From aec0ab777bde6f59cd7fc4be153caad247e973de Mon Sep 17 00:00:00 2001
From: abi_jey <git@public.abja.dev>
Date: Tue, 25 Nov 2025 19:18:41 +0000
Subject: [PATCH] feat: add GA protocol as litellm_params for realtime api on
 azure provider

---
 litellm/llms/azure/realtime/handler.py        | 54 ++++++++++++++++--
 .../realtime/test_azure_realtime_handler.py   | 56 +++++++++++++++++++
 2 files changed, 105 insertions(+), 5 deletions(-)

diff --git a/litellm/llms/azure/realtime/handler.py b/litellm/llms/azure/realtime/handler.py
index 8e5581206d..14f772ac07 100644
--- a/litellm/llms/azure/realtime/handler.py
+++ b/litellm/llms/azure/realtime/handler.py
@@ -28,16 +28,60 @@ async def forward_messages(client_ws: Any, backend_ws: Any):
 
 
 class AzureOpenAIRealtime(AzureChatCompletion):
-    def _construct_url(self, api_base: str, model: str, api_version: str) -> str:
+    def _get_realtime_protocol(self) -> str:
+        """Return the configured realtime protocol.
+
+        Supported values (case-insensitive):
+        - "beta"  -> use legacy `/openai/realtime` (current default)
+        - "v1"    -> use `/openai/v1/realtime`
+        - "ga"    -> alias for "v1" (GA path is v1)
+
+        If the parameter is missing or invalid, we fall back to the current
+        behavior for full backwards compatibility.
         """
-        Example output:
+
+        # `litellm_params` is the standard place to configure provider-specific
+        # behavior. We keep this defensive in case the attribute isn't set.
+        params: Any = getattr(self, "litellm_params", None)
+        if not isinstance(params, dict):
+            return "beta"
+
+        value = params.get("realtime_protocol")
+        if not isinstance(value, str):
+            return "beta"
+
+        value_normalized = value.lower()
+        if value_normalized in {"v1", "ga"}:
+            return "v1"
+
+        # Treat anything else (including explicit "beta") as current default
+        return "beta"
+
+    def _construct_url(
+        self,
+        api_base: str,
+        model: str,
+        api_version: str,
+    ) -> str:
+        """Construct the websocket URL for Azure OpenAI realtime.
+
+        Example default output (beta / legacy behavior):
         "wss://my-endpoint-sweden-berri992.openai.azure.com/openai/realtime?api-version=2024-10-01-preview&deployment=gpt-4o-realtime-preview";
 
+        When `realtime_protocol` is set to "v1" or "GA" via `litellm_params`,
+        this switches to `/openai/v1/realtime`.
         """
+
         api_base = api_base.replace("https://", "wss://")
-        return (
-            f"{api_base}/openai/realtime?api-version={api_version}&deployment={model}"
-        )
+
+        protocol = self._get_realtime_protocol()
+        if protocol == "v1":
+            path = "/openai/v1/realtime"
+        else:
+            # default / beta behavior
+            path = "/openai/realtime"
+
+        return f"{api_base}{path}?api-version={api_version}&deployment={model}"
 
     async def async_realtime(
         self,
diff --git a/tests/test_litellm/llms/azure/realtime/test_azure_realtime_handler.py b/tests/test_litellm/llms/azure/realtime/test_azure_realtime_handler.py
index 7bcbe37156..1446bc3df7 100644
--- a/tests/test_litellm/llms/azure/realtime/test_azure_realtime_handler.py
+++ b/tests/test_litellm/llms/azure/realtime/test_azure_realtime_handler.py
@@ -71,3 +71,59 @@ async def test_async_realtime_uses_max_size_parameter():
         mock_realtime_streaming.assert_called_once()
         mock_streaming_instance.bidirectional_forward.assert_awaited_once()
 
+
+@pytest.mark.asyncio
+async def test_construct_url_uses_legacy_realtime_by_default():
+    """By default we should keep using `/openai/realtime` (beta behavior)."""
+
+    from litellm.llms.azure.realtime.handler import AzureOpenAIRealtime
+
+    handler = AzureOpenAIRealtime()
+    api_base = "https://my-endpoint.openai.azure.com"
+    api_version = "2024-10-01-preview"
+    model = "gpt-4o-realtime-preview"
+
+    url = handler._construct_url(api_base=api_base, model=model, api_version=api_version)
+
+    assert url.startswith("wss://my-endpoint.openai.azure.com")
+    assert "/openai/realtime" in url
+    assert "/openai/v1/realtime" not in url
+
+
+@pytest.mark.asyncio
+async def test_construct_url_uses_v1_when_realtime_protocol_v1_or_ga():
+    """Setting `realtime_protocol` to v1/GA should switch to `/openai/v1/realtime`."""
+
+    from litellm.llms.azure.realtime.handler import AzureOpenAIRealtime
+
+    api_base = "https://my-endpoint.openai.azure.com"
+    api_version = "2024-10-01-preview"
+    model = "gpt-4o-realtime-preview"
+
+    # Helper to construct handler URL with a specific realtime_protocol.
+    # We avoid mutating handler attributes directly since type checkers don't
+    # know about `litellm_params` on this class. Instead, we patch the
+    # `_get_realtime_protocol` helper which is what `_construct_url` uses.
+
+    # v1 -> /openai/v1/realtime
+    handler_v1 = AzureOpenAIRealtime()
+    with patch.object(handler_v1, "_get_realtime_protocol", return_value="v1"):
+        url_v1 = handler_v1._construct_url(api_base=api_base, model=model, api_version=api_version)
+    assert "/openai/v1/realtime" in url_v1
+    assert "/openai/realtime" not in url_v1
+
+    # GA (case-insensitive) -> /openai/v1/realtime
+    handler_ga = AzureOpenAIRealtime()
+    with patch.object(handler_ga, "_get_realtime_protocol", return_value="v1"):
+        url_ga = handler_ga._construct_url(api_base=api_base, model=model, api_version=api_version)
+    assert "/openai/v1/realtime" in url_ga
+    assert "/openai/realtime" not in url_ga
+
+    # beta or any other value keeps legacy path
+    handler_beta = AzureOpenAIRealtime()
+    with patch.object(handler_beta, "_get_realtime_protocol", return_value="beta"):
+        url_beta = handler_beta._construct_url(api_base=api_base, model=model, api_version=api_version)
+    assert "/openai/realtime" in url_beta
+    assert "/openai/v1/realtime" not in url_beta
+
+