Unable to get deterministic structured JSON output for dynamic Azure ARM templates using OpenAIChatCompletionClient and response_format

[Shaik-Vaheed-UP]

I am building an application that updates Azure ARM templates based on Azure Policy definitions. The challenge is that I need the LLM to return a deterministic structured JSON output every time, but the ARM template structure changes depending on the Azure resource type.
Although most ARM resources follow a common high-level structure such as id, name, type, kind, location, tags, properties, and sku, the internal fields inside objects like properties and sku vary by resource type. Because of that, I cannot define one fixed schema that fully covers all possible ARM resources.

class TobeAzureResourceResponse(BaseModel):
model_config = {"extra": "forbid"}
id: str = Field(description="The unique identifier of the Azure resource.")
name: str = Field(description="The name of the Azure resource.")
type: str = Field(description="The type of the Azure resource, e.g., Microsoft.Compute/virtualMachines.")
kind: str = Field(None, description="The kind of the Azure resource.")
location: str = Field(None, description="The location of the Azure resource.")
sku: str = Field(None, description="The SKU of the Azure resource.")
tags: str = Field(None, description="The tags associated with the Azure resource.")
properties: str = Field(None, description="The properties of the Azure resource.")
Issue
The main issues I am facing are:

Even after passing response_format, I am still receiving the model output as a string instead of a strongly structured JSON object.
Since the ARM template schema changes with resource type, I do not know how to define a schema that is both:

strict enough to enforce deterministic response format
flexible enough to support dynamic ARM resource structures

I am not sure whether response_format alone is the correct approach for this use case, or whether I should instead use:

a Pydantic schema
a JSON schema with strict mode
function/tool calling
or another pattern for dynamic structured outputs

I also want to understand the best practice for scenarios where only a partial high-level schema is known, but nested fields are dynamic.