fix: OpenAI base_url default and reasoning effort model option.

mellea/backends/model_ids.py

@@ -17,6 +17,7 @@ class ModelIdentifier:
     ollama_name: str | None = None
     watsonx_name: str | None = None
     mlx_name: str | None = None
+    openai_name: str | None = None
 
     hf_tokenizer_name: str | None = None  # if None, is the same as hf_model_name
 
@@ -134,9 +135,9 @@ class ModelIdentifier:
 
 QWEN3_14B = ModelIdentifier(hf_model_name="Qwen/Qwen3-14B", ollama_name="qwen3:14b")
 
-######################
-#### OpenAI models ###
-######################
+###########################
+#### OpenAI open models ###
+###########################
 
 OPENAI_GPT_OSS_20B = ModelIdentifier(
     hf_model_name="openai/gpt-oss-20b", ollama_name="gpt-oss:20b"
@@ -145,6 +146,12 @@ class ModelIdentifier:
     hf_model_name="openai/gpt-oss-120b", ollama_name="gpt-oss:120b"
 )
 
+###########################
+#### OpenAI prop models ###
+###########################
+
+OPENAI_GPT_5_1 = ModelIdentifier(openai_name="gpt-5.1")
+
 #####################
 #### Misc models ####
 #####################

mellea/backends/openai.py

@@ -72,7 +72,7 @@ class OpenAIBackend(FormatterBackend, AdapterMixin):
 
     def __init__(
         self,
-        model_id: str | ModelIdentifier = model_ids.IBM_GRANITE_4_MICRO_3B,
+        model_id: str | ModelIdentifier = model_ids.OPENAI_GPT_5_1,
         formatter: Formatter | None = None,
         base_url: str | None = None,
         model_options: dict | None = None,
@@ -152,16 +152,21 @@ def __init__(
                 )
                 self._hf_model_id = model_id.hf_model_name
 
-        if base_url is None:
-            self._base_url = "http://localhost:11434/v1"  # ollama
-        else:
-            self._base_url = base_url
         if api_key is None:
+            FancyLogger.get_logger().warning(
+                "You are using an OpenAI backend with no api_key. Because no API key was provided, mellea assumes you intend to use the openai-compatible interface to your local ollama instance. If you intend to use OpenAI's platform you must specify your API key when instantiating your Mellea session/backend object."
+            )
+            self._base_url: str | None = "http://localhost:11434/v1"  # ollama
             self._api_key = "ollama"
         else:
+            self._base_url = base_url
             self._api_key = api_key
 
-        self._server_type = _server_type(self._base_url)
+        self._server_type: _ServerType = (
+            _server_type(self._base_url)
+            if self._base_url is not None
+            else _ServerType.OPENAI
+        )  # type: ignore
 
         self._openai_client_kwargs = self.filter_openai_client_kwargs(**kwargs)
 
@@ -598,14 +603,38 @@ async def _generate_from_chat_context_standard(
 
         extra_params: dict[str, Any] = {}
         if _format is not None:
-            extra_params["response_format"] = {
-                "type": "json_schema",
-                "json_schema": {
-                    "name": _format.__name__,
-                    "schema": _format.model_json_schema(),
-                    "strict": True,
-                },
-            }
+            if self._server_type == _ServerType.OPENAI:
+                # The OpenAI platform requires that additionalProperties=False on all response_format schemas.
+                # However, not all schemas generates by Mellea include additionalProperties.
+                # GenerativeSlot, in particular, does not add this property.
+                # The easiest way to address this disparity between OpenAI and other inference providers is to
+                # monkey-patch the response format exactly when we are actually using the OpenAI server.
+                #
+                # This only addresses the additionalProperties=False constraint.
+                # Other constraints we should be checking/patching are described here:
+                # https://platform.openai.com/docs/guides/structured-outputs?api-mode=chat
+                monkey_patched_response_schema = _format.model_json_schema()
+                monkey_patched_response_schema["additionalProperties"] = False
+                extra_params["response_format"] = {
+                    "type": "json_schema",
+                    "json_schema": {
+                        "name": _format.__name__,
+                        "schema": monkey_patched_response_schema,
+                        "strict": True,
+                    },
+                }
+            else:
+                FancyLogger().get_logger().warning(
+                    "Mellea assumes you are NOT using the OpenAI platform, and that other model providers have less strict requirements on support JSON schemas passed into `format=`. If you encounter a server-side error following this message, then you found an exception to this assumption. Please open an issue at github.com/generative_computing/mellea with this stack trace and your inference engine / model provider."
+                )
+                extra_params["response_format"] = {
+                    "type": "json_schema",
+                    "json_schema": {
+                        "name": _format.__name__,
+                        "schema": _format.model_json_schema(),
+                        "strict": True,
+                    },
+                }
 
         # Append tool call information if applicable.
         tools: dict[str, Callable] = dict()
@@ -631,15 +660,20 @@ async def _generate_from_chat_context_standard(
         formatted_tools = convert_tools_to_json(tools)
         use_tools = len(formatted_tools) > 0
 
+        # Build optional reasoning parameters
+        reasoning_params = {}
+        if thinking is not None:
+            reasoning_params["reasoning_effort"] = thinking
+
         chat_response: Coroutine[
             Any, Any, ChatCompletion | openai.AsyncStream[ChatCompletionChunk]
         ] = self._async_client.chat.completions.create(
             model=self._hf_model_id,
             messages=conversation,  # type: ignore
-            reasoning_effort=thinking,  # type: ignore
             tools=formatted_tools if use_tools else None,  # type: ignore
             # parallel_tool_calls=False, # We only support calling one tool per turn. But we do the choosing on our side so we leave this False.
             **extra_params,
+            **reasoning_params,  # type: ignore
             **self._make_backend_specific_and_remove(
                 model_opts, is_chat_context=ctx.is_chat_context
             ),
@@ -976,6 +1010,9 @@ def apply_chat_template(self, chat: list[dict[str, str]]):
         from transformers import AutoTokenizer
 
         if not hasattr(self, "_tokenizer"):
+            assert self._base_url, (
+                "The OpenAI Platform does not support adapters. You must specify a _base_url when using adapters."
+            )
             match _server_type(self._base_url):
                 case _ServerType.LOCALHOST:
                     self._tokenizer: "PreTrainedTokenizer" = (  # noqa: UP037

test/backends/test_openai_ollama.py

@@ -216,6 +216,46 @@ async def get_client_async():
     assert len(backend._client_cache.cache.values()) == 2
 
 
+async def test_reasoning_effort_conditional_passing(backend):
+    """Test that reasoning_effort is only passed to API when not None."""
+    from unittest.mock import AsyncMock, MagicMock, patch
+
+    ctx = ChatContext()
+    ctx = ctx.add(CBlock(value="Test"))
+
+    mock_response = MagicMock()
+    mock_response.choices = [MagicMock()]
+    mock_response.choices[0].message = MagicMock()
+    mock_response.choices[0].message.content = "Response"
+    mock_response.choices[0].message.role = "assistant"
+
+    # Test 1: reasoning_effort should NOT be passed when not specified
+    with patch.object(
+        backend._async_client.chat.completions, "create", new_callable=AsyncMock
+    ) as mock_create:
+        mock_create.return_value = mock_response
+        await backend.generate_from_chat_context(
+            CBlock(value="Hi"), ctx, model_options={}
+        )
+        call_kwargs = mock_create.call_args.kwargs
+        assert "reasoning_effort" not in call_kwargs, (
+            "reasoning_effort should not be passed when not specified"
+        )
+
+    # Test 2: reasoning_effort SHOULD be passed when specified
+    with patch.object(
+        backend._async_client.chat.completions, "create", new_callable=AsyncMock
+    ) as mock_create:
+        mock_create.return_value = mock_response
+        await backend.generate_from_chat_context(
+            CBlock(value="Hi"), ctx, model_options={ModelOption.THINKING: "medium"}
+        )
+        call_kwargs = mock_create.call_args.kwargs
+        assert call_kwargs.get("reasoning_effort") == "medium", (
+            "reasoning_effort should be passed with correct value when specified"
+        )
+
+
 if __name__ == "__main__":
     import pytest