...)
async with agent:
    agent_result = await agent.run(
	"my_prompt", deps=deps
    )

Python, Pydantic AI & LLM client version

Python 3.13.7 (v3.13.7:bcee1c32211, Aug 14 2025, 19:10:51) [Clang 16.0.0 (clang-1600.0.26.6)] on darwin
pydantic-ai==1.1.0

Using gemini-2.5-pro.

[DouweM]

@boazk485 How would you like/expect responses with RECITATION finish reason to be handled?

[boazk485]

@DouweM As it is a problem that is unlikely to solve itself in a retry of a message, I would expect an Exception to be raised with the appropriate information. As it is an odd behaviour that stops the Agent in the middle and from my experience reoccurs when trying the exact same input, I don't see any other way to manage it.

[DouweM]

@boazk485 OK, so I think we'd just need new branches here:

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

Lines 454 to 456 in bd6cbc3

	if candidate.finish_reason == 'SAFETY':
	raise UnexpectedModelBehavior('Safety settings triggered', str(response))
	else:

And here:

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

Lines 643 to 644 in bd6cbc3

	elif candidate.finish_reason == 'SAFETY': # pragma: no cover
	raise UnexpectedModelBehavior('Safety settings triggered', str(chunk))

Would you be up for filing a PR?

[boazk485]

@DouweM Thanks very much, wanted to be filing the PR and haven't gotten around to it:)