gemini.extractors

BaseExtractor

Bases: BasePrompt, Generic[BaseCallT, BaseToolT, BaseToolStreamT, ExtractedTypeT], ABC

The base abstract interface for extracting structured information using LLMs.

Source code in mirascope/base/extractors.py

class BaseExtractor(
    BasePrompt, Generic[BaseCallT, BaseToolT, BaseToolStreamT, ExtractedTypeT], ABC
):
    """The base abstract interface for extracting structured information using LLMs."""

    extract_schema: ExtractionType

    api_key: ClassVar[Optional[str]] = None
    base_url: ClassVar[Optional[str]] = None
    call_params: ClassVar[BaseCallParams] = BaseCallParams[BaseToolT](
        model="gpt-3.5-turbo-0125"
    )
    configuration: ClassVar[BaseConfig] = BaseConfig(llm_ops=[])

    @abstractmethod
    def extract(self, retries: int = 0) -> ExtractedTypeT:
        """Extracts the `extraction_schema` from an LLM call."""
        ...  # pragma: no cover

    @abstractmethod
    async def extract_async(self, retries: int = 0) -> ExtractedTypeT:
        """Asynchronously extracts the `extraction_schema` from an LLM call."""
        ...  # pragma: no cover

    # Note: only some model providers support streaming tools, so we only implement
    # streaming for those providers and do not require all extractors to implement
    # the `stream` and `stream_async` methods.
    # @abstractmethod
    # def stream(self, retries: int = 0) -> Generator[ExtractedTypeT, None, None]:
    #     """Streams extracted partial `extraction_schema` instances."""
    #     ...  # pragma: no cover

    # @abstractmethod
    # async def stream_async(
    #     self, retries: int = 0
    # ) -> AsyncGenerator[ExtractedTypeT, None]:
    #     """Asynchronously streams extracted partial `extraction_schema` instances."""
    #     ...  # pragma: no cover

    @classmethod
    def from_prompt(
        cls,
        prompt_type: type[BasePromptT],
        call_params: BaseCallParams,
        *,
        extract_schema: Optional[ExtractedType] = None,
    ) -> type[BasePromptT]:
        """Returns an extractor_type generated dynamically from this base extractor.

        Args:
            prompt_type: The prompt class to use for the extractor. Properties and class
                variables of this class will be used to create the new extractor class.
                Must be a class that can be instantiated.
            call_params: The call params to use for the extractor.
            extract_schema: The extract schema to use for the extractor. If none, the
                extractor will use the class' extract_schema.

        Returns:
            A new extractor class with new extractor type.
        """

        fields: dict[str, Any] = {
            name: (field.annotation, field.default)
            for name, field in prompt_type.model_fields.items()
        }

        if extract_schema is not None:
            fields["extract_schema"] = (type[extract_schema], extract_schema)  # type: ignore
        else:
            extract_schema = fields["extract_schema"][1]

        class_vars = {
            name: value
            for name, value in prompt_type.__dict__.items()
            if name not in prompt_type.model_fields
        }
        new_extractor = create_model(
            prompt_type.__name__,
            __base__=cls[extract_schema],  # type: ignore
            **fields,
        )

        for var_name, var_value in class_vars.items():
            setattr(new_extractor, var_name, var_value)
        setattr(new_extractor, "call_params", call_params)

        return cast(type[BasePromptT], new_extractor)

    ############################## PRIVATE METHODS ###################################

    def _extract(
        self,
        call_type: Type[BaseCallT],
        tool_type: Type[BaseToolT],
        retries: Union[int, Retrying] = 0,
        **kwargs: Any,
    ) -> ExtractedTypeT:
        """Extracts `extract_schema` from the call response.

        The `extract_schema` is converted into a tool, complete with a description of
        the tool, all of the fields, and their types. This allows us to take advantage
        of tools/function calling functionality to extract information from a prompt
        according to the context provided by the `BaseModel` schema.

        Args:
            call_type: The type of call to use for extraction. This enables shared code
                across various model providers that have slight variations but the same
                internal interfaces.
            tool_type: The type of tool to use for extraction. This enables shared code
                across various model providers that have slight variations but the same
                internal interfaces.
            retries: The number of call attempts to make on `ValidationError` before
                giving up and throwing the error to the user.
            **kwargs: Additional keyword arguments.

        Returns:
            An instance of the `extract_schema` with it's fields populated.

        Raises:
            AttributeError: if there is no tool in the call creation.
            ValidationError: if the schema cannot be instantiated from the completion.
        """

        def _extract_attempt(
            call_type: Type[BaseCallT],
            tool_type: Type[BaseToolT],
            error_messages: dict[str, Any],
            **kwargs: Any,
        ) -> ExtractedTypeT:
            kwargs, return_tool = self._setup(tool_type, kwargs)

            temp_call = self._generate_temp_call(call_type, error_messages)
            response = temp_call(
                **self.model_dump(exclude={"extract_schema"}),
            ).call(**kwargs)
            try:
                extracted_schema = self._extract_schema(
                    response.tool, self.extract_schema, return_tool, response=response
                )
                if extracted_schema is None:
                    raise AttributeError("No tool found in the completion.")
                return extracted_schema
            except (AttributeError, ValueError, ValidationError):
                raise

        if isinstance(retries, int):
            if retries > 0:
                retries = Retrying(stop=stop_after_attempt(retries))
            else:
                return _extract_attempt(call_type, tool_type, {}, **kwargs)
        try:
            error_messages: dict[str, Any] = {}
            for attempt in retries:
                with attempt:
                    try:
                        extraction = _extract_attempt(
                            call_type, tool_type, error_messages, **kwargs
                        )
                    except (AttributeError, ValueError, ValidationError) as e:
                        error_messages[str(e)] = None
                        if "logfire" in self.configuration.llm_ops:  # pragma: no cover
                            logfire.error(f"Retrying due to exception: {e}")
                        raise
        except RetryError as e:
            raise e
        return extraction

    async def _extract_async(
        self,
        call_type: Type[BaseCallT],
        tool_type: Type[BaseToolT],
        retries: Union[int, AsyncRetrying],
        **kwargs: Any,
    ) -> ExtractedTypeT:
        """Extracts `extract_schema` from the asynchronous call response.

        The `extract_schema` is converted into a tool, complete with a description of
        the tool, all of the fields, and their types. This allows us to take advantage
        of tools/function calling functionality to extract information from a prompt
        according to the context provided by the `BaseModel` schema.

        Args:
            call_type: The type of call to use for extraction. This enables shared code
                across various model providers that have slight variations but the same
                internal interfaces.
            tool_type: The type of tool to use for extraction. This enables shared code
                across various model providers that have slight variations but the same
                internal interfaces.
            retries: The number of call attempts to make on `ValidationError` before
                giving up and throwing the error to the user.
            **kwargs: Additional keyword arguments.

        Returns:
            An instance of the `extract_schema` with it's fields populated.

        Raises:
            AttributeError: if there is no tool in the call creation.
            ValidationError: if the schema cannot be instantiated from the completion.
        """

        async def _extract_attempt_async(
            call_type: Type[BaseCallT],
            tool_type: Type[BaseToolT],
            error_messages: dict[str, Any],
            **kwargs: Any,
        ) -> ExtractedTypeT:
            kwargs, return_tool = self._setup(tool_type, kwargs)

            temp_call = self._generate_temp_call(call_type, error_messages)

            response = await temp_call(
                **self.model_dump(exclude={"extract_schema"})
            ).call_async(**kwargs)
            try:
                extracted_schema = self._extract_schema(
                    response.tool, self.extract_schema, return_tool, response=response
                )
                if extracted_schema is None:
                    raise AttributeError("No tool found in the completion.")
                return extracted_schema
            except (AttributeError, ValueError, ValidationError):
                raise

        if isinstance(retries, int):
            if retries > 0:
                retries = AsyncRetrying(stop=stop_after_attempt(retries))
            else:
                return await _extract_attempt_async(call_type, tool_type, {}, **kwargs)
        try:
            error_messages: dict[str, Any] = {}
            async for attempt in retries:
                with attempt:
                    try:
                        extraction = await _extract_attempt_async(
                            call_type, tool_type, error_messages, **kwargs
                        )
                    except (AttributeError, ValueError, ValidationError) as e:
                        error_messages[str(e)] = None
                        if "logfire" in self.configuration.llm_ops:  # pragma: no cover
                            logfire.error(f"Retrying due to exception: {e}")
                        raise
        except RetryError as e:
            raise e
        return extraction

    def _stream(
        self,
        call_type: Type[BaseCallT],
        tool_type: Type[BaseToolT],
        tool_stream_type: Type[BaseToolStreamT],
        retries: Union[int, Retrying],
        **kwargs: Any,
    ) -> Generator[ExtractedTypeT, None, None]:
        """Streams partial `extract_schema` instances from the streamed chunks.

        The `extract_schema` is converted into a partial tool, complete with a
        description of the tool, all of the fields, and their types. This allows us to
        take advantage of tools/function calling functionality to stream information
        extracted from a prompt according to the context provided by the `BaseModel`
        schema.

        Args:
            call_type: The type of call to use for extraction. This enables shared code
                across various model providers that have slight variations but the same
                internal interfaces.
            tool_type: The type of tool to use for extraction. This enables shared code
                across various model providers that have slight variations but the same
                internal interfaces.
            tool_stream_type: The type of tool stream to use for streaming tools. This
                enables shared code across various model providers that have slight
                variations but the same internal interfaces.
            retries: The number of call attempts to make on `ValidationError` before
                giving up and throwing the error to the user.
            **kwargs: Additional keyword arguments.

        Yields:
            An instance of the partial `extract_schema` with it's available fields
            populated.

        Raises:
            AttributeError: if there is no tool in the call creation.
            ValidationError: if the schema cannot be instantiated from the completion.
        """

        def _stream_attempt(
            call_type: Type[BaseCallT],
            tool_type: Type[BaseToolT],
            tool_stream_type: Type[BaseToolStreamT],
            error_messages: dict[str, Any],
            **kwargs: Any,
        ) -> Generator[ExtractedTypeT, None, None]:
            kwargs, return_tool = self._setup(tool_type, kwargs)

            temp_call = self._generate_temp_call(call_type, error_messages)

            stream = temp_call(**self.model_dump(exclude={"extract_schema"})).stream(
                **kwargs
            )
            tool_stream = tool_stream_type.from_stream(stream, allow_partial=True)
            try:
                yielded = False
                for partial_tool in tool_stream:
                    extracted_schema = self._extract_schema(
                        partial_tool, self.extract_schema, return_tool, response=None
                    )
                    if extracted_schema is None:
                        break
                    yielded = True
                    yield extracted_schema

                if not yielded:
                    raise AttributeError("No tool found in the completion.")
            except (AttributeError, ValueError, ValidationError):
                raise

        if isinstance(retries, int):
            if retries > 0:
                retries = Retrying(stop=stop_after_attempt(retries))
            else:
                for partial_tool in _stream_attempt(
                    call_type,
                    tool_type,
                    tool_stream_type,
                    {},
                    **kwargs,
                ):
                    yield partial_tool
                return
        try:
            error_messages: dict[str, Any] = {}
            for attempt in retries:
                with attempt:
                    try:
                        for partial_tool in _stream_attempt(
                            call_type,
                            tool_type,
                            tool_stream_type,
                            error_messages,
                            **kwargs,
                        ):
                            yield partial_tool
                    except (AttributeError, ValueError, ValidationError) as e:
                        error_messages[str(e)] = None
                        if "logfire" in self.configuration.llm_ops:  # pragma: no cover
                            logfire.error(f"Retrying due to exception: {e}")
                        raise
        except RetryError as e:
            raise e

    async def _stream_async(
        self,
        call_type: Type[BaseCallT],
        tool_type: Type[BaseToolT],
        tool_stream_type: Type[BaseToolStreamT],
        retries: Union[int, AsyncRetrying],
        **kwargs: Any,
    ) -> AsyncGenerator[ExtractedTypeT, None]:
        """Asynchronously streams partial `extract_schema`s from streamed chunks.

        The `extract_schema` is converted into a partial tool, complete with a
        description of the tool, all of the fields, and their types. This allows us to
        take advantage of tools/function calling functionality to stream information
        extracted from a prompt according to the context provided by the `BaseModel`
        schema.

        Args:
            call_type: The type of call to use for extraction. This enables shared code
                across various model providers that have slight variations but the same
                internal interfaces.
            tool_type: The type of tool to use for extraction. This enables shared code
                across various model providers that have slight variations but the same
                internal interfaces.
            tool_stream_type: The type of tool stream to use for streaming tools. This
                enables shared code across various model providers that have slight
                variations but the same internal interfaces.
            retries: The number of call attempts to make on `ValidationError` before
                giving up and throwing the error to the user.
            **kwargs: Additional keyword arguments.

        Yields:
            An instance of the partial `extract_schema` with it's available fields
            populated.

        Raises:
            AttributeError: if there is no tool in the call creation.
            ValidationError: if the schema cannot be instantiated from the completion.
        """

        async def _stream_attempt_async(
            call_type: Type[BaseCallT],
            tool_type: Type[BaseToolT],
            tool_stream_type: Type[BaseToolStreamT],
            error_messages: dict[str, Any],
            **kwargs: Any,
        ) -> AsyncGenerator[ExtractedTypeT, None]:
            kwargs, return_tool = self._setup(tool_type, kwargs)

            temp_call = self._generate_temp_call(call_type, error_messages)

            stream = temp_call(
                **self.model_dump(exclude={"extract_schema"})
            ).stream_async(**kwargs)
            tool_stream = tool_stream_type.from_async_stream(stream, allow_partial=True)
            try:
                yielded = False
                async for partial_tool in tool_stream:
                    extracted_schema = self._extract_schema(
                        partial_tool, self.extract_schema, return_tool, response=None
                    )
                    if extracted_schema is None:
                        break
                    yielded = True
                    yield extracted_schema

                if not yielded:
                    raise AttributeError("No tool found in the completion.")
            except (AttributeError, ValueError, ValidationError):
                raise

        if isinstance(retries, int):
            if retries > 0:
                retries = AsyncRetrying(stop=stop_after_attempt(retries))
            else:
                async for partial_tool in _stream_attempt_async(
                    call_type, tool_type, tool_stream_type, {}, **kwargs
                ):
                    yield partial_tool
                return
        try:
            error_messages: dict[str, Any] = {}
            async for attempt in retries:
                with attempt:
                    try:
                        async for partial_tool in _stream_attempt_async(
                            call_type,
                            tool_type,
                            tool_stream_type,
                            error_messages,
                            **kwargs,
                        ):
                            yield partial_tool
                    except (AttributeError, ValueError, ValidationError) as e:
                        error_messages[str(e)] = None
                        if "logfire" in self.configuration.llm_ops:  # pragma: no cover
                            logfire.error(f"Retrying due to exception: {e}")
                        raise
        except RetryError as e:
            raise e

    def _generate_temp_call(
        self, call_type: Type[BaseCallT], error_messages: dict[str, Any]
    ) -> Type[BaseCallT]:
        """Returns a `TempCall` generated using the extractors definition."""
        _prompt_template = self.prompt_template
        if error_messages:
            formatted_error_messages = [
                "- " + element for element in error_messages.keys()
            ]
            error_messages_list = "\n".join(formatted_error_messages)
            _prompt_template = (
                f"{_prompt_template}\n"
                "Errors found:\n\n"
                f"{error_messages_list}\n\n"
                "Please fix the errors and try again."
            )

        class TempCall(call_type):  # type: ignore
            prompt_template = _prompt_template

            base_url = self.base_url
            api_key = self.api_key
            call_params = self.call_params
            configuration = self.configuration

            model_config = ConfigDict(extra="allow")

        properties = getmembers(self)
        for name, value in properties:
            if not hasattr(TempCall, name) or (
                name == "messages" and "messages" in self.__class__.__dict__
            ):
                setattr(TempCall, name, value)

        return TempCall

    def _extract_schema(
        self,
        tool: Optional[BaseToolT],
        schema: ExtractedType,
        return_tool: bool,
        response: Optional[Any],
    ) -> Optional[ExtractedTypeT]:
        """Returns the extracted schema extracted depending on it's extraction type.

        Due to mypy issues with all these generics, we have to type ignore a bunch
        of stuff so it doesn't complain, but each conditional properly checks types
        before doing anything specific to that type (it's just that mypy is annoying).
        """
        if tool is None:
            return None
        if return_tool:
            return tool  # type: ignore
        if _is_base_type(schema):
            return tool.value  # type: ignore
        if response:
            model = schema(**tool.model_dump())  # type: ignore
            model._response = response
        else:
            schema = partial(schema)  # type: ignore
            model = schema(**tool.model_dump())
            model._tool_call = tool.tool_call  # type: ignore
        return model

    def _setup(
        self, tool_type: Type[BaseToolT], kwargs: dict[str, Any]
    ) -> tuple[dict[str, Any], bool]:
        """Returns the call params kwargs and whether to return the tool directly."""
        call_params = self.call_params.model_copy(update=kwargs)
        kwargs = call_params.kwargs(tool_type=tool_type)
        if _is_base_type(self.extract_schema):
            tool = tool_type.from_base_type(self.extract_schema)  # type: ignore
            return_tool = False
        elif not isclass(self.extract_schema):
            tool = tool_type.from_fn(self.extract_schema)
            return_tool = True
        elif not issubclass(self.extract_schema, tool_type):
            tool = tool_type.from_model(self.extract_schema)
            return_tool = False
        else:
            tool = self.extract_schema
            return_tool = True
        kwargs["tools"] = [tool]
        return kwargs, return_tool

extract(retries=0) abstractmethod

Extracts the extraction_schema from an LLM call.

Source code in mirascope/base/extractors.py

@abstractmethod
def extract(self, retries: int = 0) -> ExtractedTypeT:
    """Extracts the `extraction_schema` from an LLM call."""
    ...  # pragma: no cover

extract_async(retries=0) abstractmethod async

Asynchronously extracts the extraction_schema from an LLM call.

Source code in mirascope/base/extractors.py

@abstractmethod
async def extract_async(self, retries: int = 0) -> ExtractedTypeT:
    """Asynchronously extracts the `extraction_schema` from an LLM call."""
    ...  # pragma: no cover

from_prompt(prompt_type, call_params, *, extract_schema=None) classmethod

Returns an extractor_type generated dynamically from this base extractor.

Parameters:

Name	Type	Description	Default
prompt_type	type[BasePromptT]	The prompt class to use for the extractor. Properties and class variables of this class will be used to create the new extractor class. Must be a class that can be instantiated.	required
call_params	BaseCallParams	The call params to use for the extractor.	required
extract_schema	Optional[ExtractedType]	The extract schema to use for the extractor. If none, the extractor will use the class' extract_schema.	None

Returns:

Type	Description
type[BasePromptT]	A new extractor class with new extractor type.

Source code in mirascope/base/extractors.py

@classmethod
def from_prompt(
    cls,
    prompt_type: type[BasePromptT],
    call_params: BaseCallParams,
    *,
    extract_schema: Optional[ExtractedType] = None,
) -> type[BasePromptT]:
    """Returns an extractor_type generated dynamically from this base extractor.

    Args:
        prompt_type: The prompt class to use for the extractor. Properties and class
            variables of this class will be used to create the new extractor class.
            Must be a class that can be instantiated.
        call_params: The call params to use for the extractor.
        extract_schema: The extract schema to use for the extractor. If none, the
            extractor will use the class' extract_schema.

    Returns:
        A new extractor class with new extractor type.
    """

    fields: dict[str, Any] = {
        name: (field.annotation, field.default)
        for name, field in prompt_type.model_fields.items()
    }

    if extract_schema is not None:
        fields["extract_schema"] = (type[extract_schema], extract_schema)  # type: ignore
    else:
        extract_schema = fields["extract_schema"][1]

    class_vars = {
        name: value
        for name, value in prompt_type.__dict__.items()
        if name not in prompt_type.model_fields
    }
    new_extractor = create_model(
        prompt_type.__name__,
        __base__=cls[extract_schema],  # type: ignore
        **fields,
    )

    for var_name, var_value in class_vars.items():
        setattr(new_extractor, var_name, var_value)
    setattr(new_extractor, "call_params", call_params)

    return cast(type[BasePromptT], new_extractor)

GeminiCall

Bases: BaseCall[GeminiCallResponse, GeminiCallResponseChunk, GeminiTool, ContentDict]

A class for prompting Google's Gemini Chat API.

This prompt supports the message types: USER, MODEL, TOOL

Example:

from google.generativeai import configure  # type: ignore
from mirascope.gemini import GeminiCall

configure(api_key="YOUR_API_KEY")


class BookRecommender(GeminiCall):
    prompt_template = """
    USER: You're the world's greatest librarian.
    MODEL: Ok, I understand I'm the world's greatest librarian. How can I help?
    USER: Please recommend some {genre} books.

    genre: str


response = BookRecommender(genre="fantasy").call()
print(response.content)
#> As the world's greatest librarian, I am delighted to recommend...

Source code in mirascope/gemini/calls.py

class GeminiCall(
    BaseCall[GeminiCallResponse, GeminiCallResponseChunk, GeminiTool, ContentDict]
):
    '''A class for prompting Google's Gemini Chat API.

    This prompt supports the message types: USER, MODEL, TOOL

    Example:

    ```python
    from google.generativeai import configure  # type: ignore
    from mirascope.gemini import GeminiCall

    configure(api_key="YOUR_API_KEY")


    class BookRecommender(GeminiCall):
        prompt_template = """
        USER: You're the world's greatest librarian.
        MODEL: Ok, I understand I'm the world's greatest librarian. How can I help?
        USER: Please recommend some {genre} books.

        genre: str


    response = BookRecommender(genre="fantasy").call()
    print(response.content)
    #> As the world's greatest librarian, I am delighted to recommend...
    ```
    '''

    call_params: ClassVar[GeminiCallParams] = GeminiCallParams()
    _provider: ClassVar[str] = "gemini"

    def messages(self) -> ContentsType:
        """Returns the `ContentsType` messages for Gemini `generate_content`.

        Raises:
            ValueError: if the docstring contains an unknown role.
        """
        return [
            {"role": message["role"], "parts": [message["content"]]}
            for message in self._parse_messages(
                [MessageRole.MODEL, MessageRole.USER, MessageRole.TOOL]
            )
        ]

    @retry
    def call(
        self, retries: Union[int, Retrying] = 0, **kwargs: Any
    ) -> GeminiCallResponse:
        """Makes an call to the model using this `GeminiCall` instance.

        Args:
            **kwargs: Additional keyword arguments that will be used for generating the
                response. These will override any existing argument settings in call
                params.

        Returns:
            A `GeminiCallResponse` instance.
        """
        kwargs, tool_types = self._setup(kwargs, GeminiTool)
        model_name = kwargs.pop("model")
        gemini_pro_model = get_wrapped_client(
            GenerativeModel(model_name=model_name), self
        )
        generate_content = get_wrapped_call(
            gemini_pro_model.generate_content,
            self,
            response_type=GeminiCallResponse,
            tool_types=tool_types,
            model_name=model_name,
        )
        messages = self.messages()
        user_message_param = self._get_possible_user_message(messages)
        start_time = datetime.datetime.now().timestamp() * 1000
        response = generate_content(
            messages,
            stream=False,
            tools=kwargs.pop("tools") if "tools" in kwargs else None,
            **kwargs,
        )
        return GeminiCallResponse(
            response=response,
            user_message_param=user_message_param,
            tool_types=tool_types,
            start_time=start_time,
            end_time=datetime.datetime.now().timestamp() * 1000,
            cost=None,
        )

    @retry
    async def call_async(
        self, retries: Union[int, AsyncRetrying] = 0, **kwargs: Any
    ) -> GeminiCallResponse:
        """Makes an asynchronous call to the model using this `GeminiCall` instance.

        Args:
            **kwargs: Additional keyword arguments that will be used for generating the
                response. These will override any existing argument settings in call
                params.

        Returns:
            A `GeminiCallResponse` instance.
        """
        kwargs, tool_types = self._setup(kwargs, GeminiTool)
        model_name = kwargs.pop("model")
        gemini_pro_model = get_wrapped_async_client(
            GenerativeModel(model_name=model_name), self
        )
        generate_content_async = get_wrapped_call(
            gemini_pro_model.generate_content_async,
            self,
            is_async=True,
            response_type=GeminiCallResponse,
            tool_types=tool_types,
            model_name=model_name,
        )
        messages = self.messages()
        user_message_param = self._get_possible_user_message(messages)
        start_time = datetime.datetime.now().timestamp() * 1000
        response = await generate_content_async(
            messages,
            stream=False,
            tools=kwargs.pop("tools") if "tools" in kwargs else None,
            **kwargs,
        )
        return GeminiCallResponse(
            response=response,
            user_message_param=user_message_param,
            tool_types=tool_types,
            start_time=start_time,
            end_time=datetime.datetime.now().timestamp() * 1000,
            cost=None,
        )

    @retry
    def stream(
        self, retries: Union[int, Retrying] = 0, **kwargs: Any
    ) -> Generator[GeminiCallResponseChunk, None, None]:
        """Streams the response for a call using this `GeminiCall`.

        Args:
            **kwargs: Additional keyword arguments parameters to pass to the call. These
                will override any existing arguments in `call_params`.

        Yields:
            A `GeminiCallResponseChunk` for each chunk of the response.
        """
        kwargs, tool_types = self._setup(kwargs, GeminiTool)
        model_name = kwargs.pop("model")
        gemini_pro_model = get_wrapped_client(
            GenerativeModel(model_name=model_name), self
        )
        generate_content = get_wrapped_call(
            gemini_pro_model.generate_content,
            self,
            response_chunk_type=GeminiCallResponseChunk,
            tool_types=tool_types,
            model_name=model_name,
        )
        messages = self.messages()
        user_message_param = self._get_possible_user_message(messages)
        stream = generate_content(
            messages,
            stream=True,
            tools=kwargs.pop("tools") if "tools" in kwargs else None,
            **kwargs,
        )
        for chunk in stream:
            yield GeminiCallResponseChunk(
                chunk=chunk,
                user_message_param=user_message_param,
                tool_types=tool_types,
            )

    @retry
    async def stream_async(
        self, retries: Union[int, AsyncRetrying] = 0, **kwargs: Any
    ) -> AsyncGenerator[GeminiCallResponseChunk, None]:
        """Streams the response asynchronously for a call using this `GeminiCall`.

        Args:
            **kwargs: Additional keyword arguments parameters to pass to the call. These
                will override any existing arguments in `call_params`.

        Yields:
            A `GeminiCallResponseChunk` for each chunk of the response.
        """
        kwargs, tool_types = self._setup(kwargs, GeminiTool)
        model_name = kwargs.pop("model")
        gemini_pro_model = get_wrapped_async_client(
            GenerativeModel(model_name=model_name), self
        )
        generate_content_async = get_wrapped_call(
            gemini_pro_model.generate_content_async,
            self,
            is_async=True,
            response_chunk_type=GeminiCallResponseChunk,
            tool_types=tool_types,
            model_name=model_name,
        )
        messages = self.messages()
        user_message_param = self._get_possible_user_message(messages)
        stream = generate_content_async(
            messages,
            stream=True,
            tools=kwargs.pop("tools") if "tools" in kwargs else None,
            **kwargs,
        )
        if inspect.iscoroutine(stream):
            stream = await stream
        async for chunk in stream:
            yield GeminiCallResponseChunk(
                chunk=chunk,
                user_message_param=user_message_param,
                tool_types=tool_types,
            )

call(retries=0, **kwargs)

Makes an call to the model using this GeminiCall instance.

Parameters:

Name	Type	Description	Default
**kwargs	Any	Additional keyword arguments that will be used for generating the response. These will override any existing argument settings in call params.	{}

Returns:

Type	Description
GeminiCallResponse	A GeminiCallResponse instance.

Source code in mirascope/gemini/calls.py

@retry
def call(
    self, retries: Union[int, Retrying] = 0, **kwargs: Any
) -> GeminiCallResponse:
    """Makes an call to the model using this `GeminiCall` instance.

    Args:
        **kwargs: Additional keyword arguments that will be used for generating the
            response. These will override any existing argument settings in call
            params.

    Returns:
        A `GeminiCallResponse` instance.
    """
    kwargs, tool_types = self._setup(kwargs, GeminiTool)
    model_name = kwargs.pop("model")
    gemini_pro_model = get_wrapped_client(
        GenerativeModel(model_name=model_name), self
    )
    generate_content = get_wrapped_call(
        gemini_pro_model.generate_content,
        self,
        response_type=GeminiCallResponse,
        tool_types=tool_types,
        model_name=model_name,
    )
    messages = self.messages()
    user_message_param = self._get_possible_user_message(messages)
    start_time = datetime.datetime.now().timestamp() * 1000
    response = generate_content(
        messages,
        stream=False,
        tools=kwargs.pop("tools") if "tools" in kwargs else None,
        **kwargs,
    )
    return GeminiCallResponse(
        response=response,
        user_message_param=user_message_param,
        tool_types=tool_types,
        start_time=start_time,
        end_time=datetime.datetime.now().timestamp() * 1000,
        cost=None,
    )

call_async(retries=0, **kwargs) async

Makes an asynchronous call to the model using this GeminiCall instance.

Parameters:

Name	Type	Description	Default
**kwargs	Any	Additional keyword arguments that will be used for generating the response. These will override any existing argument settings in call params.	{}

Returns:

Type	Description
GeminiCallResponse	A GeminiCallResponse instance.

Source code in mirascope/gemini/calls.py

@retry
async def call_async(
    self, retries: Union[int, AsyncRetrying] = 0, **kwargs: Any
) -> GeminiCallResponse:
    """Makes an asynchronous call to the model using this `GeminiCall` instance.

    Args:
        **kwargs: Additional keyword arguments that will be used for generating the
            response. These will override any existing argument settings in call
            params.

    Returns:
        A `GeminiCallResponse` instance.
    """
    kwargs, tool_types = self._setup(kwargs, GeminiTool)
    model_name = kwargs.pop("model")
    gemini_pro_model = get_wrapped_async_client(
        GenerativeModel(model_name=model_name), self
    )
    generate_content_async = get_wrapped_call(
        gemini_pro_model.generate_content_async,
        self,
        is_async=True,
        response_type=GeminiCallResponse,
        tool_types=tool_types,
        model_name=model_name,
    )
    messages = self.messages()
    user_message_param = self._get_possible_user_message(messages)
    start_time = datetime.datetime.now().timestamp() * 1000
    response = await generate_content_async(
        messages,
        stream=False,
        tools=kwargs.pop("tools") if "tools" in kwargs else None,
        **kwargs,
    )
    return GeminiCallResponse(
        response=response,
        user_message_param=user_message_param,
        tool_types=tool_types,
        start_time=start_time,
        end_time=datetime.datetime.now().timestamp() * 1000,
        cost=None,
    )

messages()

Returns the ContentsType messages for Gemini generate_content.

Raises:

Type	Description
ValueError	if the docstring contains an unknown role.

Source code in mirascope/gemini/calls.py

def messages(self) -> ContentsType:
    """Returns the `ContentsType` messages for Gemini `generate_content`.

    Raises:
        ValueError: if the docstring contains an unknown role.
    """
    return [
        {"role": message["role"], "parts": [message["content"]]}
        for message in self._parse_messages(
            [MessageRole.MODEL, MessageRole.USER, MessageRole.TOOL]
        )
    ]

stream(retries=0, **kwargs)

Streams the response for a call using this GeminiCall.

Parameters:

Name	Type	Description	Default
**kwargs	Any	Additional keyword arguments parameters to pass to the call. These will override any existing arguments in call_params.	{}

Yields:

Type	Description
GeminiCallResponseChunk	A GeminiCallResponseChunk for each chunk of the response.

Source code in mirascope/gemini/calls.py

@retry
def stream(
    self, retries: Union[int, Retrying] = 0, **kwargs: Any
) -> Generator[GeminiCallResponseChunk, None, None]:
    """Streams the response for a call using this `GeminiCall`.

    Args:
        **kwargs: Additional keyword arguments parameters to pass to the call. These
            will override any existing arguments in `call_params`.

    Yields:
        A `GeminiCallResponseChunk` for each chunk of the response.
    """
    kwargs, tool_types = self._setup(kwargs, GeminiTool)
    model_name = kwargs.pop("model")
    gemini_pro_model = get_wrapped_client(
        GenerativeModel(model_name=model_name), self
    )
    generate_content = get_wrapped_call(
        gemini_pro_model.generate_content,
        self,
        response_chunk_type=GeminiCallResponseChunk,
        tool_types=tool_types,
        model_name=model_name,
    )
    messages = self.messages()
    user_message_param = self._get_possible_user_message(messages)
    stream = generate_content(
        messages,
        stream=True,
        tools=kwargs.pop("tools") if "tools" in kwargs else None,
        **kwargs,
    )
    for chunk in stream:
        yield GeminiCallResponseChunk(
            chunk=chunk,
            user_message_param=user_message_param,
            tool_types=tool_types,
        )

stream_async(retries=0, **kwargs) async

Streams the response asynchronously for a call using this GeminiCall.

Parameters:

Name	Type	Description	Default
**kwargs	Any	Additional keyword arguments parameters to pass to the call. These will override any existing arguments in call_params.	{}

Yields:

Type	Description
AsyncGenerator[GeminiCallResponseChunk, None]	A GeminiCallResponseChunk for each chunk of the response.

Source code in mirascope/gemini/calls.py

@retry
async def stream_async(
    self, retries: Union[int, AsyncRetrying] = 0, **kwargs: Any
) -> AsyncGenerator[GeminiCallResponseChunk, None]:
    """Streams the response asynchronously for a call using this `GeminiCall`.

    Args:
        **kwargs: Additional keyword arguments parameters to pass to the call. These
            will override any existing arguments in `call_params`.

    Yields:
        A `GeminiCallResponseChunk` for each chunk of the response.
    """
    kwargs, tool_types = self._setup(kwargs, GeminiTool)
    model_name = kwargs.pop("model")
    gemini_pro_model = get_wrapped_async_client(
        GenerativeModel(model_name=model_name), self
    )
    generate_content_async = get_wrapped_call(
        gemini_pro_model.generate_content_async,
        self,
        is_async=True,
        response_chunk_type=GeminiCallResponseChunk,
        tool_types=tool_types,
        model_name=model_name,
    )
    messages = self.messages()
    user_message_param = self._get_possible_user_message(messages)
    stream = generate_content_async(
        messages,
        stream=True,
        tools=kwargs.pop("tools") if "tools" in kwargs else None,
        **kwargs,
    )
    if inspect.iscoroutine(stream):
        stream = await stream
    async for chunk in stream:
        yield GeminiCallResponseChunk(
            chunk=chunk,
            user_message_param=user_message_param,
            tool_types=tool_types,
        )

GeminiCallParams

Bases: BaseCallParams[GeminiTool]

The parameters to use when calling the Gemini API calls.

Example:

from mirascope.gemini import GeminiCall, GeminiCallParams


class BookRecommendation(GeminiPrompt):
    prompt_template = "Please recommend a {genre} book"

    genre: str

    call_params = GeminiCallParams(
        model="gemini-1.0-pro-001",
        generation_config={"candidate_count": 2},
    )


response = BookRecommender(genre="fantasy").call()
print(response.content)
#> The Name of the Wind

Source code in mirascope/gemini/types.py

class GeminiCallParams(BaseCallParams[GeminiTool]):
    """The parameters to use when calling the Gemini API calls.

    Example:

    ```python
    from mirascope.gemini import GeminiCall, GeminiCallParams


    class BookRecommendation(GeminiPrompt):
        prompt_template = "Please recommend a {genre} book"

        genre: str

        call_params = GeminiCallParams(
            model="gemini-1.0-pro-001",
            generation_config={"candidate_count": 2},
        )


    response = BookRecommender(genre="fantasy").call()
    print(response.content)
    #> The Name of the Wind
    ```
    """

    model: str = "gemini-1.0-pro"
    generation_config: Optional[dict[str, Any]] = {"candidate_count": 1}
    safety_settings: Optional[Any] = None
    request_options: Optional[dict[str, Any]] = None

GeminiExtractor

Bases: BaseExtractor[GeminiCall, GeminiTool, Any, T], Generic[T]

A class for extracting structured information using Google's Gemini Chat models.

Example:

from typing import Literal, Type
from pydantic import BaseModel
from mirascope.gemini import GeminiExtractor

class TaskDetails(BaseModel):
    title: str
    priority: Literal["low", "normal", "high"]
    due_date: str

class TaskExtractor(GeminiExtractor[TaskDetails]):
    extract_schema: Type[TaskDetails] = TaskDetails

    prompt_template = """
    USER: I need to extract task details.
    MODEL: Sure, please provide the task description.
    USER: {task}
    """

    task: str

task_description = "Prepare the budget report by next Monday. It's a high priority task."
task = TaskExtractor(task=task_description).extract(retries=3)
assert isinstance(task, TaskDetails)
print(task)
#> title='Prepare the budget report' priority='high' due_date='next Monday'

Source code in mirascope/gemini/extractors.py

class GeminiExtractor(BaseExtractor[GeminiCall, GeminiTool, Any, T], Generic[T]):
    '''A class for extracting structured information using Google's Gemini Chat models.

    Example:

    ```python
    from typing import Literal, Type
    from pydantic import BaseModel
    from mirascope.gemini import GeminiExtractor

    class TaskDetails(BaseModel):
        title: str
        priority: Literal["low", "normal", "high"]
        due_date: str

    class TaskExtractor(GeminiExtractor[TaskDetails]):
        extract_schema: Type[TaskDetails] = TaskDetails

        prompt_template = """
        USER: I need to extract task details.
        MODEL: Sure, please provide the task description.
        USER: {task}
        """

        task: str

    task_description = "Prepare the budget report by next Monday. It's a high priority task."
    task = TaskExtractor(task=task_description).extract(retries=3)
    assert isinstance(task, TaskDetails)
    print(task)
    #> title='Prepare the budget report' priority='high' due_date='next Monday'
    ```
    '''

    call_params: ClassVar[GeminiCallParams] = GeminiCallParams()
    _provider: ClassVar[str] = "gemini"

    def extract(self, retries: Union[int, Retrying] = 0, **kwargs: Any) -> T:
        """Extracts `extract_schema` from the Gemini call response.

        The `extract_schema` is converted into a `GeminiTool`, complete with a
        description of the tool, all of the fields, and their types. This allows us to
        take advantage of Gemini's tool/function calling functionality to extract
        information from a prompt according to the context provided by the `BaseModel`
        schema.

        Args:
            retries: The maximum number of times to retry the query on validation error.
            **kwargs: Additional keyword arguments parameters to pass to the call. These
                will override any existing arguments in `call_params`.

        Returns:
            The `Schema` instance extracted from the completion.

        Raises:
            AttributeError: if there is no tool in the call creation.
            ValidationError: if the schema cannot be instantiated from the completion.
            GeminiError: raises any Gemini errors.
        """
        return self._extract(GeminiCall, GeminiTool, retries, **kwargs)

    async def extract_async(
        self, retries: Union[int, AsyncRetrying] = 0, **kwargs: Any
    ) -> T:
        """Asynchronously extracts `extract_schema` from the Gemini call response.

        The `extract_schema` is converted into a `GeminiTool`, complete with a
        description of the tool, all of the fields, and their types. This allows us to
        take advantage of Gemini's tool/function calling functionality to extract
        information from a prompt according to the context provided by the `BaseModel`
        schema.

        Args:
            retries: The maximum number of times to retry the query on validation error.
            **kwargs: Additional keyword arguments parameters to pass to the call. These
                will override any existing arguments in `call_params`.

        Returns:
            The `Schema` instance extracted from the completion.

        Raises:
            AttributeError: if there is no tool in the call creation.
            ValidationError: if the schema cannot be instantiated from the completion.
            GeminiError: raises any Gemini errors.
        """
        return await self._extract_async(GeminiCall, GeminiTool, retries, **kwargs)

extract(retries=0, **kwargs)

Extracts extract_schema from the Gemini call response.

The extract_schema is converted into a GeminiTool, complete with a description of the tool, all of the fields, and their types. This allows us to take advantage of Gemini's tool/function calling functionality to extract information from a prompt according to the context provided by the BaseModel schema.

Parameters:

Name	Type	Description	Default
retries	Union[int, Retrying]	The maximum number of times to retry the query on validation error.	0
**kwargs	Any	Additional keyword arguments parameters to pass to the call. These will override any existing arguments in call_params.	{}

Returns:

Type	Description
T	The Schema instance extracted from the completion.

Raises:

Type	Description
AttributeError	if there is no tool in the call creation.
ValidationError	if the schema cannot be instantiated from the completion.
GeminiError	raises any Gemini errors.

Source code in mirascope/gemini/extractors.py

def extract(self, retries: Union[int, Retrying] = 0, **kwargs: Any) -> T:
    """Extracts `extract_schema` from the Gemini call response.

    The `extract_schema` is converted into a `GeminiTool`, complete with a
    description of the tool, all of the fields, and their types. This allows us to
    take advantage of Gemini's tool/function calling functionality to extract
    information from a prompt according to the context provided by the `BaseModel`
    schema.

    Args:
        retries: The maximum number of times to retry the query on validation error.
        **kwargs: Additional keyword arguments parameters to pass to the call. These
            will override any existing arguments in `call_params`.

    Returns:
        The `Schema` instance extracted from the completion.

    Raises:
        AttributeError: if there is no tool in the call creation.
        ValidationError: if the schema cannot be instantiated from the completion.
        GeminiError: raises any Gemini errors.
    """
    return self._extract(GeminiCall, GeminiTool, retries, **kwargs)

extract_async(retries=0, **kwargs) async

Asynchronously extracts extract_schema from the Gemini call response.

The extract_schema is converted into a GeminiTool, complete with a description of the tool, all of the fields, and their types. This allows us to take advantage of Gemini's tool/function calling functionality to extract information from a prompt according to the context provided by the BaseModel schema.

Parameters:

Name	Type	Description	Default
retries	Union[int, AsyncRetrying]	The maximum number of times to retry the query on validation error.	0
**kwargs	Any	Additional keyword arguments parameters to pass to the call. These will override any existing arguments in call_params.	{}

Returns:

Type	Description
T	The Schema instance extracted from the completion.

Raises:

Type	Description
AttributeError	if there is no tool in the call creation.
ValidationError	if the schema cannot be instantiated from the completion.
GeminiError	raises any Gemini errors.

Source code in mirascope/gemini/extractors.py

async def extract_async(
    self, retries: Union[int, AsyncRetrying] = 0, **kwargs: Any
) -> T:
    """Asynchronously extracts `extract_schema` from the Gemini call response.

    The `extract_schema` is converted into a `GeminiTool`, complete with a
    description of the tool, all of the fields, and their types. This allows us to
    take advantage of Gemini's tool/function calling functionality to extract
    information from a prompt according to the context provided by the `BaseModel`
    schema.

    Args:
        retries: The maximum number of times to retry the query on validation error.
        **kwargs: Additional keyword arguments parameters to pass to the call. These
            will override any existing arguments in `call_params`.

    Returns:
        The `Schema` instance extracted from the completion.

    Raises:
        AttributeError: if there is no tool in the call creation.
        ValidationError: if the schema cannot be instantiated from the completion.
        GeminiError: raises any Gemini errors.
    """
    return await self._extract_async(GeminiCall, GeminiTool, retries, **kwargs)

GeminiTool

Bases: BaseTool[FunctionCall]

A base class for easy use of tools with the Gemini API.

GeminiTool internally handles the logic that allows you to use tools with simple calls such as GeminiCompletion.tool or GeminiTool.fn, as seen in the examples below.

Example:

from mirascope.gemini import GeminiCall, GeminiCallParams, GeminiTool


class CurrentWeather(GeminiTool):
    """A tool for getting the current weather in a location."""

    location: str


class WeatherForecast(GeminiPrompt):
    prompt_template = "What is the current weather in {city}?"

    city: str

    call_params = GeminiCallParams(
        model="gemini-pro",
        tools=[CurrentWeather],
    )


prompt = WeatherPrompt()
forecast = WeatherForecast(city="Tokyo").call().tool
print(forecast.location)
#> Tokyo

Source code in mirascope/gemini/tools.py

class GeminiTool(BaseTool[FunctionCall]):
    '''A base class for easy use of tools with the Gemini API.

    `GeminiTool` internally handles the logic that allows you to use tools with simple
    calls such as `GeminiCompletion.tool` or `GeminiTool.fn`, as seen in the
    examples below.

    Example:

    ```python
    from mirascope.gemini import GeminiCall, GeminiCallParams, GeminiTool


    class CurrentWeather(GeminiTool):
        """A tool for getting the current weather in a location."""

        location: str


    class WeatherForecast(GeminiPrompt):
        prompt_template = "What is the current weather in {city}?"

        city: str

        call_params = GeminiCallParams(
            model="gemini-pro",
            tools=[CurrentWeather],
        )


    prompt = WeatherPrompt()
    forecast = WeatherForecast(city="Tokyo").call().tool
    print(forecast.location)
    #> Tokyo
    ```
    '''

    model_config = ConfigDict(arbitrary_types_allowed=True)

    @classmethod
    def tool_schema(cls) -> Tool:
        """Constructs a tool schema for use with the Gemini API.

        A Mirascope `GeminiTool` is deconstructed into a `Tool` schema for use with the
        Gemini API.

        Returns:
            The constructed `Tool` schema.
        """
        tool_schema = super().tool_schema()
        if "parameters" in tool_schema:
            if "$defs" in tool_schema["parameters"]:
                raise ValueError(
                    "Unfortunately Google's Gemini API cannot handle nested structures "
                    "with $defs."
                )
            tool_schema["parameters"]["properties"] = {
                prop: {
                    key: value for key, value in prop_schema.items() if key != "title"
                }
                for prop, prop_schema in tool_schema["parameters"]["properties"].items()
            }
        return Tool(function_declarations=[FunctionDeclaration(**tool_schema)])

    @classmethod
    def from_tool_call(cls, tool_call: FunctionCall) -> GeminiTool:
        """Extracts an instance of the tool constructed from a tool call response.

        Given a `GenerateContentResponse` from a Gemini chat completion response, this
        method extracts the tool call and constructs an instance of the tool.

        Args:
            tool_call: The `GenerateContentResponse` from which to extract the tool.

        Returns:
            An instance of the tool constructed from the tool call.

        Raises:
            ValueError: if the tool call doesn't have any arguments.
            ValidationError: if the tool call doesn't match the tool schema.
        """
        if not tool_call.args:
            raise ValueError("Tool call doesn't have any arguments.")
        model_json = {key: value for key, value in tool_call.args.items()}
        model_json["tool_call"] = tool_call
        return cls.model_validate(model_json)

    @classmethod
    def from_model(cls, model: Type[BaseModel]) -> Type[GeminiTool]:
        """Constructs a `GeminiTool` type from a `BaseModel` type."""
        return convert_base_model_to_tool(model, GeminiTool)

    @classmethod
    def from_fn(cls, fn: Callable) -> Type[GeminiTool]:
        """Constructs a `GeminiTool` type from a function."""
        return convert_function_to_tool(fn, GeminiTool)

    @classmethod
    def from_base_type(cls, base_type: Type[BaseType]) -> Type[GeminiTool]:
        """Constructs a `GeminiTool` type from a `BaseType` type."""
        return convert_base_type_to_tool(base_type, GeminiTool)

from_base_type(base_type) classmethod

Constructs a GeminiTool type from a BaseType type.

Source code in mirascope/gemini/tools.py

@classmethod
def from_base_type(cls, base_type: Type[BaseType]) -> Type[GeminiTool]:
    """Constructs a `GeminiTool` type from a `BaseType` type."""
    return convert_base_type_to_tool(base_type, GeminiTool)

from_fn(fn) classmethod

Constructs a GeminiTool type from a function.

Source code in mirascope/gemini/tools.py

@classmethod
def from_fn(cls, fn: Callable) -> Type[GeminiTool]:
    """Constructs a `GeminiTool` type from a function."""
    return convert_function_to_tool(fn, GeminiTool)

from_model(model) classmethod

Constructs a GeminiTool type from a BaseModel type.

Source code in mirascope/gemini/tools.py

@classmethod
def from_model(cls, model: Type[BaseModel]) -> Type[GeminiTool]:
    """Constructs a `GeminiTool` type from a `BaseModel` type."""
    return convert_base_model_to_tool(model, GeminiTool)

from_tool_call(tool_call) classmethod

Extracts an instance of the tool constructed from a tool call response.

Given a GenerateContentResponse from a Gemini chat completion response, this method extracts the tool call and constructs an instance of the tool.

Parameters:

Name	Type	Description	Default
tool_call	FunctionCall	The GenerateContentResponse from which to extract the tool.	required

Returns:

Type	Description
GeminiTool	An instance of the tool constructed from the tool call.

Raises:

Type	Description
ValueError	if the tool call doesn't have any arguments.
ValidationError	if the tool call doesn't match the tool schema.

Source code in mirascope/gemini/tools.py

@classmethod
def from_tool_call(cls, tool_call: FunctionCall) -> GeminiTool:
    """Extracts an instance of the tool constructed from a tool call response.

    Given a `GenerateContentResponse` from a Gemini chat completion response, this
    method extracts the tool call and constructs an instance of the tool.

    Args:
        tool_call: The `GenerateContentResponse` from which to extract the tool.

    Returns:
        An instance of the tool constructed from the tool call.

    Raises:
        ValueError: if the tool call doesn't have any arguments.
        ValidationError: if the tool call doesn't match the tool schema.
    """
    if not tool_call.args:
        raise ValueError("Tool call doesn't have any arguments.")
    model_json = {key: value for key, value in tool_call.args.items()}
    model_json["tool_call"] = tool_call
    return cls.model_validate(model_json)

tool_schema() classmethod

Constructs a tool schema for use with the Gemini API.

A Mirascope GeminiTool is deconstructed into a Tool schema for use with the Gemini API.

Returns:

Type	Description
Tool	The constructed Tool schema.

Source code in mirascope/gemini/tools.py

@classmethod
def tool_schema(cls) -> Tool:
    """Constructs a tool schema for use with the Gemini API.

    A Mirascope `GeminiTool` is deconstructed into a `Tool` schema for use with the
    Gemini API.

    Returns:
        The constructed `Tool` schema.
    """
    tool_schema = super().tool_schema()
    if "parameters" in tool_schema:
        if "$defs" in tool_schema["parameters"]:
            raise ValueError(
                "Unfortunately Google's Gemini API cannot handle nested structures "
                "with $defs."
            )
        tool_schema["parameters"]["properties"] = {
            prop: {
                key: value for key, value in prop_schema.items() if key != "title"
            }
            for prop, prop_schema in tool_schema["parameters"]["properties"].items()
        }
    return Tool(function_declarations=[FunctionDeclaration(**tool_schema)])