From 4c6af2d1b2cad7b169d287f1cf228788fc84043f Mon Sep 17 00:00:00 2001
From: Mason Daugherty <mason@langchain.dev>
Date: Tue, 9 Sep 2025 11:37:50 -0400
Subject: [PATCH] fix(openai): structured output (#32551)

---
 .../langchain_openai/chat_models/base.py      | 27 +++++------
 .../tests/unit_tests/chat_models/test_base.py | 48 ++++++++++++++++++-
 2 files changed, 60 insertions(+), 15 deletions(-)

diff --git a/libs/partners/openai/langchain_openai/chat_models/base.py b/libs/partners/openai/langchain_openai/chat_models/base.py
index 38a4568d90e..1cc33ca0075 100644
--- a/libs/partners/openai/langchain_openai/chat_models/base.py
+++ b/libs/partners/openai/langchain_openai/chat_models/base.py
@@ -3586,13 +3586,6 @@ def _construct_responses_api_payload(
 
     # Structured output
     if schema := payload.pop("response_format", None):
-        if payload.get("text"):
-            text = payload["text"]
-            raise ValueError(
-                "Can specify at most one of 'response_format' or 'text', received both:"
-                f"\n{schema=}\n{text=}"
-            )
-
         # For pydantic + non-streaming case, we use responses.parse.
         # Otherwise, we use responses.create.
         strict = payload.pop("strict", None)
@@ -3605,7 +3598,10 @@ def _construct_responses_api_payload(
             else:
                 schema_dict = schema
             if schema_dict == {"type": "json_object"}:  # JSON mode
-                payload["text"] = {"format": {"type": "json_object"}}
+                if "text" in payload and isinstance(payload["text"], dict):
+                    payload["text"]["format"] = {"type": "json_object"}
+                else:
+                    payload["text"] = {"format": {"type": "json_object"}}
             elif (
                 (
                     response_format := _convert_to_openai_response_format(
@@ -3615,17 +3611,20 @@ def _construct_responses_api_payload(
                 and (isinstance(response_format, dict))
                 and (response_format["type"] == "json_schema")
             ):
-                payload["text"] = {
-                    "format": {"type": "json_schema", **response_format["json_schema"]}
-                }
+                format_value = {"type": "json_schema", **response_format["json_schema"]}
+                if "text" in payload and isinstance(payload["text"], dict):
+                    payload["text"]["format"] = format_value
+                else:
+                    payload["text"] = {"format": format_value}
             else:
                 pass
 
     verbosity = payload.pop("verbosity", None)
     if verbosity is not None:
-        if "text" not in payload:
-            payload["text"] = {"format": {"type": "text"}}
-        payload["text"]["verbosity"] = verbosity
+        if "text" in payload and isinstance(payload["text"], dict):
+            payload["text"]["verbosity"] = verbosity
+        else:
+            payload["text"] = {"verbosity": verbosity}
 
     return payload
 
diff --git a/libs/partners/openai/tests/unit_tests/chat_models/test_base.py b/libs/partners/openai/tests/unit_tests/chat_models/test_base.py
index cf24d857de4..01a1e1cae98 100644
--- a/libs/partners/openai/tests/unit_tests/chat_models/test_base.py
+++ b/libs/partners/openai/tests/unit_tests/chat_models/test_base.py
@@ -23,6 +23,7 @@ from langchain_core.messages import (
 from langchain_core.messages.ai import UsageMetadata
 from langchain_core.outputs import ChatGeneration, ChatResult
 from langchain_core.runnables import RunnableLambda
+from langchain_core.runnables.base import RunnableBinding, RunnableSequence
 from langchain_core.tracers.base import BaseTracer
 from langchain_core.tracers.schemas import Run
 from openai.types.responses import ResponseOutputMessage, ResponseReasoningItem
@@ -1201,7 +1202,6 @@ def test_verbosity_parameter_payload() -> None:
     payload = llm._get_request_payload(messages, stop=None)
 
     assert payload["text"]["verbosity"] == "high"
-    assert payload["text"]["format"]["type"] == "text"
 
 
 def test_structured_output_old_model() -> None:
@@ -2718,6 +2718,52 @@ def test_extra_body_with_model_kwargs() -> None:
     assert payload["temperature"] == 0.5
 
 
+@pytest.mark.parametrize("verbosity_format", ["model_kwargs", "top_level"])
+@pytest.mark.parametrize("streaming", [False, True])
+@pytest.mark.parametrize("schema_format", ["pydantic", "dict"])
+def test_structured_output_verbosity(
+    verbosity_format: str, streaming: bool, schema_format: str
+) -> None:
+    class MySchema(BaseModel):
+        foo: str
+
+    if verbosity_format == "model_kwargs":
+        init_params: dict[str, Any] = {"model_kwargs": {"text": {"verbosity": "high"}}}
+    else:
+        init_params = {"verbosity": "high"}
+
+    if streaming:
+        init_params["streaming"] = True
+
+    llm = ChatOpenAI(model="gpt-5", use_responses_api=True, **init_params)
+
+    if schema_format == "pydantic":
+        schema: Any = MySchema
+    else:
+        schema = MySchema.model_json_schema()
+
+    structured_llm = llm.with_structured_output(schema)
+    sequence = cast(RunnableSequence, structured_llm)
+    binding = cast(RunnableBinding, sequence.first)
+    bound_llm = cast(ChatOpenAI, binding.bound)
+    bound_kwargs = binding.kwargs
+
+    messages = [HumanMessage(content="Hello")]
+    payload = bound_llm._get_request_payload(messages, **bound_kwargs)
+
+    # Verify that verbosity is present in `text` param
+    assert "text" in payload
+    assert "verbosity" in payload["text"]
+    assert payload["text"]["verbosity"] == "high"
+
+    # Verify that schema is passed correctly
+    if schema_format == "pydantic" and not streaming:
+        assert payload["text_format"] == schema
+    else:
+        assert "format" in payload["text"]
+        assert payload["text"]["format"]["type"] == "json_schema"
+
+
 @pytest.mark.parametrize("use_responses_api", [False, True])
 def test_gpt_5_temperature(use_responses_api: bool) -> None:
     llm = ChatOpenAI(