llama-index-llms-openai 0.3.28__py3-none-any.whl → 0.3.29__py3-none-any.whl

llama_index/llms/openai/__init__.py

@@ -1,3 +1,4 @@
 from llama_index.llms.openai.base import AsyncOpenAI, OpenAI, SyncOpenAI, Tokenizer
+from llama_index.llms.openai.responses import OpenAIResponses
 
-__all__ = ["OpenAI", "Tokenizer", "SyncOpenAI", "AsyncOpenAI"]
+__all__ = ["OpenAI", "OpenAIResponses", "Tokenizer", "SyncOpenAI", "AsyncOpenAI"]

llama_index/llms/openai/responses.py

@@ -0,0 +1,952 @@
+import functools
+import httpx
+import tiktoken
+from openai import AsyncOpenAI, AzureOpenAI
+from openai import OpenAI as SyncOpenAI
+from openai.types.responses import (
+    Response,
+    ResponseStreamEvent,
+    ResponseCompletedEvent,
+    ResponseCreatedEvent,
+    ResponseFileSearchCallCompletedEvent,
+    ResponseFunctionCallArgumentsDeltaEvent,
+    ResponseFunctionCallArgumentsDoneEvent,
+    ResponseInProgressEvent,
+    ResponseOutputItemAddedEvent,
+    ResponseTextAnnotationDeltaEvent,
+    ResponseTextDeltaEvent,
+    ResponseWebSearchCallCompletedEvent,
+    ResponseOutputItem,
+    ResponseOutputMessage,
+    ResponseFileSearchToolCall,
+    ResponseFunctionToolCall,
+    ResponseFunctionWebSearch,
+    ResponseComputerToolCall,
+    ResponseReasoningItem,
+)
+from typing import (
+    TYPE_CHECKING,
+    Any,
+    AsyncGenerator,
+    Callable,
+    Dict,
+    Generator,
+    List,
+    Literal,
+    Optional,
+    Protocol,
+    Sequence,
+    Tuple,
+    Type,
+    Union,
+    runtime_checkable,
+)
+
+import llama_index.core.instrumentation as instrument
+from llama_index.core.base.llms.generic_utils import (
+    achat_to_completion_decorator,
+    astream_chat_to_completion_decorator,
+    chat_to_completion_decorator,
+    stream_chat_to_completion_decorator,
+)
+from llama_index.core.base.llms.types import (
+    ChatMessage,
+    ChatResponse,
+    ChatResponseAsyncGen,
+    ChatResponseGen,
+    CompletionResponse,
+    CompletionResponseAsyncGen,
+    CompletionResponseGen,
+    LLMMetadata,
+    MessageRole,
+    TextBlock,
+)
+from llama_index.core.bridge.pydantic import (
+    Field,
+    PrivateAttr,
+)
+from llama_index.core.constants import (
+    DEFAULT_TEMPERATURE,
+)
+from llama_index.core.llms.callbacks import (
+    llm_chat_callback,
+    llm_completion_callback,
+)
+from llama_index.core.llms.function_calling import FunctionCallingLLM
+from llama_index.core.llms.llm import ToolSelection, Model
+from llama_index.core.llms.utils import parse_partial_json
+from llama_index.core.prompts import PromptTemplate
+from llama_index.core.program.utils import FlexibleModel
+from llama_index.llms.openai.utils import (
+    O1_MODELS,
+    create_retry_decorator,
+    is_function_calling_model,
+    openai_modelname_to_contextsize,
+    resolve_openai_credentials,
+    resolve_tool_choice,
+    to_openai_message_dicts,
+)
+
+
+dispatcher = instrument.get_dispatcher(__name__)
+
+if TYPE_CHECKING:
+    from llama_index.core.tools.types import BaseTool
+
+DEFAULT_OPENAI_MODEL = "gpt-4o-mini"
+
+
+def llm_retry_decorator(f: Callable[..., Any]) -> Callable[..., Any]:
+    @functools.wraps(f)
+    def wrapper(self, *args: Any, **kwargs: Any) -> Any:
+        max_retries = getattr(self, "max_retries", 0)
+        if max_retries <= 0:
+            return f(self, *args, **kwargs)
+
+        retry = create_retry_decorator(
+            max_retries=max_retries,
+            random_exponential=True,
+            stop_after_delay_seconds=60,
+            min_seconds=1,
+            max_seconds=20,
+        )
+        return retry(f)(self, *args, **kwargs)
+
+    return wrapper
+
+
+@runtime_checkable
+class Tokenizer(Protocol):
+    """Tokenizers support an encode function that returns a list of ints."""
+
+    def encode(self, text: str) -> List[int]:  # fmt: skip
+        ...
+
+
+def force_single_tool_call(response: ChatResponse) -> None:
+    tool_calls = response.message.additional_kwargs.get("tool_calls", [])
+    if len(tool_calls) > 1:
+        response.message.additional_kwargs["tool_calls"] = [tool_calls[0]]
+
+
+class OpenAIResponses(FunctionCallingLLM):
+    """
+    OpenAI Responses LLM.
+
+    Args:
+        model: name of the OpenAI model to use.
+        temperature: a float from 0 to 1 controlling randomness in generation; higher will lead to more creative, less deterministic responses.
+        max_output_tokens: the maximum number of tokens to generate.
+        include: Additional output data to include in the model response.
+        instructions: Instructions for the model to follow.
+        track_previous_responses: Whether to track previous responses. If true, the LLM class will statefully track previous responses.
+        store: Whether to store previous responses in OpenAI's storage.
+        built_in_tools: The built-in tools to use for the model to augment responses.
+        truncation: Whether to auto-truncate the input if it exceeds the model's context window.
+        user: An optional identifier to help track the user's requests for abuse.
+        strict: Whether to enforce strict validation of the structured output.
+        additional_kwargs: Add additional parameters to OpenAI request body.
+        max_retries: How many times to retry the API call if it fails.
+        timeout: How long to wait, in seconds, for an API call before failing.
+        api_key: Your OpenAI api key
+        api_base: The base URL of the API to call
+        api_version: the version of the API to call
+        default_headers: override the default headers for API requests.
+        http_client: pass in your own httpx.Client instance.
+        async_http_client: pass in your own httpx.AsyncClient instance.
+
+    Examples:
+        `pip install llama-index-llms-openai`
+
+        ```python
+        from llama_index.llms.openai import OpenAIResponses
+
+        llm = OpenAIResponses(model="gpt-4o-mini", api_key="sk-...")
+
+        response = llm.complete("Hi, write a short story")
+        print(response.text)
+        ```
+    """
+
+    model: str = Field(
+        default=DEFAULT_OPENAI_MODEL, description="The OpenAI model to use."
+    )
+    temperature: float = Field(
+        default=DEFAULT_TEMPERATURE,
+        description="The temperature to use during generation.",
+        ge=0.0,
+        le=2.0,
+    )
+    top_p: float = Field(
+        default=1.0,
+        description="The top-p value to use during generation.",
+        ge=0.0,
+        le=1.0,
+    )
+    max_output_tokens: Optional[int] = Field(
+        description="The maximum number of tokens to generate.",
+        gt=0,
+    )
+    include: Optional[List[str]] = Field(
+        default=None,
+        description="Additional output data to include in the model response.",
+    )
+    instructions: Optional[str] = Field(
+        default=None,
+        description="Instructions for the model to follow.",
+    )
+    track_previous_responses: bool = Field(
+        default=False,
+        description="Whether to track previous responses. If true, the LLM class will statefully track previous responses.",
+    )
+    store: bool = Field(
+        default=False,
+        description="Whether to store previous responses in OpenAI's storage.",
+    )
+    built_in_tools: Optional[List[dict]] = Field(
+        default=None,
+        description="The built-in tools to use for the model to augment responses.",
+    )
+    truncation: str = Field(
+        default="disabled",
+        description="Whether to auto-truncate the input if it exceeds the model's context window.",
+    )
+    user: Optional[str] = Field(
+        default=None,
+        description="An optional identifier to help track the user's requests for abuse.",
+    )
+    call_metadata: Optional[Dict[str, Any]] = Field(
+        default=None,
+        description="Metadata to include in the API call.",
+    )
+    additional_kwargs: Dict[str, Any] = Field(
+        default_factory=dict,
+        description="Additional kwargs for the OpenAI API at inference time.",
+    )
+    max_retries: int = Field(
+        default=3,
+        description="The maximum number of API retries.",
+        ge=0,
+    )
+    timeout: float = Field(
+        default=60.0,
+        description="The timeout, in seconds, for API requests.",
+        ge=0,
+    )
+    strict: bool = Field(
+        default=False,
+        description="Whether to enforce strict validation of the structured output.",
+    )
+    default_headers: Optional[Dict[str, str]] = Field(
+        default=None, description="The default headers for API requests."
+    )
+    api_key: str = Field(default=None, description="The OpenAI API key.")
+    api_base: str = Field(description="The base URL for OpenAI API.")
+    api_version: str = Field(description="The API version for OpenAI API.")
+    reasoning_effort: Optional[Literal["low", "medium", "high"]] = Field(
+        default=None,
+        description="The effort to use for reasoning models.",
+    )
+
+    _client: SyncOpenAI = PrivateAttr()
+    _aclient: AsyncOpenAI = PrivateAttr()
+    _http_client: Optional[httpx.Client] = PrivateAttr()
+    _async_http_client: Optional[httpx.AsyncClient] = PrivateAttr()
+    _previous_response_id: Optional[str] = PrivateAttr()
+
+    def __init__(
+        self,
+        model: str = DEFAULT_OPENAI_MODEL,
+        temperature: float = DEFAULT_TEMPERATURE,
+        max_output_tokens: Optional[int] = None,
+        reasoning_effort: Optional[Literal["low", "medium", "high"]] = None,
+        include: Optional[List[str]] = None,
+        instructions: Optional[str] = None,
+        track_previous_responses: bool = False,
+        store: bool = False,
+        built_in_tools: Optional[List[dict]] = None,
+        truncation: str = "disabled",
+        user: Optional[str] = None,
+        previous_response_id: Optional[str] = None,
+        call_metadata: Optional[Dict[str, Any]] = None,
+        strict: bool = False,
+        additional_kwargs: Optional[Dict[str, Any]] = None,
+        max_retries: int = 3,
+        timeout: float = 60.0,
+        api_key: Optional[str] = None,
+        api_base: Optional[str] = None,
+        api_version: Optional[str] = None,
+        default_headers: Optional[Dict[str, str]] = None,
+        http_client: Optional[httpx.Client] = None,
+        async_http_client: Optional[httpx.AsyncClient] = None,
+        openai_client: Optional[SyncOpenAI] = None,
+        async_openai_client: Optional[AsyncOpenAI] = None,
+        **kwargs: Any,
+    ) -> None:
+        additional_kwargs = additional_kwargs or {}
+
+        api_key, api_base, api_version = resolve_openai_credentials(
+            api_key=api_key,
+            api_base=api_base,
+            api_version=api_version,
+        )
+
+        # TODO: Temp forced to 1.0 for o1
+        if model in O1_MODELS:
+            temperature = 1.0
+
+        super().__init__(
+            model=model,
+            temperature=temperature,
+            max_output_tokens=max_output_tokens,
+            reasoning_effort=reasoning_effort,
+            include=include,
+            instructions=instructions,
+            track_previous_responses=track_previous_responses,
+            store=store,
+            built_in_tools=built_in_tools,
+            truncation=truncation,
+            user=user,
+            additional_kwargs=additional_kwargs,
+            max_retries=max_retries,
+            api_key=api_key,
+            api_version=api_version,
+            api_base=api_base,
+            timeout=timeout,
+            default_headers=default_headers,
+            call_metadata=call_metadata,
+            strict=strict,
+            **kwargs,
+        )
+
+        self._previous_response_id = previous_response_id
+
+        # store is set to true if track_previous_responses is true
+        if self.track_previous_responses:
+            self.store = True
+
+        self._http_client = http_client
+        self._async_http_client = async_http_client
+        self._client = openai_client or SyncOpenAI(**self._get_credential_kwargs())
+        self._aclient = async_openai_client or AsyncOpenAI(
+            **self._get_credential_kwargs(is_async=True)
+        )
+
+    @classmethod
+    def class_name(cls) -> str:
+        return "openai_responses_llm"
+
+    @property
+    def metadata(self) -> LLMMetadata:
+        return LLMMetadata(
+            context_window=openai_modelname_to_contextsize(self._get_model_name()),
+            num_output=self.max_output_tokens or -1,
+            is_chat_model=True,
+            is_function_calling_model=is_function_calling_model(
+                model=self._get_model_name()
+            ),
+            model_name=self.model,
+        )
+
+    @property
+    def _tokenizer(self) -> Optional[Tokenizer]:
+        """
+        Get a tokenizer for this model, or None if a tokenizing method is unknown.
+
+        OpenAI can do this using the tiktoken package, subclasses may not have
+        this convenience.
+        """
+        return tiktoken.encoding_for_model(self._get_model_name())
+
+    def _get_model_name(self) -> str:
+        model_name = self.model
+        if "ft-" in model_name:  # legacy fine-tuning
+            model_name = model_name.split(":")[0]
+        elif model_name.startswith("ft:"):
+            model_name = model_name.split(":")[1]
+        return model_name
+
+    def _is_azure_client(self) -> bool:
+        return isinstance(self._get_client(), AzureOpenAI)
+
+    def _get_credential_kwargs(self, is_async: bool = False) -> Dict[str, Any]:
+        return {
+            "api_key": self.api_key,
+            "base_url": self.api_base,
+            "max_retries": self.max_retries,
+            "timeout": self.timeout,
+            "default_headers": self.default_headers,
+            "http_client": self._async_http_client if is_async else self._http_client,
+        }
+
+    def _get_model_kwargs(self, **kwargs: Any) -> Dict[str, Any]:
+        model_kwargs = {
+            "model": self.model,
+            "include": self.include,
+            "instructions": self.instructions,
+            "max_output_tokens": self.max_output_tokens,
+            "metadata": self.call_metadata,
+            "previous_response_id": self._previous_response_id,
+            "store": self.store,
+            "temperature": self.temperature,
+            "tools": self.built_in_tools,
+            "top_p": self.top_p,
+            "truncation": self.truncation,
+            "user": self.user,
+        }
+
+        if self.model in O1_MODELS and self.reasoning_effort is not None:
+            # O1 models support reasoning_effort of low, medium, high
+            model_kwargs["reasoning_effort"] = {"effort": self.reasoning_effort}
+
+        # add tools or extend openai tools
+        if "tools" in kwargs:
+            if isinstance(model_kwargs["tools"], list):
+                model_kwargs["tools"].extend(kwargs.pop("tools"))
+            else:
+                model_kwargs["tools"] = kwargs.pop("tools")
+
+        # priority is class args > additional_kwargs > runtime args
+        model_kwargs.update(self.additional_kwargs)
+
+        kwargs = kwargs or {}
+        model_kwargs.update(kwargs)
+
+        return model_kwargs
+
+    @llm_chat_callback()
+    def chat(self, messages: Sequence[ChatMessage], **kwargs: Any) -> ChatResponse:
+        return self._chat(messages, **kwargs)
+
+    @llm_chat_callback()
+    def stream_chat(
+        self, messages: Sequence[ChatMessage], **kwargs: Any
+    ) -> ChatResponseGen:
+        return self._stream_chat(messages, **kwargs)
+
+    @llm_completion_callback()
+    def complete(
+        self, prompt: str, formatted: bool = False, **kwargs: Any
+    ) -> CompletionResponse:
+        complete_fn = chat_to_completion_decorator(self._chat)
+
+        return complete_fn(prompt, **kwargs)
+
+    @llm_completion_callback()
+    def stream_complete(
+        self, prompt: str, formatted: bool = False, **kwargs: Any
+    ) -> CompletionResponseGen:
+        stream_complete_fn = stream_chat_to_completion_decorator(self._stream_chat)
+
+        return stream_complete_fn(prompt, **kwargs)
+
+    def _parse_response_output(self, output: List[ResponseOutputItem]) -> ChatResponse:
+        message = ChatMessage(role=MessageRole.ASSISTANT, blocks=[])
+        additional_kwargs = {"built_in_tool_calls": []}
+        tool_calls = []
+        for item in output:
+            if isinstance(item, ResponseOutputMessage):
+                blocks = []
+                for part in item.content:
+                    if hasattr(part, "text"):
+                        blocks.append(TextBlock(text=part.text))
+                    if hasattr(part, "annotations"):
+                        additional_kwargs["annotations"] = part.annotations
+                    if hasattr(part, "refusal"):
+                        additional_kwargs["refusal"] = part.refusal
+
+                message.blocks.extend(blocks)
+            elif isinstance(item, ResponseFileSearchToolCall):
+                additional_kwargs["built_in_tool_calls"].append(item)
+            elif isinstance(item, ResponseFunctionToolCall):
+                tool_calls.append(item)
+            elif isinstance(item, ResponseFunctionWebSearch):
+                additional_kwargs["built_in_tool_calls"].append(item)
+            elif isinstance(item, ResponseComputerToolCall):
+                additional_kwargs["built_in_tool_calls"].append(item)
+            elif isinstance(item, ResponseReasoningItem):
+                additional_kwargs["reasoning"] = item
+
+        if tool_calls and message:
+            message.additional_kwargs["tool_calls"] = tool_calls
+
+        return ChatResponse(message=message, additional_kwargs=additional_kwargs)
+
+    @llm_retry_decorator
+    def _chat(self, messages: Sequence[ChatMessage], **kwargs: Any) -> ChatResponse:
+        message_dicts = to_openai_message_dicts(
+            messages,
+            model=self.model,
+            is_responses_api=True,
+        )
+
+        response: Response = self._client.responses.create(
+            input=message_dicts,
+            stream=False,
+            **self._get_model_kwargs(**kwargs),
+        )
+
+        if self.track_previous_responses:
+            self._previous_response_id = response.id
+
+        chat_response = self._parse_response_output(response.output)
+        chat_response.raw = response
+        chat_response.additional_kwargs["usage"] = response.usage
+
+        return chat_response
+
+    @staticmethod
+    def process_response_event(
+        event: ResponseStreamEvent,
+        content: str,
+        tool_calls: List[ResponseFunctionToolCall],
+        built_in_tool_calls: List[Any],
+        additional_kwargs: Dict[str, Any],
+        current_tool_call: Optional[ResponseFunctionToolCall],
+        track_previous_responses: bool,
+        previous_response_id: Optional[str] = None,
+    ) -> Tuple[
+        str,
+        List[ResponseFunctionToolCall],
+        List[Any],
+        Dict[str, Any],
+        Optional[ResponseFunctionToolCall],
+        Optional[str],
+        str,
+    ]:
+        """
+        Process a ResponseStreamEvent and update the state accordingly.
+
+        Args:
+            event: The response stream event to process
+            content: Current accumulated content string
+            tool_calls: List of completed tool calls
+            built_in_tool_calls: List of built-in tool calls
+            additional_kwargs: Additional keyword arguments to include in ChatResponse
+            current_tool_call: The currently in-progress tool call, if any
+            track_previous_responses: Whether to track previous response IDs
+            previous_response_id: Previous response ID if tracking
+
+        Returns:
+            A tuple containing the updated state:
+            (content, tool_calls, built_in_tool_calls, additional_kwargs, current_tool_call, updated_previous_response_id, delta)
+        """
+        delta = ""
+        updated_previous_response_id = previous_response_id
+
+        if isinstance(event, ResponseCreatedEvent) or isinstance(
+            event, ResponseInProgressEvent
+        ):
+            # Initial events, track the response id
+            if track_previous_responses:
+                updated_previous_response_id = event.response.id
+        elif isinstance(event, ResponseOutputItemAddedEvent):
+            # New output item (message, tool call, etc.)
+            if isinstance(event.item, ResponseFunctionToolCall):
+                current_tool_call = event.item
+        elif isinstance(event, ResponseTextDeltaEvent):
+            # Text content is being added
+            delta = event.delta
+            content += delta
+        elif isinstance(event, ResponseFunctionCallArgumentsDeltaEvent):
+            # Function call arguments are being streamed
+            if current_tool_call is not None:
+                current_tool_call.arguments += event.delta
+        elif isinstance(event, ResponseFunctionCallArgumentsDoneEvent):
+            # Function call arguments are complete
+            if current_tool_call is not None:
+                current_tool_call.arguments = event.arguments
+                current_tool_call.status = "completed"
+
+                # append a copy of the tool call to the list
+                tool_calls.append(
+                    ResponseFunctionToolCall(**current_tool_call.model_dump())
+                )
+
+                # clear the current tool call
+                current_tool_call = None
+        elif isinstance(event, ResponseTextAnnotationDeltaEvent):
+            # Annotations for the text
+            annotations = additional_kwargs.get("annotations", [])
+            annotations.append(event.annotation)
+            additional_kwargs["annotations"] = annotations
+        elif isinstance(event, ResponseFileSearchCallCompletedEvent):
+            # File search tool call completed
+            built_in_tool_calls.append(event)
+        elif isinstance(event, ResponseWebSearchCallCompletedEvent):
+            # Web search tool call completed
+            built_in_tool_calls.append(event)
+        elif isinstance(event, ResponseReasoningItem):
+            # Reasoning information
+            additional_kwargs["reasoning"] = event
+        elif isinstance(event, ResponseCompletedEvent):
+            # Response is complete
+            if hasattr(event, "response") and hasattr(event.response, "usage"):
+                additional_kwargs["usage"] = event.response.usage
+
+        return (
+            content,
+            tool_calls,
+            built_in_tool_calls,
+            additional_kwargs,
+            current_tool_call,
+            updated_previous_response_id,
+            delta,
+        )
+
+    @llm_retry_decorator
+    def _stream_chat(
+        self, messages: Sequence[ChatMessage], **kwargs: Any
+    ) -> ChatResponseGen:
+        message_dicts = to_openai_message_dicts(
+            messages,
+            model=self.model,
+            is_responses_api=True,
+        )
+
+        def gen() -> ChatResponseGen:
+            content = ""
+            tool_calls = []
+            built_in_tool_calls = []
+            additional_kwargs = {"built_in_tool_calls": []}
+            current_tool_call: Optional[ResponseFunctionToolCall] = None
+            local_previous_response_id = self._previous_response_id
+
+            for event in self._client.responses.create(
+                input=message_dicts,
+                stream=True,
+                **self._get_model_kwargs(**kwargs),
+            ):
+                # Process the event and update state
+                (
+                    content,
+                    tool_calls,
+                    built_in_tool_calls,
+                    additional_kwargs,
+                    current_tool_call,
+                    local_previous_response_id,
+                    delta,
+                ) = OpenAIResponses.process_response_event(
+                    event=event,
+                    content=content,
+                    tool_calls=tool_calls,
+                    built_in_tool_calls=built_in_tool_calls,
+                    additional_kwargs=additional_kwargs,
+                    current_tool_call=current_tool_call,
+                    track_previous_responses=self.track_previous_responses,
+                    previous_response_id=local_previous_response_id,
+                )
+
+                if (
+                    self.track_previous_responses
+                    and local_previous_response_id != self._previous_response_id
+                ):
+                    self._previous_response_id = local_previous_response_id
+
+                if built_in_tool_calls:
+                    additional_kwargs["built_in_tool_calls"] = built_in_tool_calls
+
+                # For any event, yield a ChatResponse with the current state
+                yield ChatResponse(
+                    message=ChatMessage(
+                        role=MessageRole.ASSISTANT,
+                        content=content,
+                        additional_kwargs={"tool_calls": tool_calls}
+                        if tool_calls
+                        else {},
+                    ),
+                    delta=delta,
+                    raw=event,
+                    additional_kwargs=additional_kwargs,
+                )
+
+        return gen()
+
+    # ===== Async Endpoints =====
+    @llm_chat_callback()
+    async def achat(
+        self,
+        messages: Sequence[ChatMessage],
+        **kwargs: Any,
+    ) -> ChatResponse:
+        return await self._achat(messages, **kwargs)
+
+    @llm_chat_callback()
+    async def astream_chat(
+        self,
+        messages: Sequence[ChatMessage],
+        **kwargs: Any,
+    ) -> ChatResponseAsyncGen:
+        return await self._astream_chat(messages, **kwargs)
+
+    @llm_completion_callback()
+    async def acomplete(
+        self, prompt: str, formatted: bool = False, **kwargs: Any
+    ) -> CompletionResponse:
+        acomplete_fn = achat_to_completion_decorator(self._achat)
+
+        return await acomplete_fn(prompt, **kwargs)
+
+    @llm_completion_callback()
+    async def astream_complete(
+        self, prompt: str, formatted: bool = False, **kwargs: Any
+    ) -> CompletionResponseAsyncGen:
+        astream_complete_fn = astream_chat_to_completion_decorator(self._astream_chat)
+
+        return await astream_complete_fn(prompt, **kwargs)
+
+    @llm_retry_decorator
+    async def _achat(
+        self, messages: Sequence[ChatMessage], **kwargs: Any
+    ) -> ChatResponse:
+        message_dicts = to_openai_message_dicts(
+            messages,
+            model=self.model,
+            is_responses_api=True,
+        )
+
+        response: Response = await self._aclient.responses.create(
+            input=message_dicts,
+            stream=False,
+            **self._get_model_kwargs(**kwargs),
+        )
+
+        if self.track_previous_responses:
+            self._previous_response_id = response.id
+
+        chat_response = self._parse_response_output(response.output)
+        chat_response.raw = response
+        chat_response.additional_kwargs["usage"] = response.usage
+
+        return chat_response
+
+    @llm_retry_decorator
+    async def _astream_chat(
+        self, messages: Sequence[ChatMessage], **kwargs: Any
+    ) -> ChatResponseAsyncGen:
+        message_dicts = to_openai_message_dicts(
+            messages,
+            model=self.model,
+            is_responses_api=True,
+        )
+
+        async def gen() -> ChatResponseAsyncGen:
+            content = ""
+            tool_calls = []
+            built_in_tool_calls = []
+            additional_kwargs = {"built_in_tool_calls": []}
+            current_tool_call: Optional[ResponseFunctionToolCall] = None
+            local_previous_response_id = self._previous_response_id
+
+            response_stream = await self._aclient.responses.create(
+                input=message_dicts,
+                stream=True,
+                **self._get_model_kwargs(**kwargs),
+            )
+
+            async for event in response_stream:
+                # Process the event and update state
+                (
+                    content,
+                    tool_calls,
+                    built_in_tool_calls,
+                    additional_kwargs,
+                    current_tool_call,
+                    local_previous_response_id,
+                    delta,
+                ) = OpenAIResponses.process_response_event(
+                    event=event,
+                    content=content,
+                    tool_calls=tool_calls,
+                    built_in_tool_calls=built_in_tool_calls,
+                    additional_kwargs=additional_kwargs,
+                    current_tool_call=current_tool_call,
+                    track_previous_responses=self.track_previous_responses,
+                    previous_response_id=local_previous_response_id,
+                )
+
+                if (
+                    self.track_previous_responses
+                    and local_previous_response_id != self._previous_response_id
+                ):
+                    self._previous_response_id = local_previous_response_id
+
+                if built_in_tool_calls:
+                    additional_kwargs["built_in_tool_calls"] = built_in_tool_calls
+
+                # For any event, yield a ChatResponse with the current state
+                yield ChatResponse(
+                    message=ChatMessage(
+                        role=MessageRole.ASSISTANT,
+                        content=content,
+                        additional_kwargs={"tool_calls": tool_calls}
+                        if tool_calls
+                        else {},
+                    ),
+                    delta=delta,
+                    raw=event,
+                    additional_kwargs=additional_kwargs,
+                )
+
+        return gen()
+
+    def _prepare_chat_with_tools(
+        self,
+        tools: Sequence["BaseTool"],
+        user_msg: Optional[Union[str, ChatMessage]] = None,
+        chat_history: Optional[List[ChatMessage]] = None,
+        allow_parallel_tool_calls: bool = True,
+        tool_choice: Union[str, dict] = "auto",
+        verbose: bool = False,
+        strict: Optional[bool] = None,
+        **kwargs: Any,
+    ) -> Dict[str, Any]:
+        """Predict and call the tool."""
+
+        # openai responses api has a slightly different tool spec format
+        tool_specs = [
+            {"type": "function", **tool.metadata.to_openai_tool()["function"]}
+            for tool in tools
+        ]
+
+        if strict is not None:
+            strict = strict
+        else:
+            strict = self.strict
+
+        if strict:
+            for tool_spec in tool_specs:
+                tool_spec["strict"] = True
+                tool_spec["parameters"]["additionalProperties"] = False
+
+        if isinstance(user_msg, str):
+            user_msg = ChatMessage(role=MessageRole.USER, content=user_msg)
+
+        messages = chat_history or []
+        if user_msg:
+            messages.append(user_msg)
+
+        return {
+            "messages": messages,
+            "tools": tool_specs or None,
+            "tool_choice": resolve_tool_choice(tool_choice) if tool_specs else None,
+            "parallel_tool_calls": allow_parallel_tool_calls,
+            **kwargs,
+        }
+
+    def get_tool_calls_from_response(
+        self,
+        response: "ChatResponse",
+        error_on_no_tool_call: bool = True,
+        **kwargs: Any,
+    ) -> List[ToolSelection]:
+        """Predict and call the tool."""
+        tool_calls: List[
+            ResponseFunctionToolCall
+        ] = response.message.additional_kwargs.get("tool_calls", [])
+
+        if len(tool_calls) < 1:
+            if error_on_no_tool_call:
+                raise ValueError(
+                    f"Expected at least one tool call, but got {len(tool_calls)} tool calls."
+                )
+            else:
+                return []
+
+        tool_selections = []
+        for tool_call in tool_calls:
+            # this should handle both complete and partial jsons
+            try:
+                argument_dict = parse_partial_json(tool_call.arguments)
+            except ValueError:
+                argument_dict = {}
+
+            tool_selections.append(
+                ToolSelection(
+                    tool_id=tool_call.call_id,
+                    tool_name=tool_call.name,
+                    tool_kwargs=argument_dict,
+                )
+            )
+
+        return tool_selections
+
+    @dispatcher.span
+    def structured_predict(
+        self,
+        output_cls: Type[Model],
+        prompt: PromptTemplate,
+        llm_kwargs: Optional[Dict[str, Any]] = None,
+        **prompt_args: Any,
+    ) -> Model:
+        """Structured predict."""
+        llm_kwargs = llm_kwargs or {}
+
+        llm_kwargs["tool_choice"] = (
+            "required" if "tool_choice" not in llm_kwargs else llm_kwargs["tool_choice"]
+        )
+        # by default structured prediction uses function calling to extract structured outputs
+        # here we force tool_choice to be required
+        return super().structured_predict(
+            output_cls, prompt, llm_kwargs=llm_kwargs, **prompt_args
+        )
+
+    @dispatcher.span
+    async def astructured_predict(
+        self,
+        output_cls: Type[Model],
+        prompt: PromptTemplate,
+        llm_kwargs: Optional[Dict[str, Any]] = None,
+        **prompt_args: Any,
+    ) -> Model:
+        """Structured predict."""
+        llm_kwargs = llm_kwargs or {}
+
+        llm_kwargs["tool_choice"] = (
+            "required" if "tool_choice" not in llm_kwargs else llm_kwargs["tool_choice"]
+        )
+        # by default structured prediction uses function calling to extract structured outputs
+        # here we force tool_choice to be required
+        return await super().astructured_predict(
+            output_cls, prompt, llm_kwargs=llm_kwargs, **prompt_args
+        )
+
+    @dispatcher.span
+    def stream_structured_predict(
+        self,
+        output_cls: Type[Model],
+        prompt: PromptTemplate,
+        llm_kwargs: Optional[Dict[str, Any]] = None,
+        **prompt_args: Any,
+    ) -> Generator[Union[Model, FlexibleModel], None, None]:
+        """Stream structured predict."""
+        llm_kwargs = llm_kwargs or {}
+
+        llm_kwargs["tool_choice"] = (
+            "required" if "tool_choice" not in llm_kwargs else llm_kwargs["tool_choice"]
+        )
+        # by default structured prediction uses function calling to extract structured outputs
+        # here we force tool_choice to be required
+        return super().stream_structured_predict(
+            output_cls, prompt, llm_kwargs=llm_kwargs, **prompt_args
+        )
+
+    @dispatcher.span
+    async def astream_structured_predict(
+        self,
+        output_cls: Type[Model],
+        prompt: PromptTemplate,
+        llm_kwargs: Optional[Dict[str, Any]] = None,
+        **prompt_args: Any,
+    ) -> AsyncGenerator[Union[Model, FlexibleModel], None]:
+        """Stream structured predict."""
+        llm_kwargs = llm_kwargs or {}
+
+        llm_kwargs["tool_choice"] = (
+            "required" if "tool_choice" not in llm_kwargs else llm_kwargs["tool_choice"]
+        )
+        # by default structured prediction uses function calling to extract structured outputs
+        # here we force tool_choice to be required
+        return await super().astream_structured_predict(
+            output_cls, prompt, llm_kwargs=llm_kwargs, **prompt_args
+        )

llama_index/llms/openai/utils.py

@@ -391,20 +391,142 @@ def to_openai_message_dict(
     return message_dict  # type: ignore
 
 
+def to_openai_responses_message_dict(
+    message: ChatMessage,
+    drop_none: bool = False,
+    model: Optional[str] = None,
+) -> Union[Dict[str, Any], List[Dict[str, Any]]]:
+    """Convert a ChatMessage to an OpenAI message dict."""
+    content = []
+    content_txt = ""
+
+    for block in message.blocks:
+        if isinstance(block, TextBlock):
+            content.append({"type": "input_text", "text": block.text})
+            content_txt += block.text
+        elif isinstance(block, ImageBlock):
+            if block.url:
+                content.append(
+                    {
+                        "type": "input_image",
+                        "image_url": str(block.url),
+                        "detail": block.detail or "auto",
+                    }
+                )
+            else:
+                img_bytes = block.resolve_image(as_base64=True).read()
+                img_str = img_bytes.decode("utf-8")
+                content.append(
+                    {
+                        "type": "input_image",
+                        "image_url": f"data:{block.image_mimetype};base64,{img_str}",
+                        "detail": block.detail or "auto",
+                    }
+                )
+        else:
+            msg = f"Unsupported content block type: {type(block).__name__}"
+            raise ValueError(msg)
+
+    # NOTE: Sending a null value (None) for Tool Message to OpenAI will cause error
+    # It's only Allowed to send None if it's an Assistant Message and either a function call or tool calls were performed
+    # Reference: https://platform.openai.com/docs/api-reference/chat/create
+    content_txt = (
+        None
+        if content_txt == ""
+        and message.role == MessageRole.ASSISTANT
+        and (
+            "function_call" in message.additional_kwargs
+            or "tool_calls" in message.additional_kwargs
+        )
+        else content_txt
+    )
+
+    # NOTE: Despite what the openai docs say, if the role is ASSISTANT, SYSTEM
+    # or TOOL, 'content' cannot be a list and must be string instead.
+    # Furthermore, if all blocks are text blocks, we can use the content_txt
+    # as the content. This will avoid breaking openai-like APIs.
+    if message.role.value == "tool":
+        call_id = message.additional_kwargs.get(
+            "tool_call_id", message.additional_kwargs.get("call_id")
+        )
+        if call_id is None:
+            raise ValueError(
+                "tool_call_id or call_id is required in additional_kwargs for tool messages"
+            )
+
+        message_dict = {
+            "type": "function_call_output",
+            "output": content_txt,
+            "call_id": call_id,
+        }
+
+        return message_dict
+    elif "tool_calls" in message.additional_kwargs:
+        message_dicts = [
+            tool_call if isinstance(tool_call, dict) else tool_call.model_dump()
+            for tool_call in message.additional_kwargs["tool_calls"]
+        ]
+
+        return message_dicts
+    else:
+        message_dict = {
+            "role": message.role.value,
+            "content": (
+                content_txt
+                if message.role.value in ("assistant", "system", "developer")
+                or all(isinstance(block, TextBlock) for block in message.blocks)
+                else content
+            ),
+        }
+
+    # TODO: O1 models do not support system prompts
+    if (
+        model is not None
+        and model in O1_MODELS
+        and model not in O1_MODELS_WITHOUT_FUNCTION_CALLING
+    ):
+        if message_dict["role"] == "system":
+            message_dict["role"] = "developer"
+
+    null_keys = [key for key, value in message_dict.items() if value is None]
+    # if drop_none is True, remove keys with None values
+    if drop_none:
+        for key in null_keys:
+            message_dict.pop(key)
+
+    return message_dict  # type: ignore
+
+
 def to_openai_message_dicts(
     messages: Sequence[ChatMessage],
     drop_none: bool = False,
     model: Optional[str] = None,
+    is_responses_api: bool = False,
 ) -> List[ChatCompletionMessageParam]:
     """Convert generic messages to OpenAI message dicts."""
-    return [
-        to_openai_message_dict(
-            message,
-            drop_none=drop_none,
-            model=model,
-        )
-        for message in messages
-    ]
+    if is_responses_api:
+        final_message_dicts = []
+        for message in messages:
+            message_dicts = to_openai_responses_message_dict(
+                message,
+                drop_none=drop_none,
+                model=model,
+            )
+            if isinstance(message_dicts, list):
+                final_message_dicts.extend(message_dicts)
+            else:
+                final_message_dicts.append(message_dicts)
+
+        return final_message_dicts
+    else:
+        return [
+            to_openai_message_dict(
+                message,
+                drop_none=drop_none,
+                model=model,
+            )
+            for message in messages
+        ]
 
 
 def from_openai_message(

{llama_index_llms_openai-0.3.28.dist-info → llama_index_llms_openai-0.3.29.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: llama-index-llms-openai
-Version: 0.3.28
+Version: 0.3.29
 Summary: llama-index llms openai integration
 License: MIT
 Author: llama-index
@@ -12,7 +12,7 @@ Classifier: Programming Language :: Python :: 3.10
 Classifier: Programming Language :: Python :: 3.11
 Classifier: Programming Language :: Python :: 3.12
 Requires-Dist: llama-index-core (>=0.12.17,<0.13.0)
-Requires-Dist: openai (>=1.58.1,<2.0.0)
+Requires-Dist: openai (>=1.66.3,<2.0.0)
 Description-Content-Type: text/markdown
 
 # LlamaIndex Llms Integration: Openai

llama_index_llms_openai-0.3.29.dist-info/RECORD

@@ -0,0 +1,9 @@
+llama_index/llms/openai/__init__.py,sha256=8nmgixeXifQ4eVSgtCic54WxXqrrpXQPL4rhACWCSFs,229
+llama_index/llms/openai/base.py,sha256=RjkISrh-RvbrQWOfdNdH4nimDQN0byUFm_n6r703jdM,38609
+llama_index/llms/openai/py.typed,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+llama_index/llms/openai/responses.py,sha256=TikqnpW-UQmgjmMYGznsBx7eEo5prNIaxE81RO1ZZjE,34465
+llama_index/llms/openai/utils.py,sha256=qp9qpXY7HbUnUsVDx6TgK98feibzTRi-bLdq_F3S0fo,26017
+llama_index_llms_openai-0.3.29.dist-info/LICENSE,sha256=JPQLUZD9rKvCTdu192Nk0V5PAwklIg6jANii3UmTyMs,1065
+llama_index_llms_openai-0.3.29.dist-info/METADATA,sha256=g0FtAtfto485JxdoqCDwulLfoKYoPcglr1ETMZ_lFjg,3322
+llama_index_llms_openai-0.3.29.dist-info/WHEEL,sha256=sP946D7jFCHeNz5Iq4fL4Lu-PrWrFsgfLXbbkciIZwg,88
+llama_index_llms_openai-0.3.29.dist-info/RECORD,,

llama_index_llms_openai-0.3.28.dist-info/RECORD

@@ -1,8 +0,0 @@
-llama_index/llms/openai/__init__.py,sha256=vm3cIBSGkBFlE77GyfyN0EhpJcnJZN95QMhPN53EkbE,148
-llama_index/llms/openai/base.py,sha256=RjkISrh-RvbrQWOfdNdH4nimDQN0byUFm_n6r703jdM,38609
-llama_index/llms/openai/py.typed,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-llama_index/llms/openai/utils.py,sha256=n7GEv864j34idRUR2ouu0McSBqBTVX2Tko9vf1YOl-k,21624
-llama_index_llms_openai-0.3.28.dist-info/LICENSE,sha256=JPQLUZD9rKvCTdu192Nk0V5PAwklIg6jANii3UmTyMs,1065
-llama_index_llms_openai-0.3.28.dist-info/METADATA,sha256=Bmx7FvGOcpdMpwP5gx-Wx2Dlbkp_42N7QC-zVodKXuw,3322
-llama_index_llms_openai-0.3.28.dist-info/WHEEL,sha256=sP946D7jFCHeNz5Iq4fL4Lu-PrWrFsgfLXbbkciIZwg,88
-llama_index_llms_openai-0.3.28.dist-info/RECORD,,