langchain/libs/partners/mistralai/langchain_mistralai/chat_models.py

from __future__ import annotations

import logging
import uuid
from operator import itemgetter
from typing import (
    Any,
    AsyncContextManager,
    AsyncIterator,
    Callable,
    Dict,
    Iterator,
    List,
    Optional,
    Sequence,
    Tuple,
    Type,
    Union,
    cast,
)

import httpx
from httpx_sse import EventSource, aconnect_sse, connect_sse
from langchain_core._api import beta
from langchain_core.callbacks import (
    AsyncCallbackManagerForLLMRun,
    CallbackManagerForLLMRun,
)
from langchain_core.language_models import LanguageModelInput
from langchain_core.language_models.chat_models import (
    BaseChatModel,
    agenerate_from_stream,
    generate_from_stream,
)
from langchain_core.language_models.llms import create_base_retry_decorator
from langchain_core.messages import (
    AIMessage,
    AIMessageChunk,
    BaseMessage,
    BaseMessageChunk,
    ChatMessage,
    ChatMessageChunk,
    HumanMessage,
    HumanMessageChunk,
    SystemMessage,
    SystemMessageChunk,
    ToolMessage,
)
from langchain_core.output_parsers.base import OutputParserLike
from langchain_core.output_parsers.openai_tools import (
    JsonOutputKeyToolsParser,
    PydanticToolsParser,
    make_invalid_tool_call,
    parse_tool_call,
)
from langchain_core.outputs import ChatGeneration, ChatGenerationChunk, ChatResult
from langchain_core.pydantic_v1 import BaseModel, Field, SecretStr, root_validator
from langchain_core.runnables import Runnable, RunnableMap, RunnablePassthrough
from langchain_core.tools import BaseTool
from langchain_core.utils import convert_to_secret_str, get_from_dict_or_env
from langchain_core.utils.function_calling import convert_to_openai_tool

logger = logging.getLogger(__name__)


def _create_retry_decorator(
    llm: ChatMistralAI,
    run_manager: Optional[
        Union[AsyncCallbackManagerForLLMRun, CallbackManagerForLLMRun]
    ] = None,
) -> Callable[[Any], Any]:
    """Returns a tenacity retry decorator, preconfigured to handle exceptions"""

    errors = [httpx.RequestError, httpx.StreamError]
    return create_base_retry_decorator(
        error_types=errors, max_retries=llm.max_retries, run_manager=run_manager
    )


def _convert_mistral_chat_message_to_message(
    _message: Dict,
) -> BaseMessage:
    role = _message["role"]
    assert role == "assistant", f"Expected role to be 'assistant', got {role}"
    content = cast(str, _message["content"])

    additional_kwargs: Dict = {}
    tool_calls = []
    invalid_tool_calls = []
    if raw_tool_calls := _message.get("tool_calls"):
        additional_kwargs["tool_calls"] = raw_tool_calls
        for raw_tool_call in raw_tool_calls:
            try:
                parsed: dict = cast(
                    dict, parse_tool_call(raw_tool_call, return_id=True)
                )
                if not parsed["id"]:
                    tool_call_id = uuid.uuid4().hex[:]
                    tool_calls.append(
                        {
                            **parsed,
                            **{"id": tool_call_id},
                        },
                    )
                else:
                    tool_calls.append(parsed)
            except Exception as e:
                invalid_tool_calls.append(
                    dict(make_invalid_tool_call(raw_tool_call, str(e)))
                )
    return AIMessage(
        content=content,
        additional_kwargs=additional_kwargs,
        tool_calls=tool_calls,
        invalid_tool_calls=invalid_tool_calls,
    )


def _raise_on_error(response: httpx.Response) -> None:
    """Raise an error if the response is an error."""
    if httpx.codes.is_error(response.status_code):
        error_message = response.read().decode("utf-8")
        raise httpx.HTTPStatusError(
            f"Error response {response.status_code} "
            f"while fetching {response.url}: {error_message}",
            request=response.request,
            response=response,
        )


async def _araise_on_error(response: httpx.Response) -> None:
    """Raise an error if the response is an error."""
    if httpx.codes.is_error(response.status_code):
        error_message = (await response.aread()).decode("utf-8")
        raise httpx.HTTPStatusError(
            f"Error response {response.status_code} "
            f"while fetching {response.url}: {error_message}",
            request=response.request,
            response=response,
        )


async def _aiter_sse(
    event_source_mgr: AsyncContextManager[EventSource],
) -> AsyncIterator[Dict]:
    """Iterate over the server-sent events."""
    async with event_source_mgr as event_source:
        # TODO(Team): Remove after this is fixed in httpx dependency
        # https://github.com/florimondmanca/httpx-sse/pull/25/files
        await _araise_on_error(event_source._response)
        async for event in event_source.aiter_sse():
            if event.data == "[DONE]":
                return
            yield event.json()


async def acompletion_with_retry(
    llm: ChatMistralAI,
    run_manager: Optional[AsyncCallbackManagerForLLMRun] = None,
    **kwargs: Any,
) -> Any:
    """Use tenacity to retry the async completion call."""
    retry_decorator = _create_retry_decorator(llm, run_manager=run_manager)

    @retry_decorator
    async def _completion_with_retry(**kwargs: Any) -> Any:
        if "stream" not in kwargs:
            kwargs["stream"] = False
        stream = kwargs["stream"]
        if stream:
            event_source = aconnect_sse(
                llm.async_client, "POST", "/chat/completions", json=kwargs
            )
            return _aiter_sse(event_source)
        else:
            response = await llm.async_client.post(url="/chat/completions", json=kwargs)
            await _araise_on_error(response)
            return response.json()

    return await _completion_with_retry(**kwargs)


def _convert_delta_to_message_chunk(
    _delta: Dict, default_class: Type[BaseMessageChunk]
) -> BaseMessageChunk:
    role = _delta.get("role")
    content = _delta.get("content") or ""
    if role == "user" or default_class == HumanMessageChunk:
        return HumanMessageChunk(content=content)
    elif role == "assistant" or default_class == AIMessageChunk:
        additional_kwargs: Dict = {}
        if raw_tool_calls := _delta.get("tool_calls"):
            additional_kwargs["tool_calls"] = raw_tool_calls
            try:
                tool_call_chunks = []
                for raw_tool_call in raw_tool_calls:
                    if not raw_tool_call.get("index") and not raw_tool_call.get("id"):
                        tool_call_id = uuid.uuid4().hex[:]
                    else:
                        tool_call_id = raw_tool_call.get("id")
                    tool_call_chunks.append(
                        {
                            "name": raw_tool_call["function"].get("name"),
                            "args": raw_tool_call["function"].get("arguments"),
                            "id": tool_call_id,
                            "index": raw_tool_call.get("index"),
                        }
                    )
            except KeyError:
                pass
        else:
            tool_call_chunks = []
        return AIMessageChunk(
            content=content,
            additional_kwargs=additional_kwargs,
            tool_call_chunks=tool_call_chunks,
        )
    elif role == "system" or default_class == SystemMessageChunk:
        return SystemMessageChunk(content=content)
    elif role or default_class == ChatMessageChunk:
        return ChatMessageChunk(content=content, role=role)
    else:
        return default_class(content=content)


def _convert_message_to_mistral_chat_message(
    message: BaseMessage,
) -> Dict:
    if isinstance(message, ChatMessage):
        return dict(role=message.role, content=message.content)
    elif isinstance(message, HumanMessage):
        return dict(role="user", content=message.content)
    elif isinstance(message, AIMessage):
        if "tool_calls" in message.additional_kwargs:
            tool_calls = []
            for tc in message.additional_kwargs["tool_calls"]:
                chunk = {
                    "function": {
                        "name": tc["function"]["name"],
                        "arguments": tc["function"]["arguments"],
                    }
                }
                if _id := tc.get("id"):
                    chunk["id"] = _id
                tool_calls.append(chunk)
        else:
            tool_calls = []
        return {
            "role": "assistant",
            "content": message.content,
            "tool_calls": tool_calls,
        }
    elif isinstance(message, SystemMessage):
        return dict(role="system", content=message.content)
    elif isinstance(message, ToolMessage):
        return {
            "role": "tool",
            "content": message.content,
            "name": message.name,
        }
    else:
        raise ValueError(f"Got unknown type {message}")


class ChatMistralAI(BaseChatModel):
    """A chat model that uses the MistralAI API."""

    client: httpx.Client = Field(default=None)  #: :meta private:
    async_client: httpx.AsyncClient = Field(default=None)  #: :meta private:
    mistral_api_key: Optional[SecretStr] = Field(default=None, alias="api_key")
    endpoint: str = "https://api.mistral.ai/v1"
    max_retries: int = 5
    timeout: int = 120
    max_concurrent_requests: int = 64

    model: str = "mistral-small"
    temperature: float = 0.7
    max_tokens: Optional[int] = None
    top_p: float = 1
    """Decode using nucleus sampling: consider the smallest set of tokens whose
       probability sum is at least top_p. Must be in the closed interval [0.0, 1.0]."""
    random_seed: Optional[int] = None
    safe_mode: bool = False
    streaming: bool = False

    class Config:
        """Configuration for this pydantic object."""

        allow_population_by_field_name = True
        arbitrary_types_allowed = True

    @property
    def _default_params(self) -> Dict[str, Any]:
        """Get the default parameters for calling the API."""
        defaults = {
            "model": self.model,
            "temperature": self.temperature,
            "max_tokens": self.max_tokens,
            "top_p": self.top_p,
            "random_seed": self.random_seed,
            "safe_prompt": self.safe_mode,
        }
        filtered = {k: v for k, v in defaults.items() if v is not None}
        return filtered

    @property
    def _client_params(self) -> Dict[str, Any]:
        """Get the parameters used for the client."""
        return self._default_params

    def completion_with_retry(
        self, run_manager: Optional[CallbackManagerForLLMRun] = None, **kwargs: Any
    ) -> Any:
        """Use tenacity to retry the completion call."""
        # retry_decorator = _create_retry_decorator(self, run_manager=run_manager)

        # @retry_decorator
        def _completion_with_retry(**kwargs: Any) -> Any:
            if "stream" not in kwargs:
                kwargs["stream"] = False
            stream = kwargs["stream"]
            if stream:

                def iter_sse() -> Iterator[Dict]:
                    with connect_sse(
                        self.client, "POST", "/chat/completions", json=kwargs
                    ) as event_source:
                        # TODO(Team): Remove after this is fixed in httpx dependency
                        # https://github.com/florimondmanca/httpx-sse/pull/25/files
                        _raise_on_error(event_source._response)
                        for event in event_source.iter_sse():
                            if event.data == "[DONE]":
                                return
                            yield event.json()

                return iter_sse()
            else:
                response = self.client.post(url="/chat/completions", json=kwargs)
                _raise_on_error(response)
                return response.json()

        rtn = _completion_with_retry(**kwargs)
        return rtn

    def _combine_llm_outputs(self, llm_outputs: List[Optional[dict]]) -> dict:
        overall_token_usage: dict = {}
        for output in llm_outputs:
            if output is None:
                # Happens in streaming
                continue
            token_usage = output["token_usage"]
            if token_usage is not None:
                for k, v in token_usage.items():
                    if k in overall_token_usage:
                        overall_token_usage[k] += v
                    else:
                        overall_token_usage[k] = v
        combined = {"token_usage": overall_token_usage, "model_name": self.model}
        return combined

    @root_validator()
    def validate_environment(cls, values: Dict) -> Dict:
        """Validate api key, python package exists, temperature, and top_p."""

        values["mistral_api_key"] = convert_to_secret_str(
            get_from_dict_or_env(
                values, "mistral_api_key", "MISTRAL_API_KEY", default=""
            )
        )
        api_key_str = values["mistral_api_key"].get_secret_value()
        # todo: handle retries
        values["client"] = httpx.Client(
            base_url=values["endpoint"],
            headers={
                "Content-Type": "application/json",
                "Accept": "application/json",
                "Authorization": f"Bearer {api_key_str}",
            },
            timeout=values["timeout"],
        )
        # todo: handle retries and max_concurrency
        values["async_client"] = httpx.AsyncClient(
            base_url=values["endpoint"],
            headers={
                "Content-Type": "application/json",
                "Accept": "application/json",
                "Authorization": f"Bearer {api_key_str}",
            },
            timeout=values["timeout"],
        )

        if values["temperature"] is not None and not 0 <= values["temperature"] <= 1:
            raise ValueError("temperature must be in the range [0.0, 1.0]")

        if values["top_p"] is not None and not 0 <= values["top_p"] <= 1:
            raise ValueError("top_p must be in the range [0.0, 1.0]")

        return values

    def _generate(
        self,
        messages: List[BaseMessage],
        stop: Optional[List[str]] = None,
        run_manager: Optional[CallbackManagerForLLMRun] = None,
        stream: Optional[bool] = None,
        **kwargs: Any,
    ) -> ChatResult:
        should_stream = stream if stream is not None else self.streaming
        if should_stream:
            stream_iter = self._stream(
                messages, stop=stop, run_manager=run_manager, **kwargs
            )
            return generate_from_stream(stream_iter)

        message_dicts, params = self._create_message_dicts(messages, stop)
        params = {**params, **kwargs}
        response = self.completion_with_retry(
            messages=message_dicts, run_manager=run_manager, **params
        )
        return self._create_chat_result(response)

    def _create_chat_result(self, response: Dict) -> ChatResult:
        generations = []
        for res in response["choices"]:
            finish_reason = res.get("finish_reason")
            gen = ChatGeneration(
                message=_convert_mistral_chat_message_to_message(res["message"]),
                generation_info={"finish_reason": finish_reason},
            )
            generations.append(gen)
        token_usage = response.get("usage", {})

        llm_output = {"token_usage": token_usage, "model": self.model}
        return ChatResult(generations=generations, llm_output=llm_output)

    def _create_message_dicts(
        self, messages: List[BaseMessage], stop: Optional[List[str]]
    ) -> Tuple[List[Dict], Dict[str, Any]]:
        params = self._client_params
        if stop is not None or "stop" in params:
            if "stop" in params:
                params.pop("stop")
            logger.warning(
                "Parameter `stop` not yet supported (https://docs.mistral.ai/api)"
            )
        message_dicts = [_convert_message_to_mistral_chat_message(m) for m in messages]
        return message_dicts, params

    def _stream(
        self,
        messages: List[BaseMessage],
        stop: Optional[List[str]] = None,
        run_manager: Optional[CallbackManagerForLLMRun] = None,
        **kwargs: Any,
    ) -> Iterator[ChatGenerationChunk]:
        message_dicts, params = self._create_message_dicts(messages, stop)
        params = {**params, **kwargs, "stream": True}

        default_chunk_class: Type[BaseMessageChunk] = AIMessageChunk
        for chunk in self.completion_with_retry(
            messages=message_dicts, run_manager=run_manager, **params
        ):
            if len(chunk["choices"]) == 0:
                continue
            delta = chunk["choices"][0]["delta"]
            new_chunk = _convert_delta_to_message_chunk(delta, default_chunk_class)
            # make future chunks same type as first chunk
            default_chunk_class = new_chunk.__class__
            gen_chunk = ChatGenerationChunk(message=new_chunk)
            if run_manager:
                run_manager.on_llm_new_token(
                    token=cast(str, new_chunk.content), chunk=gen_chunk
                )
            yield gen_chunk

    async def _astream(
        self,
        messages: List[BaseMessage],
        stop: Optional[List[str]] = None,
        run_manager: Optional[AsyncCallbackManagerForLLMRun] = None,
        **kwargs: Any,
    ) -> AsyncIterator[ChatGenerationChunk]:
        message_dicts, params = self._create_message_dicts(messages, stop)
        params = {**params, **kwargs, "stream": True}

        default_chunk_class: Type[BaseMessageChunk] = AIMessageChunk
        async for chunk in await acompletion_with_retry(
            self, messages=message_dicts, run_manager=run_manager, **params
        ):
            if len(chunk["choices"]) == 0:
                continue
            delta = chunk["choices"][0]["delta"]
            new_chunk = _convert_delta_to_message_chunk(delta, default_chunk_class)
            # make future chunks same type as first chunk
            default_chunk_class = new_chunk.__class__
            gen_chunk = ChatGenerationChunk(message=new_chunk)
            if run_manager:
                await run_manager.on_llm_new_token(
                    token=cast(str, new_chunk.content), chunk=gen_chunk
                )
            yield gen_chunk

    async def _agenerate(
        self,
        messages: List[BaseMessage],
        stop: Optional[List[str]] = None,
        run_manager: Optional[AsyncCallbackManagerForLLMRun] = None,
        stream: Optional[bool] = None,
        **kwargs: Any,
    ) -> ChatResult:
        should_stream = stream if stream is not None else False
        if should_stream:
            stream_iter = self._astream(
                messages=messages, stop=stop, run_manager=run_manager, **kwargs
            )
            return await agenerate_from_stream(stream_iter)

        message_dicts, params = self._create_message_dicts(messages, stop)
        params = {**params, **kwargs}
        response = await acompletion_with_retry(
            self, messages=message_dicts, run_manager=run_manager, **params
        )
        return self._create_chat_result(response)

    def bind_tools(
        self,
        tools: Sequence[Union[Dict[str, Any], Type[BaseModel], Callable, BaseTool]],
        **kwargs: Any,
    ) -> Runnable[LanguageModelInput, BaseMessage]:
        """Bind tool-like objects to this chat model.

        Assumes model is compatible with OpenAI tool-calling API.

        Args:
            tools: A list of tool definitions to bind to this chat model.
                Can be  a dictionary, pydantic model, callable, or BaseTool. Pydantic
                models, callables, and BaseTools will be automatically converted to
                their schema dictionary representation.
            tool_choice: Which tool to require the model to call.
                Must be the name of the single provided function or
                "auto" to automatically determine which function to call
                (if any), or a dict of the form:
                {"type": "function", "function": {"name": <<tool_name>>}}.
            **kwargs: Any additional parameters to pass to the
                :class:`~langchain.runnable.Runnable` constructor.
        """

        formatted_tools = [convert_to_openai_tool(tool) for tool in tools]
        return super().bind(tools=formatted_tools, **kwargs)

    @beta()
    def with_structured_output(
        self,
        schema: Union[Dict, Type[BaseModel]],
        *,
        include_raw: bool = False,
        **kwargs: Any,
    ) -> Runnable[LanguageModelInput, Union[Dict, BaseModel]]:
        """Model wrapper that returns outputs formatted to match the given schema.

        Args:
            schema: The output schema as a dict or a Pydantic class. If a Pydantic class
                then the model output will be an object of that class. If a dict then
                the model output will be a dict. With a Pydantic class the returned
                attributes will be validated, whereas with a dict they will not be. If
                `method` is "function_calling" and `schema` is a dict, then the dict
                must match the OpenAI function-calling spec.
            include_raw: If False then only the parsed structured output is returned. If
                an error occurs during model output parsing it will be raised. If True
                then both the raw model response (a BaseMessage) and the parsed model
                response will be returned. If an error occurs during output parsing it
                will be caught and returned as well. The final output is always a dict
                with keys "raw", "parsed", and "parsing_error".

        Returns:
            A Runnable that takes any ChatModel input and returns as output:

                If include_raw is True then a dict with keys:
                    raw: BaseMessage
                    parsed: Optional[_DictOrPydantic]
                    parsing_error: Optional[BaseException]

                If include_raw is False then just _DictOrPydantic is returned,
                where _DictOrPydantic depends on the schema:

                If schema is a Pydantic class then _DictOrPydantic is the Pydantic
                    class.

                If schema is a dict then _DictOrPydantic is a dict.

        Example: Function-calling, Pydantic schema (method="function_calling", include_raw=False):
            .. code-block:: python

                from langchain_mistralai import ChatMistralAI
                from langchain_core.pydantic_v1 import BaseModel

                class AnswerWithJustification(BaseModel):
                    '''An answer to the user question along with justification for the answer.'''
                    answer: str
                    justification: str

                llm = ChatMistralAI(model="mistral-large-latest", temperature=0)
                structured_llm = llm.with_structured_output(AnswerWithJustification)

                structured_llm.invoke("What weighs more a pound of bricks or a pound of feathers")

                # -> AnswerWithJustification(
                #     answer='They weigh the same',
                #     justification='Both a pound of bricks and a pound of feathers weigh one pound. The weight is the same, but the volume or density of the objects may differ.'
                # )

        Example: Function-calling, Pydantic schema (method="function_calling", include_raw=True):
            .. code-block:: python

                from langchain_mistralai import ChatMistralAI
                from langchain_core.pydantic_v1 import BaseModel

                class AnswerWithJustification(BaseModel):
                    '''An answer to the user question along with justification for the answer.'''
                    answer: str
                    justification: str

                llm = ChatMistralAI(model="mistral-large-latest", temperature=0)
                structured_llm = llm.with_structured_output(AnswerWithJustification, include_raw=True)

                structured_llm.invoke("What weighs more a pound of bricks or a pound of feathers")
                # -> {
                #     'raw': AIMessage(content='', additional_kwargs={'tool_calls': [{'id': 'call_Ao02pnFYXD6GN1yzc0uXPsvF', 'function': {'arguments': '{"answer":"They weigh the same.","justification":"Both a pound of bricks and a pound of feathers weigh one pound. The weight is the same, but the volume or density of the objects may differ."}', 'name': 'AnswerWithJustification'}, 'type': 'function'}]}),
                #     'parsed': AnswerWithJustification(answer='They weigh the same.', justification='Both a pound of bricks and a pound of feathers weigh one pound. The weight is the same, but the volume or density of the objects may differ.'),
                #     'parsing_error': None
                # }

        Example: Function-calling, dict schema (method="function_calling", include_raw=False):
            .. code-block:: python

                from langchain_mistralai import ChatMistralAI
                from langchain_core.pydantic_v1 import BaseModel
                from langchain_core.utils.function_calling import convert_to_openai_tool

                class AnswerWithJustification(BaseModel):
                    '''An answer to the user question along with justification for the answer.'''
                    answer: str
                    justification: str

                dict_schema = convert_to_openai_tool(AnswerWithJustification)
                llm = ChatMistralAI(model="mistral-large-latest", temperature=0)
                structured_llm = llm.with_structured_output(dict_schema)

                structured_llm.invoke("What weighs more a pound of bricks or a pound of feathers")
                # -> {
                #     'answer': 'They weigh the same',
                #     'justification': 'Both a pound of bricks and a pound of feathers weigh one pound. The weight is the same, but the volume and density of the two substances differ.'
                # }

        """  # noqa: E501
        if kwargs:
            raise ValueError(f"Received unsupported arguments {kwargs}")
        is_pydantic_schema = isinstance(schema, type) and issubclass(schema, BaseModel)
        llm = self.bind_tools([schema], tool_choice="any")
        if is_pydantic_schema:
            output_parser: OutputParserLike = PydanticToolsParser(
                tools=[schema], first_tool_only=True
            )
        else:
            key_name = convert_to_openai_tool(schema)["function"]["name"]
            output_parser = JsonOutputKeyToolsParser(
                key_name=key_name, first_tool_only=True
            )

        if include_raw:
            parser_assign = RunnablePassthrough.assign(
                parsed=itemgetter("raw") | output_parser, parsing_error=lambda _: None
            )
            parser_none = RunnablePassthrough.assign(parsed=lambda _: None)
            parser_with_fallback = parser_assign.with_fallbacks(
                [parser_none], exception_key="parsing_error"
            )
            return RunnableMap(raw=llm) | parser_with_fallback
        else:
            return llm | output_parser

    @property
    def _identifying_params(self) -> Dict[str, Any]:
        """Get the identifying parameters."""
        return self._default_params

    @property
    def _llm_type(self) -> str:
        """Return type of chat model."""
        return "mistralai-chat"

    @property
    def lc_secrets(self) -> Dict[str, str]:
        return {"mistral_api_key": "MISTRAL_API_KEY"}

    @classmethod
    def is_lc_serializable(cls) -> bool:
        """Return whether this model can be serialized by Langchain."""
        return True

    @classmethod
    def get_lc_namespace(cls) -> List[str]:
        """Get the namespace of the langchain object."""
        return ["langchain", "chat_models", "mistralai"]