feat: Structured output for Responses API

This adds the missing `text` parameter to the Responses API that is
how users control structured outputs. All we do with that parameter is
map it to the corresponding chat completion response_format.

The unit tests exercise the various permutations allowed for this
property, while a couple of new verification tests actually use it for
real to verify the model outputs are following the format as expected.

Signed-off-by: Ben Browning <bbrownin@redhat.com>

docs/_static/llama-stack-spec.html

@@ -7241,6 +7241,79 @@
                 ],
                 "title": "OpenAIResponseOutputMessageWebSearchToolCall"
             },
+            "OpenAIResponseText": {
+                "type": "object",
+                "properties": {
+                    "format": {
+                        "type": "object",
+                        "properties": {
+                            "type": {
+                                "oneOf": [
+                                    {
+                                        "type": "string",
+                                        "const": "text"
+                                    },
+                                    {
+                                        "type": "string",
+                                        "const": "json_schema"
+                                    },
+                                    {
+                                        "type": "string",
+                                        "const": "json_object"
+                                    }
+                                ],
+                                "description": "Must be \"text\", \"json_schema\", or \"json_object\" to identify the format type"
+                            },
+                            "name": {
+                                "type": "string",
+                                "description": "The name of the response format. Only used for json_schema."
+                            },
+                            "schema": {
+                                "type": "object",
+                                "additionalProperties": {
+                                    "oneOf": [
+                                        {
+                                            "type": "null"
+                                        },
+                                        {
+                                            "type": "boolean"
+                                        },
+                                        {
+                                            "type": "number"
+                                        },
+                                        {
+                                            "type": "string"
+                                        },
+                                        {
+                                            "type": "array"
+                                        },
+                                        {
+                                            "type": "object"
+                                        }
+                                    ]
+                                },
+                                "description": "The JSON schema the response should conform to. In a Python SDK, this is often a `pydantic` model. Only used for json_schema."
+                            },
+                            "description": {
+                                "type": "string",
+                                "description": "(Optional) A description of the response format. Only used for json_schema."
+                            },
+                            "strict": {
+                                "type": "boolean",
+                                "description": "(Optional) Whether to strictly enforce the JSON schema. If true, the response must match the schema exactly. Only used for json_schema."
+                            }
+                        },
+                        "additionalProperties": false,
+                        "required": [
+                            "type"
+                        ],
+                        "title": "OpenAIResponseTextFormat",
+                        "description": "Configuration for Responses API text format."
+                    }
+                },
+                "additionalProperties": false,
+                "title": "OpenAIResponseText"
+            },
             "CreateOpenaiResponseRequest": {
                 "type": "object",
                 "properties": {
@@ -7278,6 +7351,9 @@
                     "temperature": {
                         "type": "number"
                     },
+                    "text": {
+                        "$ref": "#/components/schemas/OpenAIResponseText"
+                    },
                     "tools": {
                         "type": "array",
                         "items": {
@@ -7351,6 +7427,9 @@
                     "temperature": {
                         "type": "number"
                     },
+                    "text": {
+                        "$ref": "#/components/schemas/OpenAIResponseText"
+                    },
                     "top_p": {
                         "type": "number"
                     },
@@ -7369,7 +7448,8 @@
                     "object",
                     "output",
                     "parallel_tool_calls",
-                    "status"
+                    "status",
+                    "text"
                 ],
                 "title": "OpenAIResponseObject"
             },
@@ -10406,6 +10486,9 @@
                     "temperature": {
                         "type": "number"
                     },
+                    "text": {
+                        "$ref": "#/components/schemas/OpenAIResponseText"
+                    },
                     "top_p": {
                         "type": "number"
                     },
@@ -10431,6 +10514,7 @@
                     "output",
                     "parallel_tool_calls",
                     "status",
+                    "text",
                     "input"
                 ],
                 "title": "OpenAIResponseObjectWithInput"

docs/_static/llama-stack-spec.yaml

@@ -5118,6 +5118,57 @@ components:
         - type
       title: >-
         OpenAIResponseOutputMessageWebSearchToolCall
+    OpenAIResponseText:
+      type: object
+      properties:
+        format:
+          type: object
+          properties:
+            type:
+              oneOf:
+                - type: string
+                  const: text
+                - type: string
+                  const: json_schema
+                - type: string
+                  const: json_object
+              description: >-
+                Must be "text", "json_schema", or "json_object" to identify the format
+                type
+            name:
+              type: string
+              description: >-
+                The name of the response format. Only used for json_schema.
+            schema:
+              type: object
+              additionalProperties:
+                oneOf:
+                  - type: 'null'
+                  - type: boolean
+                  - type: number
+                  - type: string
+                  - type: array
+                  - type: object
+              description: >-
+                The JSON schema the response should conform to. In a Python SDK, this
+                is often a `pydantic` model. Only used for json_schema.
+            description:
+              type: string
+              description: >-
+                (Optional) A description of the response format. Only used for json_schema.
+            strict:
+              type: boolean
+              description: >-
+                (Optional) Whether to strictly enforce the JSON schema. If true, the
+                response must match the schema exactly. Only used for json_schema.
+          additionalProperties: false
+          required:
+            - type
+          title: OpenAIResponseTextFormat
+          description: >-
+            Configuration for Responses API text format.
+      additionalProperties: false
+      title: OpenAIResponseText
     CreateOpenaiResponseRequest:
       type: object
       properties:
@@ -5145,6 +5196,8 @@ components:
           type: boolean
         temperature:
           type: number
+        text:
+          $ref: '#/components/schemas/OpenAIResponseText'
         tools:
           type: array
           items:
@@ -5196,6 +5249,8 @@ components:
           type: string
         temperature:
           type: number
+        text:
+          $ref: '#/components/schemas/OpenAIResponseText'
         top_p:
           type: number
         truncation:
@@ -5211,6 +5266,7 @@ components:
         - output
         - parallel_tool_calls
         - status
+        - text
       title: OpenAIResponseObject
     OpenAIResponseOutput:
       oneOf:
@@ -7288,6 +7344,8 @@ components:
           type: string
         temperature:
           type: number
+        text:
+          $ref: '#/components/schemas/OpenAIResponseText'
         top_p:
           type: number
         truncation:
@@ -7307,6 +7365,7 @@ components:
         - output
         - parallel_tool_calls
         - status
+        - text
         - input
       title: OpenAIResponseObjectWithInput
     ListProvidersResponse:

llama_stack/apis/agents/agents.py

@@ -37,6 +37,7 @@ from .openai_responses import (
     OpenAIResponseInputTool,
     OpenAIResponseObject,
     OpenAIResponseObjectStream,
+    OpenAIResponseText,
 )
 
 # TODO: use enum.StrEnum when we drop support for python 3.10
@@ -603,6 +604,7 @@ class Agents(Protocol):
         store: bool | None = True,
         stream: bool | None = False,
         temperature: float | None = None,
+        text: OpenAIResponseText | None = None,
         tools: list[OpenAIResponseInputTool] | None = None,
         max_infer_iters: int | None = 10,  # this is an extension to the OpenAI API
     ) -> OpenAIResponseObject | AsyncIterator[OpenAIResponseObjectStream]:

llama_stack/apis/agents/openai_responses.py

@@ -7,6 +7,7 @@
 from typing import Annotated, Any, Literal
 
 from pydantic import BaseModel, Field
+from typing_extensions import TypedDict
 
 from llama_stack.schema_utils import json_schema_type, register_schema
 
@@ -126,6 +127,32 @@ OpenAIResponseOutput = Annotated[
 register_schema(OpenAIResponseOutput, name="OpenAIResponseOutput")
 
 
+# This has to be a TypedDict because we need a "schema" field and our strong
+# typing code in the schema generator doesn't support Pydantic aliases. That also
+# means we can't use a discriminator field here, because TypedDicts don't support
+# default values which the strong typing code requires for discriminators.
+class OpenAIResponseTextFormat(TypedDict, total=False):
+    """Configuration for Responses API text format.
+
+    :param type: Must be "text", "json_schema", or "json_object" to identify the format type
+    :param name: The name of the response format. Only used for json_schema.
+    :param schema: The JSON schema the response should conform to. In a Python SDK, this is often a `pydantic` model. Only used for json_schema.
+    :param description: (Optional) A description of the response format. Only used for json_schema.
+    :param strict: (Optional) Whether to strictly enforce the JSON schema. If true, the response must match the schema exactly. Only used for json_schema.
+    """
+
+    type: Literal["text"] | Literal["json_schema"] | Literal["json_object"]
+    name: str | None
+    schema: dict[str, Any] | None
+    description: str | None
+    strict: bool | None
+
+
+@json_schema_type
+class OpenAIResponseText(BaseModel):
+    format: OpenAIResponseTextFormat | None = None
+
+
 @json_schema_type
 class OpenAIResponseObject(BaseModel):
     created_at: int
@@ -138,6 +165,9 @@ class OpenAIResponseObject(BaseModel):
     previous_response_id: str | None = None
     status: str
     temperature: float | None = None
+    # Default to text format to avoid breaking the loading of old responses
+    # before the field was added. New responses will have this set always.
+    text: OpenAIResponseText = OpenAIResponseText(format=OpenAIResponseTextFormat(type="text"))
     top_p: float | None = None
     truncation: str | None = None
     user: str | None = None

llama_stack/providers/inline/agents/meta_reference/agents.py

@@ -29,6 +29,7 @@ from llama_stack.apis.agents import (
     Session,
     Turn,
 )
+from llama_stack.apis.agents.openai_responses import OpenAIResponseText
 from llama_stack.apis.common.responses import PaginatedResponse
 from llama_stack.apis.inference import (
     Inference,
@@ -324,11 +325,12 @@ class MetaReferenceAgentsImpl(Agents):
         store: bool | None = True,
         stream: bool | None = False,
         temperature: float | None = None,
+        text: OpenAIResponseText | None = None,
         tools: list[OpenAIResponseInputTool] | None = None,
         max_infer_iters: int | None = 10,
     ) -> OpenAIResponseObject:
         return await self.openai_responses_impl.create_openai_response(
-            input, model, instructions, previous_response_id, store, stream, temperature, tools, max_infer_iters
+            input, model, instructions, previous_response_id, store, stream, temperature, text, tools, max_infer_iters
         )
 
     async def list_openai_responses(

llama_stack/providers/inline/agents/meta_reference/openai_responses.py

@@ -37,6 +37,8 @@ from llama_stack.apis.agents.openai_responses import (
     OpenAIResponseOutputMessageFunctionToolCall,
     OpenAIResponseOutputMessageMCPListTools,
     OpenAIResponseOutputMessageWebSearchToolCall,
+    OpenAIResponseText,
+    OpenAIResponseTextFormat,
 )
 from llama_stack.apis.inference.inference import (
     Inference,
@@ -50,7 +52,12 @@ from llama_stack.apis.inference.inference import (
     OpenAIChoice,
     OpenAIDeveloperMessageParam,
     OpenAIImageURL,
+    OpenAIJSONSchema,
     OpenAIMessageParam,
+    OpenAIResponseFormatJSONObject,
+    OpenAIResponseFormatJSONSchema,
+    OpenAIResponseFormatParam,
+    OpenAIResponseFormatText,
     OpenAISystemMessageParam,
     OpenAIToolMessageParam,
     OpenAIUserMessageParam,
@@ -158,6 +165,21 @@ async def _convert_chat_choice_to_response_message(choice: OpenAIChoice) -> Open
     )
 
 
+async def _convert_response_text_to_chat_response_format(text: OpenAIResponseText) -> OpenAIResponseFormatParam:
+    """
+    Convert an OpenAI Response text parameter into an OpenAI Chat Completion response format.
+    """
+    if not text.format or text.format["type"] == "text":
+        return OpenAIResponseFormatText(type="text")
+    if text.format["type"] == "json_object":
+        return OpenAIResponseFormatJSONObject()
+    if text.format["type"] == "json_schema":
+        return OpenAIResponseFormatJSONSchema(
+            json_schema=OpenAIJSONSchema(name=text.format["name"], schema=text.format["schema"])
+        )
+    raise ValueError(f"Unsupported text format: {text.format}")
+
+
 async def _get_message_type_by_role(role: str):
     role_to_type = {
         "user": OpenAIUserMessageParam,
@@ -180,6 +202,7 @@ class ChatCompletionContext(BaseModel):
     mcp_tool_to_server: dict[str, OpenAIResponseInputToolMCP]
     stream: bool
     temperature: float | None
+    response_format: OpenAIResponseFormatParam
 
 
 class OpenAIResponsesImpl:
@@ -343,10 +366,12 @@ class OpenAIResponsesImpl:
         store: bool | None = True,
         stream: bool | None = False,
         temperature: float | None = None,
+        text: OpenAIResponseText | None = None,
         tools: list[OpenAIResponseInputTool] | None = None,
         max_infer_iters: int | None = 10,
     ):
         stream = False if stream is None else stream
+        text = OpenAIResponseText(format=OpenAIResponseTextFormat(type="text")) if text is None else text
 
         output_messages: list[OpenAIResponseOutput] = []
 
@@ -355,6 +380,9 @@ class OpenAIResponsesImpl:
         messages = await _convert_response_input_to_chat_messages(input)
         await self._prepend_instructions(messages, instructions)
 
+        # Structured outputs
+        response_format = await _convert_response_text_to_chat_response_format(text)
+
         # Tool setup
         chat_tools, mcp_tool_to_server, mcp_list_message = (
             await self._convert_response_tools_to_chat_tools(tools) if tools else (None, {}, None)
@@ -369,6 +397,7 @@ class OpenAIResponsesImpl:
             mcp_tool_to_server=mcp_tool_to_server,
             stream=stream,
             temperature=temperature,
+            response_format=response_format,
         )
 
         # Fork to streaming vs non-streaming - let each handle ALL inference rounds
@@ -379,6 +408,7 @@ class OpenAIResponsesImpl:
                 input=input,
                 model=model,
                 store=store,
+                text=text,
                 tools=tools,
                 max_infer_iters=max_infer_iters,
             )
@@ -389,6 +419,7 @@ class OpenAIResponsesImpl:
                 input=input,
                 model=model,
                 store=store,
+                text=text,
                 tools=tools,
                 max_infer_iters=max_infer_iters,
             )
@@ -400,6 +431,7 @@ class OpenAIResponsesImpl:
         input: str | list[OpenAIResponseInput],
         model: str,
         store: bool | None,
+        text: OpenAIResponseText,
         tools: list[OpenAIResponseInputTool] | None,
         max_infer_iters: int | None,
     ) -> OpenAIResponseObject:
@@ -416,6 +448,7 @@ class OpenAIResponsesImpl:
                 tools=ctx.tools,
                 stream=False,
                 temperature=ctx.temperature,
+                response_format=ctx.response_format,
             )
             current_response = OpenAIChatCompletion(**inference_result.model_dump())
 
@@ -470,6 +503,7 @@ class OpenAIResponsesImpl:
             object="response",
             status="completed",
             output=output_messages,
+            text=text,
         )
         logger.debug(f"OpenAI Responses response: {response}")
 
@@ -489,6 +523,7 @@ class OpenAIResponsesImpl:
         input: str | list[OpenAIResponseInput],
         model: str,
         store: bool | None,
+        text: OpenAIResponseText,
         tools: list[OpenAIResponseInputTool] | None,
         max_infer_iters: int | None,
     ) -> AsyncIterator[OpenAIResponseObjectStream]:
@@ -503,6 +538,7 @@ class OpenAIResponsesImpl:
             object="response",
             status="in_progress",
             output=output_messages.copy(),
+            text=text,
         )
 
         # Emit response.created immediately
@@ -520,6 +556,7 @@ class OpenAIResponsesImpl:
                 tools=ctx.tools,
                 stream=True,
                 temperature=ctx.temperature,
+                response_format=ctx.response_format,
             )
 
             # Process streaming chunks and build complete response
@@ -645,6 +682,7 @@ class OpenAIResponsesImpl:
             model=model,
             object="response",
             status="completed",
+            text=text,
             output=output_messages,
         )
 

tests/unit/providers/agents/meta_reference/test_openai_responses.py

@@ -25,11 +25,17 @@ from llama_stack.apis.agents.openai_responses import (
     OpenAIResponseObjectWithInput,
     OpenAIResponseOutputMessageContentOutputText,
     OpenAIResponseOutputMessageWebSearchToolCall,
+    OpenAIResponseText,
+    OpenAIResponseTextFormat,
 )
 from llama_stack.apis.inference.inference import (
     OpenAIAssistantMessageParam,
     OpenAIChatCompletionContentPartTextParam,
     OpenAIDeveloperMessageParam,
+    OpenAIJSONSchema,
+    OpenAIResponseFormatJSONObject,
+    OpenAIResponseFormatJSONSchema,
+    OpenAIResponseFormatText,
     OpenAIUserMessageParam,
 )
 from llama_stack.apis.tools.tools import Tool, ToolGroups, ToolInvocationResult, ToolParameter, ToolRuntime
@@ -96,6 +102,7 @@ async def test_create_openai_response_with_string_input(openai_responses_impl, m
     mock_inference_api.openai_chat_completion.assert_called_once_with(
         model=model,
         messages=[OpenAIUserMessageParam(role="user", content="What is the capital of Ireland?", name=None)],
+        response_format=OpenAIResponseFormatText(),
         tools=None,
         stream=False,
         temperature=0.1,
@@ -320,6 +327,7 @@ async def test_prepend_previous_response_basic(openai_responses_impl, mock_respo
         model="fake_model",
         output=[response_output_message],
         status="completed",
+        text=OpenAIResponseText(format=OpenAIResponseTextFormat(type="text")),
         input=[input_item_message],
     )
     mock_responses_store.get_response_object.return_value = previous_response
@@ -362,6 +370,7 @@ async def test_prepend_previous_response_web_search(openai_responses_impl, mock_
         model="fake_model",
         output=[output_web_search, output_message],
         status="completed",
+        text=OpenAIResponseText(format=OpenAIResponseTextFormat(type="text")),
         input=[input_item_message],
     )
     mock_responses_store.get_response_object.return_value = response
@@ -483,6 +492,7 @@ async def test_create_openai_response_with_instructions_and_previous_response(
         model="fake_model",
         output=[response_output_message],
         status="completed",
+        text=OpenAIResponseText(format=OpenAIResponseTextFormat(type="text")),
         input=[input_item_message],
     )
     mock_responses_store.get_response_object.return_value = response
@@ -576,6 +586,7 @@ async def test_responses_store_list_input_items_logic():
         object="response",
         status="completed",
         output=[],
+        text=OpenAIResponseText(format=(OpenAIResponseTextFormat(type="text"))),
         input=input_items,
     )
 
@@ -644,6 +655,7 @@ async def test_store_response_uses_rehydrated_input_with_previous_response(
         created_at=1234567890,
         model="meta-llama/Llama-3.1-8B-Instruct",
         status="completed",
+        text=OpenAIResponseText(format=OpenAIResponseTextFormat(type="text")),
         input=[
             OpenAIResponseMessage(
                 id="msg-prev-user", role="user", content=[OpenAIResponseInputMessageContentText(text="What is 2+2?")]
@@ -694,3 +706,61 @@ async def test_store_response_uses_rehydrated_input_with_previous_response(
     # Verify the response itself is correct
     assert result.model == model
     assert result.status == "completed"
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    "text_format, response_format",
+    [
+        (OpenAIResponseText(format=OpenAIResponseTextFormat(type="text")), OpenAIResponseFormatText()),
+        (
+            OpenAIResponseText(format=OpenAIResponseTextFormat(name="Test", schema={"foo": "bar"}, type="json_schema")),
+            OpenAIResponseFormatJSONSchema(json_schema=OpenAIJSONSchema(name="Test", schema={"foo": "bar"})),
+        ),
+        (OpenAIResponseText(format=OpenAIResponseTextFormat(type="json_object")), OpenAIResponseFormatJSONObject()),
+        # ensure text param with no format specified defaults to text
+        (OpenAIResponseText(format=None), OpenAIResponseFormatText()),
+        # ensure text param of None defaults to text
+        (None, OpenAIResponseFormatText()),
+    ],
+)
+async def test_create_openai_response_with_text_format(
+    openai_responses_impl, mock_inference_api, text_format, response_format
+):
+    """Test creating Responses with text formats."""
+    # Setup
+    input_text = "How hot it is in San Francisco today?"
+    model = "meta-llama/Llama-3.1-8B-Instruct"
+
+    # Load the chat completion fixture
+    mock_chat_completion = load_chat_completion_fixture("simple_chat_completion.yaml")
+    mock_inference_api.openai_chat_completion.return_value = mock_chat_completion
+
+    # Execute
+    _result = await openai_responses_impl.create_openai_response(
+        input=input_text,
+        model=model,
+        text=text_format,
+    )
+
+    # Verify
+    first_call = mock_inference_api.openai_chat_completion.call_args_list[0]
+    assert first_call.kwargs["messages"][0].content == input_text
+    assert first_call.kwargs["response_format"] is not None
+    assert first_call.kwargs["response_format"] == response_format
+
+
+@pytest.mark.asyncio
+async def test_create_openai_response_with_invalid_text_format(openai_responses_impl, mock_inference_api):
+    """Test creating an OpenAI response with an invalid text format."""
+    # Setup
+    input_text = "How hot it is in San Francisco today?"
+    model = "meta-llama/Llama-3.1-8B-Instruct"
+
+    # Execute
+    with pytest.raises(ValueError):
+        _result = await openai_responses_impl.create_openai_response(
+            input=input_text,
+            model=model,
+            text=OpenAIResponseText(format={"type": "invalid"}),
+        )

tests/verifications/openai_api/test_responses.py

@@ -546,3 +546,39 @@ async def test_response_streaming_multi_turn_tool_execution(
             assert expected_output.lower() in final_response.output_text.lower(), (
                 f"Expected '{expected_output}' to appear in response: {final_response.output_text}"
             )
+
+
+@pytest.mark.parametrize(
+    "text_format",
+    # Not testing json_object because most providers don't actually support it.
+    [
+        {"type": "text"},
+        {
+            "type": "json_schema",
+            "name": "capitals",
+            "description": "A schema for the capital of each country",
+            "schema": {"type": "object", "properties": {"capital": {"type": "string"}}},
+            "strict": True,
+        },
+    ],
+)
+def test_response_text_format(request, openai_client, model, provider, verification_config, text_format):
+    if isinstance(openai_client, LlamaStackAsLibraryClient):
+        pytest.skip("Responses API text format is not yet supported in library client.")
+
+    test_name_base = get_base_test_name(request)
+    if should_skip_test(verification_config, provider, model, test_name_base):
+        pytest.skip(f"Skipping {test_name_base} for model {model} on provider {provider} based on config.")
+
+    stream = False
+    response = openai_client.responses.create(
+        model=model,
+        input="What is the capital of France?",
+        stream=stream,
+        text={"format": text_format},
+    )
+    # by_alias=True is needed because otherwise Pydantic renames our "schema" field
+    assert response.text.format.model_dump(exclude_none=True, by_alias=True) == text_format
+    assert "paris" in response.output_text.lower()
+    if text_format["type"] == "json_schema":
+        assert "paris" in json.loads(response.output_text)["capital"].lower()