fix: forward base_url and api_key from LLM to instructor client (#5204)

InternalInstructor._create_instructor_client() was discarding base_url
and api_key when calling instructor.from_provider(), causing requests
to go to api.openai.com instead of the configured endpoint.

Added _get_llm_extra_kwargs() to extract base_url and api_key from the
LLM object and forward them as kwargs to instructor.from_provider().
This fixes OpenAI-compatible endpoints (vLLM, Ollama, etc.) when using
output_pydantic or output_json on tasks.

Co-Authored-By: João <joao@crewai.com>

lib/crewai/src/crewai/utilities/internal_instructor.py

@@ -98,7 +98,28 @@ class InternalInstructor(Generic[T]):
         else:
             provider = "openai"  # Default fallback
 
-        return instructor.from_provider(f"{provider}/{model_string}")
+        extra_kwargs = self._get_llm_extra_kwargs()
+        return instructor.from_provider(f"{provider}/{model_string}", **extra_kwargs)
+
+    def _get_llm_extra_kwargs(self) -> dict[str, Any]:
+        """Extract extra keyword arguments from the LLM to forward to instructor.
+
+        Collects configuration such as base_url and api_key from the LLM object
+        so they can be forwarded to the instructor client. This ensures that
+        OpenAI-compatible endpoints with custom base_url are respected.
+
+        Returns:
+            Dictionary of extra keyword arguments for instructor.from_provider()
+        """
+        if isinstance(self.llm, str) or self.llm is None:
+            return {}
+
+        extra: dict[str, Any] = {}
+        for attr in ("base_url", "api_key"):
+            value = getattr(self.llm, attr, None)
+            if value is not None:
+                extra[attr] = value
+        return extra
 
     def _extract_provider(self) -> str:
         """Extract provider from LLM model name.

lib/crewai/tests/utilities/test_converter.py

@@ -952,3 +952,187 @@ def test_internal_instructor_real_unsupported_provider() -> None:
 
     # Verify it's a configuration error about unsupported provider
     assert "Unsupported provider" in str(exc_info.value) or "unsupported" in str(exc_info.value).lower()
+
+
+def test_internal_instructor_forwards_base_url_to_provider() -> None:
+    """Test that InternalInstructor forwards base_url from LLM to instructor.from_provider.
+
+    This is the core regression test for issue #5204: when using an OpenAI-compatible
+    provider with a custom base_url (e.g. vLLM, Ollama remote), the base_url was
+    silently discarded, causing requests to go to api.openai.com instead.
+    """
+    from crewai.utilities.internal_instructor import InternalInstructor
+
+    mock_llm = Mock()
+    mock_llm.is_litellm = False
+    mock_llm.model = "gpt-4o"
+    mock_llm.provider = "openai"
+    mock_llm.base_url = "http://localhost:8000/v1"
+    mock_llm.api_key = "test-key-123"
+
+    mock_client = Mock()
+
+    with patch("instructor.from_provider", return_value=mock_client) as mock_from_provider:
+        InternalInstructor(
+            content="Test content",
+            model=SimpleModel,
+            llm=mock_llm,
+        )
+
+        mock_from_provider.assert_called_once_with(
+            "openai/gpt-4o",
+            base_url="http://localhost:8000/v1",
+            api_key="test-key-123",
+        )
+
+
+def test_internal_instructor_forwards_base_url_without_api_key() -> None:
+    """Test that base_url is forwarded even when api_key is not set."""
+    from crewai.utilities.internal_instructor import InternalInstructor
+
+    mock_llm = Mock()
+    mock_llm.is_litellm = False
+    mock_llm.model = "gpt-4o"
+    mock_llm.provider = "openai"
+    mock_llm.base_url = "http://my-vllm-server:8000/v1"
+    mock_llm.api_key = None
+
+    mock_client = Mock()
+
+    with patch("instructor.from_provider", return_value=mock_client) as mock_from_provider:
+        InternalInstructor(
+            content="Test content",
+            model=SimpleModel,
+            llm=mock_llm,
+        )
+
+        mock_from_provider.assert_called_once_with(
+            "openai/gpt-4o",
+            base_url="http://my-vllm-server:8000/v1",
+        )
+
+
+def test_internal_instructor_no_extra_kwargs_when_no_base_url() -> None:
+    """Test that no extra kwargs are passed when LLM has no base_url or api_key."""
+    from crewai.utilities.internal_instructor import InternalInstructor
+
+    mock_llm = Mock()
+    mock_llm.is_litellm = False
+    mock_llm.model = "gpt-4o"
+    mock_llm.provider = "openai"
+    mock_llm.base_url = None
+    mock_llm.api_key = None
+
+    mock_client = Mock()
+
+    with patch("instructor.from_provider", return_value=mock_client) as mock_from_provider:
+        InternalInstructor(
+            content="Test content",
+            model=SimpleModel,
+            llm=mock_llm,
+        )
+
+        mock_from_provider.assert_called_once_with("openai/gpt-4o")
+
+
+def test_internal_instructor_forwards_base_url_for_ollama_provider() -> None:
+    """Test that base_url is forwarded for Ollama provider with custom endpoint."""
+    from crewai.utilities.internal_instructor import InternalInstructor
+
+    mock_llm = Mock()
+    mock_llm.is_litellm = False
+    mock_llm.model = "llama3.1"
+    mock_llm.provider = "ollama"
+    mock_llm.base_url = "http://remote-ollama:11434/v1"
+    mock_llm.api_key = None
+
+    mock_client = Mock()
+
+    with patch("instructor.from_provider", return_value=mock_client) as mock_from_provider:
+        InternalInstructor(
+            content="Test content",
+            model=SimpleModel,
+            llm=mock_llm,
+        )
+
+        mock_from_provider.assert_called_once_with(
+            "ollama/llama3.1",
+            base_url="http://remote-ollama:11434/v1",
+        )
+
+
+def test_internal_instructor_string_llm_no_extra_kwargs() -> None:
+    """Test that string LLM does not produce extra kwargs."""
+    from crewai.utilities.internal_instructor import InternalInstructor
+
+    mock_client = Mock()
+
+    with patch("instructor.from_provider", return_value=mock_client) as mock_from_provider:
+        InternalInstructor(
+            content="Test content",
+            model=SimpleModel,
+            llm="openai/gpt-4o",
+        )
+
+        # String LLM "openai/gpt-4o": _extract_provider returns "openai",
+        # then from_provider is called with "openai/openai/gpt-4o"
+        mock_from_provider.assert_called_once_with("openai/openai/gpt-4o")
+
+
+def test_internal_instructor_get_llm_extra_kwargs() -> None:
+    """Test _get_llm_extra_kwargs extracts the right attributes from the LLM."""
+    from crewai.utilities.internal_instructor import InternalInstructor
+
+    mock_client = Mock()
+
+    with patch("instructor.from_provider", return_value=mock_client):
+        # LLM with both base_url and api_key
+        mock_llm = Mock()
+        mock_llm.is_litellm = False
+        mock_llm.model = "gpt-4o"
+        mock_llm.provider = "openai"
+        mock_llm.base_url = "http://localhost:8000/v1"
+        mock_llm.api_key = "sk-test"
+
+        inst = InternalInstructor(
+            content="Test", model=SimpleModel, llm=mock_llm
+        )
+        assert inst._get_llm_extra_kwargs() == {
+            "base_url": "http://localhost:8000/v1",
+            "api_key": "sk-test",
+        }
+
+        # LLM with only base_url
+        mock_llm2 = Mock()
+        mock_llm2.is_litellm = False
+        mock_llm2.model = "gpt-4o"
+        mock_llm2.provider = "openai"
+        mock_llm2.base_url = "http://localhost:8000/v1"
+        mock_llm2.api_key = None
+
+        inst2 = InternalInstructor(
+            content="Test", model=SimpleModel, llm=mock_llm2
+        )
+        assert inst2._get_llm_extra_kwargs() == {
+            "base_url": "http://localhost:8000/v1",
+        }
+
+        # LLM with neither
+        mock_llm3 = Mock()
+        mock_llm3.is_litellm = False
+        mock_llm3.model = "gpt-4o"
+        mock_llm3.provider = "openai"
+        mock_llm3.base_url = None
+        mock_llm3.api_key = None
+
+        inst3 = InternalInstructor(
+            content="Test", model=SimpleModel, llm=mock_llm3
+        )
+        assert inst3._get_llm_extra_kwargs() == {}
+
+    # String LLM
+    with patch("instructor.from_provider", return_value=mock_client):
+        inst4 = InternalInstructor(
+            content="Test", model=SimpleModel, llm="openai/gpt-4o"
+        )
+        assert inst4._get_llm_extra_kwargs() == {}