Impove derivation of the model tag

[adubovik]

Currently the model is derived from request.model field, which is not reliable, since the user may have provided anything in this field.

ai-dial-analytics-realtime/aidial_analytics_realtime/app.py

Lines 105 to 107 in 68ab942

	request_body = json.loads(request_body_str)
	stream = request_body.get("stream", False)
	model = request_body.get("model", deployment)

We should rather look at response.model for chat completion requests because this field is populated by the model (or adapter) itself.
The model/adapter must know better which model has been actually called.

---

Note also that model field doesn't exist in the chat completion request according to Azure OpenAI API:

https://github.com/Azure/azure-rest-api-specs/blob/main/specification/cognitiveservices/data-plane/AzureOpenAI/inference/preview/2024-12-01-preview/inference.yaml#L2188

At the same time, model fields is a required field in the chat completion response:

https://github.com/Azure/azure-rest-api-specs/blob/main/specification/cognitiveservices/data-plane/AzureOpenAI/inference/preview/2024-12-01-preview/inference.yaml#L4152

---

We should probably keep request.model as a fallback from a missing response.model for the special case of the assistant deployment:

assistant service receives the deployment id of the model it's going to call via the request.model field.

Whether it makes any sense to report the model field for non-model deployments (applications and assisstant) in the first place, remains unclear.

[adubovik]

Related to

• fix: add model field to chat/completion responses ai-dial-adapter-bedrock#213
• fix: add model field to chat completion responses ai-dial-adapter-vertexai#174