Merge branch 'master' into html-simple-split

libs/partners/openai/Makefile

@@ -21,7 +21,7 @@ test tests:
 	TIKTOKEN_CACHE_DIR=tiktoken_cache poetry run pytest --disable-socket --allow-unix-socket $(TEST_FILE)
 
 integration_test integration_tests:
-	poetry run pytest $(TEST_FILE)
+	poetry run pytest -n auto $(TEST_FILE)
 
 test_watch:
 	poetry run ptw --snapshot-update --now . -- -vv $(TEST_FILE)

libs/partners/openai/langchain_openai/chat_models/base.py

@@ -1354,19 +1354,33 @@ def with_structured_output(
             )
         is_pydantic_schema = _is_pydantic_class(schema)
 
-        # Check for Pydantic BaseModel V1
-        if (
-            method == "json_schema"
-            and is_pydantic_schema
-            and issubclass(schema, BaseModelV1)  # type: ignore[arg-type]
-        ):
-            warnings.warn(
-                "Received a Pydantic BaseModel V1 schema. This is not supported by "
-                'method="json_schema". Please use method="function_calling" '
-                "or specify schema via JSON Schema or Pydantic V2 BaseModel. "
-                'Overriding to method="function_calling".'
-            )
-            method = "function_calling"
+        if method == "json_schema":
+            # Check for Pydantic BaseModel V1
+            if (
+                is_pydantic_schema and issubclass(schema, BaseModelV1)  # type: ignore[arg-type]
+            ):
+                warnings.warn(
+                    "Received a Pydantic BaseModel V1 schema. This is not supported by "
+                    'method="json_schema". Please use method="function_calling" '
+                    "or specify schema via JSON Schema or Pydantic V2 BaseModel. "
+                    'Overriding to method="function_calling".'
+                )
+                method = "function_calling"
+            # Check for incompatible model
+            if self.model_name and (
+                self.model_name.startswith("gpt-3")
+                or self.model_name.startswith("gpt-4-")
+                or self.model_name == "gpt-4"
+            ):
+                warnings.warn(
+                    f"Cannot use method='json_schema' with model {self.model_name} "
+                    f"since it doesn't support OpenAI's Structured Output API. You can "
+                    f"see supported models here: "
+                    f"https://platform.openai.com/docs/guides/structured-outputs#supported-models. "  # noqa: E501
+                    "To fix this warning, set `method='function_calling'. "
+                    "Overriding to method='function_calling'."
+                )
+                method = "function_calling"
 
         if method == "function_calling":
             if schema is None:

libs/partners/openai/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "poetry.core.masonry.api"
 
 [tool.poetry]
 name = "langchain-openai"
-version = "0.3.1"
+version = "0.3.2"
 description = "An integration package connecting OpenAI and LangChain"
 authors = []
 readme = "README.md"
@@ -23,7 +23,7 @@ ignore_missing_imports = true
 
 [tool.poetry.dependencies]
 python = ">=3.9,<4.0"
-langchain-core = "^0.3.30"
+langchain-core = "^0.3.31"
 openai = "^1.58.1"
 tiktoken = ">=0.7,<1"
 
@@ -67,6 +67,7 @@ pytest-watcher = "^0.3.4"
 pytest-asyncio = "^0.21.1"
 pytest-cov = "^4.1.0"
 pytest-socket = "^0.6.0"
+pytest-xdist = "^3.6.1"
 [[tool.poetry.group.test.dependencies.numpy]]
 version = "^1"
 python = "<3.12"

libs/partners/openai/tests/integration_tests/chat_models/test_base.py

@@ -642,7 +642,7 @@ def test_disable_parallel_tool_calling() -> None:
     assert len(result.tool_calls) == 1
 
 
-@pytest.mark.parametrize("model", ["gpt-4o-mini", "o1"])
+@pytest.mark.parametrize("model", ["gpt-4o-mini", "o1", "gpt-4"])
 def test_openai_structured_output(model: str) -> None:
     class MyModel(BaseModel):
         """A Person"""
@@ -657,24 +657,6 @@ class MyModel(BaseModel):
     assert result.age == 27
 
 
-def test_structured_output_errors_with_legacy_models() -> None:
-    class MyModel(BaseModel):
-        """A Person"""
-
-        name: str
-        age: int
-
-    llm = ChatOpenAI(model="gpt-4").with_structured_output(MyModel)
-
-    with pytest.warns(UserWarning, match="with_structured_output"):
-        with pytest.raises(openai.BadRequestError):
-            _ = llm.invoke("I'm a 27 year old named Erick")
-
-    with pytest.warns(UserWarning, match="with_structured_output"):
-        with pytest.raises(openai.BadRequestError):
-            _ = list(llm.stream("I'm a 27 year old named Erick"))
-
-
 def test_openai_proxy() -> None:
     """Test ChatOpenAI with proxy."""
     chat_openai = ChatOpenAI(openai_proxy="http://localhost:8080")

libs/partners/openai/tests/unit_tests/chat_models/test_azure.py

@@ -3,6 +3,9 @@
 import os
 from unittest import mock
 
+import pytest
+from typing_extensions import TypedDict
+
 from langchain_openai import AzureChatOpenAI
 
 
@@ -59,3 +62,22 @@ def test_initialize_azure_openai_with_openai_api_base_set() -> None:
         ls_params = llm._get_ls_params()
         assert ls_params["ls_provider"] == "azure"
         assert ls_params["ls_model_name"] == "35-turbo-dev"
+
+
+def test_structured_output_old_model() -> None:
+    class Output(TypedDict):
+        """output."""
+
+        foo: str
+
+    with pytest.warns(match="Cannot use method='json_schema'"):
+        llm = AzureChatOpenAI(  # type: ignore[call-arg]
+            model="gpt-35-turbo",
+            azure_deployment="35-turbo-dev",
+            openai_api_version="2023-05-15",
+            azure_endpoint="my-base-url",
+        ).with_structured_output(Output)
+
+    # assert tool calling was used instead of json_schema
+    assert "tools" in llm.steps[0].kwargs  # type: ignore
+    assert "response_format" not in llm.steps[0].kwargs  # type: ignore

libs/partners/openai/tests/unit_tests/chat_models/test_base.py

@@ -746,7 +746,7 @@ class Foo(BaseModel):
 def test_schema_from_with_structured_output(schema: Type) -> None:
     """Test schema from with_structured_output."""
 
-    llm = ChatOpenAI()
+    llm = ChatOpenAI(model="gpt-4o")
 
     structured_llm = llm.with_structured_output(
         schema, method="json_schema", strict=True
@@ -886,3 +886,16 @@ def test_init_o1() -> None:
     with pytest.warns(None) as record:  # type: ignore[call-overload]
         ChatOpenAI(model="o1", reasoning_effort="medium")
     assert len(record) == 0
+
+
+def test_structured_output_old_model() -> None:
+    class Output(TypedDict):
+        """output."""
+
+        foo: str
+
+    with pytest.warns(match="Cannot use method='json_schema'"):
+        llm = ChatOpenAI(model="gpt-4").with_structured_output(Output)
+    # assert tool calling was used instead of json_schema
+    assert "tools" in llm.steps[0].kwargs  # type: ignore
+    assert "response_format" not in llm.steps[0].kwargs  # type: ignore