Check for existing issues

• I have searched the existing issues and checked that my issue is not a duplicate.

What happened?

Summary

When using LiteLLM with the azure_ai provider and passing response_format containing a JSON schema (structured output), requests would fail or return unstructured responses for the vast majority of Azure AI Foundry models.

Problem Description

LiteLLM supports structured output via the response_format parameter, which accepts a JSON schema:

import litellm

response = litellm.completion(
    model="azure_ai/gpt-4o",
    messages=[{"role": "user", "content": "Return a JSON object with name and age."}],
    response_format={
        "type": "json_schema",
        "json_schema": {
            "name": "person",
            "schema": {
                "type": "object",
                "properties": {
                    "name": {"type": "string"},
                    "age": {"type": "integer"}
                },
                "required": ["name", "age"]
            }
        }
    }
)

For the azure_ai provider (Azure AI Foundry endpoints, *.services.ai.azure.com), this would either:

• Fail with an API error — the model endpoint rejects the raw response_format payload because it does not natively support structured output.
• Return unstructured output — the schema was silently ignored.

The azure provider (Azure OpenAI) worked correctly. Only azure_ai (Azure AI Foundry) was affected.

Root Cause

AzureAIStudioConfig (in litellm/llms/azure_ai/chat/transformation.py) inherits map_openai_params() from OpenAIConfig → OpenAIGPTConfig._map_openai_params(), which is a pure passthrough — it copies response_format into the outgoing request body as-is with no structured output logic whatsoever.

By contrast, AzureOpenAIConfig (in litellm/llms/azure/chat/gpt_transformation.py) already had proper handling:

1. Check if the model supports native structured output via supports_response_schema().
2. If supported → pass response_format through natively.
3. If not supported → convert the JSON schema to a forced tool call via BaseConfig._add_response_format_to_tools().

AzureAIStudioConfig had none of this logic, so every request with a JSON schema was sent raw to the endpoint.

Additionally, 43 out of 75 azure_ai/* models in model_prices_and_context_window.json were missing the supports_response_schema flag entirely, meaning the helper supports_response_schema() had no data to make the right decision for those models.

Impact

All Azure AI Foundry models without native structured output support (Llama, Phi, Mistral, DeepSeek, JAIS, Jamba, Kimi, MAI-DS, Ministral, and others) would fail or misbehave when response_format with a JSON schema was used.

Fix

1. litellm/llms/azure_ai/chat/transformation.py

Added a map_openai_params() override to AzureAIStudioConfig with the following logic:

• JSON schema + supported model → pass response_format through natively.
• JSON schema + unsupported model → convert to a forced tool call via _add_response_format_to_tools() (same fallback used by Azure OpenAI).
• Plain {"type": "json_object"} (no schema) → pass through directly.
• All other params → copy through unchanged.

2. model_prices_and_context_window.json

Added supports_response_schema flags for 43 previously-untagged azure_ai chat models. All were set to false since none of them (Llama, Phi, Mistral, DeepSeek, etc.) support native structured output on Foundry endpoints.

3. tests/test_litellm/llms/azure_ai/chat/test_azure_ai_transformation.py

Added 6 unit tests covering the new behavior:

Files Changed

Steps to Reproduce

1. Create connection between Azure AI Foundry and LiteLLM
2. Configure your application that is using structured output to retrieve data from LiteLLM
3. No structured output will be retrieved

Relevant log output

What part of LiteLLM is this about?

Proxy

What LiteLLM version are you on ?

v1.81.16

Twitter / LinkedIn details

No response