Source code for components.model_client.google_client

"""Google GenAI ModelClient integration."""

import os
from typing import Dict, Sequence, Optional, Any
import backoff
import logging


from adalflow.core.model_client import ModelClient
from adalflow.core.types import CompletionUsage, ModelType, GeneratorOutput

from adalflow.utils.lazy_import import safe_import, OptionalPackages

# optional import
google = safe_import(
    OptionalPackages.GOOGLE_GENERATIVEAI.value[0],
    OptionalPackages.GOOGLE_GENERATIVEAI.value[1],
)
import google.generativeai as genai
from google.api_core.exceptions import (
    InternalServerError,
    BadRequest,
    GoogleAPICallError,
)
from google.generativeai.types import GenerateContentResponse

log = logging.getLogger(__name__)




class GoogleGenAIClient(ModelClient):
    __doc__ = r"""A component wrapper for the Google GenAI API client.

    Visit https://cloud.google.com/vertex-ai/generative-ai/docs/model-reference/inference for more api details.

    Info: 8/1/2024
    Tested: gemini-1.0-pro, gemini-1.5-pro-latest

    .. code-block:: python

        class UsageMetadata(proto.Message):

            prompt_token_count: int = proto.Field(
                proto.INT32,
                number=1,
            )
            cached_content_token_count: int = proto.Field(
                proto.INT32,
                number=4,
            )
            candidates_token_count: int = proto.Field(
                proto.INT32,
                number=2,
            )
            total_token_count: int = proto.Field(
                proto.INT32,
                number=3,
            )
    """

    def __init__(self, api_key: Optional[str] = None):
        r"""It is recommended to set the GOOGLE_API_KEY environment variable instead of passing it as an argument."""
        super().__init__()
        self._api_key = api_key
        self.sync_client = self.init_sync_client()
        self.async_client = None  # only initialize if the async call is called
        self.tested_llm_models = ["gemini-1.0-pro", "gemini-1.5-pro-latest"]



    def init_sync_client(self):
        api_key = self._api_key or os.getenv("GOOGLE_API_KEY")
        if not api_key:
            raise ValueError("Environment variable GOOGLE_API_KEY must be set")
        genai.configure(api_key=api_key)
        return genai




    def parse_chat_completion(
        self, completion: GenerateContentResponse
    ) -> "GeneratorOutput":
        """
        Parse the completion to a structure your sytem standarizes. (here is str)
        """
        log.debug(f"completion: {completion}")
        try:
            data = completion.text
            usage = self.track_completion_usage(completion)
            return GeneratorOutput(data=None, usage=usage, raw_response=data)
        except Exception as e:
            log.error(f"Error parsing completion: {e}")
            return GeneratorOutput(
                data=None, error=str(e), raw_response=str(completion)
            )




    def track_completion_usage(
        self, completion: GenerateContentResponse
    ) -> CompletionUsage:
        return CompletionUsage(
            completion_tokens=completion.usage_metadata.candidates_token_count,
            prompt_tokens=completion.usage_metadata.prompt_token_count,
            total_tokens=completion.usage_metadata.total_token_count,
        )




    def convert_inputs_to_api_kwargs(
        self,
        input: Optional[Any] = None,
        model_kwargs: Dict = {},
        model_type: ModelType = ModelType.UNDEFINED,
    ) -> Dict:
        r"""
        Specify the API input type and output api_kwargs that will be used in _call and _acall methods.
        Convert the Component's standard input, and system_input(chat model) and model_kwargs into API-specific format
        """
        final_model_kwargs = model_kwargs.copy()
        if model_type == ModelType.EMBEDDER:
            if isinstance(input, str):
                input = [input]
            # convert input to input
            if not isinstance(input, Sequence):
                raise TypeError("input must be a sequence of text")
            final_model_kwargs["input"] = input
        elif model_type == ModelType.LLM:

            final_model_kwargs["prompt"] = input
        else:
            raise ValueError(f"model_type {model_type} is not supported")
        return final_model_kwargs




    @backoff.on_exception(
        backoff.expo,
        (
            InternalServerError,
            BadRequest,
            GoogleAPICallError,
        ),
        max_time=5,
    )
    def call(self, api_kwargs: Dict = {}, model_type: ModelType = ModelType.UNDEFINED):
        """
        kwargs is the combined input and model_kwargs
        """

        if model_type == ModelType.LLM:
            # remove model from api_kwargs
            model = api_kwargs.pop("model")
            prompt = api_kwargs.pop("prompt")

            assert model != "", "model must be specified"
            assert prompt != "", "prompt must be specified"

            config = genai.GenerationConfig(**api_kwargs)
            llm = genai.GenerativeModel(model_name=model, generation_config=config)
            return llm.generate_content(contents=prompt)
        else:
            raise ValueError(f"model_type {model_type} is not supported")




if __name__ == "__main__":
    from adalflow.utils import setup_env

    setup_env()
    google_client = GoogleGenAIClient()