fix(llms): add is_reasoning_model config override for versioned deployments (#5327)

Co-authored-by: kartik-mem0 <kartik.labhshetwar@mem0.ai>

tests/llms/test_azure_openai.py

@@ -186,6 +186,72 @@ def test_azure_config_accepts_reasoning_effort():
     assert config.model == "o3-mini"
 
 
+def test_is_reasoning_model_override_forces_reasoning_path(mock_openai_client):
+    """Versioned Azure gpt-5.x deployments can opt in via is_reasoning_model=True.
+
+    Regression test for https://github.com/mem0ai/mem0/issues/5296 — the
+    name-based heuristic does not recognize dated deployment names like
+    ``gpt-5.4-nano-2026-03-17``, so the call sent max_tokens and Azure replied
+    400. The explicit override forces the reasoning-model parameter set, which
+    drops max_tokens (and temperature).
+    """
+    config = AzureOpenAIConfig(
+        model="gpt-5.4-nano-2026-03-17",
+        temperature=TEMPERATURE,
+        max_tokens=MAX_TOKENS,
+        is_reasoning_model=True,
+    )
+    llm = AzureOpenAILLM(config)
+    messages = [{"role": "user", "content": "I have oily skin."}]
+
+    mock_response = Mock()
+    mock_response.choices = [Mock(message=Mock(content="ok"))]
+    mock_openai_client.chat.completions.create.return_value = mock_response
+
+    llm.generate_response(messages)
+
+    call_kwargs = mock_openai_client.chat.completions.create.call_args[1]
+    assert "max_tokens" not in call_kwargs
+    assert "temperature" not in call_kwargs
+
+
+def test_is_reasoning_model_override_false_keeps_standard_params(mock_openai_client):
+    """is_reasoning_model=False forces the standard param set even for o-series names."""
+    config = AzureOpenAIConfig(
+        model="o3-mini",
+        temperature=TEMPERATURE,
+        max_tokens=MAX_TOKENS,
+        top_p=TOP_P,
+        is_reasoning_model=False,
+    )
+    llm = AzureOpenAILLM(config)
+    messages = [{"role": "user", "content": "Hello"}]
+
+    mock_response = Mock()
+    mock_response.choices = [Mock(message=Mock(content="ok"))]
+    mock_openai_client.chat.completions.create.return_value = mock_response
+
+    llm.generate_response(messages)
+
+    call_kwargs = mock_openai_client.chat.completions.create.call_args[1]
+    assert call_kwargs["max_tokens"] == MAX_TOKENS
+    assert call_kwargs["temperature"] == TEMPERATURE
+
+
+def test_is_reasoning_model_defaults_to_name_heuristic(mock_openai_client):
+    """When is_reasoning_model is None (default), classification stays name-based."""
+    config = AzureOpenAIConfig(
+        model="gpt-5.4-nano-2026-03-17",
+        temperature=TEMPERATURE,
+        max_tokens=MAX_TOKENS,
+        top_p=TOP_P,
+    )
+    llm = AzureOpenAILLM(config)
+    # Unrecognized versioned name -> heuristic says "not reasoning" (unchanged).
+    assert config.is_reasoning_model is None
+    assert llm._is_reasoning_model("gpt-5.4-nano-2026-03-17") is False
+
+
 @pytest.mark.parametrize(
     "default_headers",
     [None, {"Firstkey": "FirstVal", "SecondKey": "SecondVal"}],

tests/llms/test_openai.py

@@ -328,6 +328,53 @@ def test_is_reasoning_model_classification(mock_openai_client):
     assert llm._is_reasoning_model("gpt-4.1-nano-2025-04-14") is False
 
 
+def test_is_reasoning_model_explicit_override(mock_openai_client):
+    """Explicit is_reasoning_model overrides the name-based heuristic both ways.
+
+    See https://github.com/mem0ai/mem0/issues/5296 — deployments with custom or
+    versioned names need to opt in/out without relying on string matching.
+    None (default) must preserve the existing heuristic.
+    """
+    # Force True: a name the heuristic would reject is now reasoning.
+    config_true = OpenAIConfig(model="gpt-5.4-nano-2026-03-17", is_reasoning_model=True)
+    llm_true = OpenAILLM(config_true)
+    assert llm_true._is_reasoning_model("gpt-5.4-nano-2026-03-17") is True
+
+    # Force False: an o-series name the heuristic would accept is now standard.
+    config_false = OpenAIConfig(model="o3-mini", is_reasoning_model=False)
+    llm_false = OpenAILLM(config_false)
+    assert llm_false._is_reasoning_model("o3-mini") is False
+
+    # None (default) preserves the existing heuristic.
+    config_none = OpenAIConfig(model="gpt-4.1")
+    llm_none = OpenAILLM(config_none)
+    assert config_none.is_reasoning_model is None
+    assert llm_none._is_reasoning_model("o3-mini") is True
+    assert llm_none._is_reasoning_model("gpt-5.4-mini") is False
+
+
+def test_is_reasoning_model_override_generates_correct_params(mock_openai_client):
+    """End-to-end: is_reasoning_model=True drops max_tokens/temperature from the actual API call."""
+    config = OpenAIConfig(
+        model="gpt-5.4-nano-2026-03-17",
+        temperature=0.7,
+        max_tokens=100,
+        is_reasoning_model=True,
+    )
+    llm = OpenAILLM(config)
+    messages = [{"role": "user", "content": "Hello"}]
+
+    mock_response = Mock()
+    mock_response.choices = [Mock(message=Mock(content="ok"))]
+    mock_openai_client.chat.completions.create.return_value = mock_response
+
+    llm.generate_response(messages)
+
+    call_kwargs = mock_openai_client.chat.completions.create.call_args[1]
+    assert "max_tokens" not in call_kwargs
+    assert "temperature" not in call_kwargs
+
+
 def test_callback_with_tools(mock_openai_client):
     mock_callback = Mock()
     config = OpenAIConfig(model="gpt-4.1-nano-2025-04-14", response_callback=mock_callback)