feat: Support inject_schema_instructions_in_native_mode

[xcpky]

closes #3490

Add a field "inject_schema_instructions_in_native_mode" in ModelProfile

The implementation is basically as per @DouweM , with the following improvement:

NativeOutput now can accept a template and json schema instruction is injected regardless of inject_schema_instructions_in_native_mode in that case. For example:

result = await agent.run("how can I solve 8x + 7 = -23", output_type=NativeOutput(MathReasoning, template="{schema}"))

[DouweM]

@xcpky Thanks! Can you please add a test for the model profile behavior + new template field? If we update OutlinesModel/Provider to use the new model profile field, it may be sufficient to add a test to test_outlines.py.

[DouweM]

We need to clarify that this is only used if 1) explicitly provided or 2) the model requires it

[xcpky]

I've added more clarification.

[DouweM]

I have a mild preference for describing traits of the models/providers, rather than our behavior:

Suggested change

	inject_schema_instructions_in_native_mode: bool = False
	json_schema_output_requires_schema_in_instructions: bool = False

[DouweM]

Should we we set this to True on any profiles already?

[DouweM]

we may be able to remove this:

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

Lines 528 to 534 in c1213e4

	def customize_request_parameters(self, model_request_parameters: ModelRequestParameters) -> ModelRequestParameters:
	"""Customize the model request parameters for the model."""
	if model_request_parameters.output_mode in ('auto', 'native'):
	# This way the JSON schema will be included in the instructions.
	return replace(model_request_parameters, output_mode='prompted')
	else:
	return model_request_parameters

if we set the new field here:

pydantic-ai/pydantic_ai_slim/pydantic_ai/providers/outlines.py

Lines 33 to 40 in c1213e4

	def model_profile(self, model_name: str) -> ModelProfile | None:
	"""The model profile for the named model, if available."""
	return ModelProfile(
	supports_tools=False,
	supports_json_schema_output=True,
	supports_json_object_output=True,
	default_structured_output_mode='native',
	)

[xcpky]

I have a mild preference for describing traits of the models/providers, rather than our behavior:

I think native_output_requires_schema_in_instructions is better

we may be able to remove this:

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

Lines 528 to 534 in c1213e4

	def customize_request_parameters(self, model_request_parameters: ModelRequestParameters) -> ModelRequestParameters:
	"""Customize the model request parameters for the model."""
	if model_request_parameters.output_mode in ('auto', 'native'):
	# This way the JSON schema will be included in the instructions.
	return replace(model_request_parameters, output_mode='prompted')
	else:
	return model_request_parameters

if we set the new field here:

pydantic-ai/pydantic_ai_slim/pydantic_ai/providers/outlines.py

Lines 33 to 40 in c1213e4

	def model_profile(self, model_name: str) -> ModelProfile | None:
	"""The model profile for the named model, if available."""
	return ModelProfile(
	supports_tools=False,
	supports_json_schema_output=True,
	supports_json_object_output=True,
	default_structured_output_mode='native',
	)

Done

[DouweM]

@xcpky Thanks! Can you please add a test for the model profile behavior + new template field? If we update OutlinesModel/Provider to use the new model profile field, it may be sufficient to add a test to test_outlines.py.

[xcpky]

a

I've added the test, plz review.

[DouweM]

Users mostly don't deal with profile fields directly, and should be able to assume that we've done our best to make them correct out of the box so that they don't have to worry about them, so I'd rather say:

Suggested change

	If not specified, the default template specified on the model's profile will be used only if `native_output_requires_schema_in_instructions` on the profile is True.
	If explicitly provided, it's used regardless of `native_output_requires_schema_in_instructions`.
	If no template is specified but the model's profile indicates that it requires the schema to be sent as a prompt, the default template specified on the profile will be used.