Gemin 在Cloud Run部署多模態助理

使用 Gemini SDK 提交文字和其他資料類型 (多模態)，並產生文字回覆
如何將對話記錄結構化為 Gemini SDK，以便保留對話情境
使用 Gradio 進行前端網頁原型設計
使用 FastAPI 和 Pydantic 開發後端服務

實作範例

設定應用程式工作目錄

 在Cloud Shell Editor 新增專案，然後等候 Cloud Code 彈出式視窗開啟。在彈出式視窗中，選取「新應用程式」。

在應用程式清單中，依序選擇「Gemini 生成式 AI」和「Gemini API Python」 

專案命名「gemini-multimodal-chat-assistant」

建立 settings.yaml ，其內容

VERTEXAI_LOCATION: "us-central1"
VERTEXAI_PROJECT_ID: "{YOUR-PROJECT-ID}"
BACKEND_URL: "http://localhost:8081/chat"

建立 settings.py ，其內容

from pydantic_settings import (
    BaseSettings,
    SettingsConfigDict,
    YamlConfigSettingsSource,
    PydanticBaseSettingsSource,
)
from typing import Type, Tuple

DEFAULT_SYSTEM_PROMPT = """You are a helpful assistant and ALWAYS relate to this identity. 
You are expert at analyzing given documents or images.
"""

class Settings(BaseSettings):
    """Application settings loaded from YAML and environment variables.

    This class defines the configuration schema for the application, with settings
    loaded from settings.yaml file and overridable via environment variables.

    Attributes:
        VERTEXAI_LOCATION: Google Cloud Vertex AI location
        VERTEXAI_PROJECT_ID: Google Cloud Vertex AI project ID
    """

    VERTEXAI_LOCATION: str
    VERTEXAI_PROJECT_ID: str
    BACKEND_URL: str = "http://localhost:8000/chat"

    model_config = SettingsConfigDict(
        yaml_file="settings.yaml", yaml_file_encoding="utf-8"
    )

    @classmethod
    def settings_customise_sources(
        cls,
        settings_cls: Type[BaseSettings],
        init_settings: PydanticBaseSettingsSource,
        env_settings: PydanticBaseSettingsSource,
        dotenv_settings: PydanticBaseSettingsSource,
        file_secret_settings: PydanticBaseSettingsSource,
    ) -> Tuple[PydanticBaseSettingsSource, ...]:
        """Customize the settings sources and their priority order.

        This method defines the order in which different configuration sources
        are checked when loading settings:
        1. Constructor-provided values
        2. YAML configuration file
        3. Environment variables

        Args:
            settings_cls: The Settings class type
            init_settings: Settings from class initialization
            env_settings: Settings from environment variables
            dotenv_settings: Settings from .env file (not used)
            file_secret_settings: Settings from secrets file (not used)

        Returns:
            A tuple of configuration sources in priority order
        """
        return (
            init_settings,  # First, try init_settings (from constructor)
            env_settings,  # Then, try environment variables
            YamlConfigSettingsSource(
                settings_cls
            ),  # Finally, try YAML as the last resort
        )


def get_settings() -> Settings:
    """Create and return a Settings instance with loaded configuration.

    Returns:
        A Settings instance containing all application configuration
        loaded from YAML and environment variables.
    """
    return Settings()

將main.py 重新命名為 frontend.py ，其內容

import gradio as gr
import requests
import base64
from pathlib import Path
from typing import List, Dict, Any
from settings import get_settings, DEFAULT_SYSTEM_PROMPT

settings = get_settings()

IMAGE_SUFFIX_MIME_MAP = {
    ".png": "image/png",
    ".jpg": "image/jpeg",
    ".jpeg": "image/jpeg",
    ".heic": "image/heic",
    ".heif": "image/heif",
    ".webp": "image/webp",
}
DOCUMENT_SUFFIX_MIME_MAP = {
    ".pdf": "application/pdf",
}


def get_mime_type(filepath: str) -> str:
    """Get the MIME type for a file based on its extension.

    Args:
        filepath: Path to the file.

    Returns:
        str: The MIME type of the file.

    Raises:
        ValueError: If the file type is not supported.
    """
    filepath = Path(filepath)
    suffix = filepath.suffix

    # modify ".jpg" suffix to ".jpeg" to unify the mime type
    suffix = suffix if suffix != ".jpg" else ".jpeg"

    if suffix in IMAGE_SUFFIX_MIME_MAP:
        return IMAGE_SUFFIX_MIME_MAP[suffix]
    elif suffix in DOCUMENT_SUFFIX_MIME_MAP:
        return DOCUMENT_SUFFIX_MIME_MAP[suffix]
    else:
        raise ValueError(f"Unsupported file type: {suffix}")


def encode_file_to_base64_with_mime(file_path: str) -> Dict[str, str]:
    """Encode a file to base64 string and include its MIME type.

    Args:
        file_path: Path to the file to encode.

    Returns:
        Dict[str, str]: Dictionary with 'data' and 'mime_type' keys.
    """
    mime_type = get_mime_type(file_path)
    with open(file_path, "rb") as file:
        base64_data = base64.b64encode(file.read()).decode("utf-8")

    return {"data": base64_data, "mime_type": mime_type}


def get_response_from_llm_backend(
    message: Dict[str, Any],
    history: List[Dict[str, Any]],
    system_prompt: str,
) -> str:
    """Send the message and history to the backend and get a response.

    Args:
        message: Dictionary containing the current message with 'text' and optional 'files' keys.
        history: List of previous message dictionaries in the conversation.
        system_prompt: The system prompt to be sent to the backend.

    Returns:
        str: The text response from the backend service.
    """

    # Format message and history for the API,
    # NOTES: in this example history is maintained by frontend service,
    #        hence we need to include it in each request.
    #        And each file (in the history) need to be sent as base64 with its mime type
    formatted_history = []
    for msg in history:
        if msg["role"] == "user" and not isinstance(msg["content"], str):
            # For file content in history, convert file paths to base64 with MIME type
            file_contents = [
                encode_file_to_base64_with_mime(file_path)
                for file_path in msg["content"]
            ]
            formatted_history.append({"role": msg["role"], "content": file_contents})
        else:
            formatted_history.append({"role": msg["role"], "content": msg["content"]})

    # Extract files and convert to base64 with MIME type
    files_with_mime = []
    if uploaded_files := message.get("files", []):
        for file_path in uploaded_files:
            files_with_mime.append(encode_file_to_base64_with_mime(file_path))

    # Prepare the request payload
    message["text"] = message["text"] if message["text"] != "" else " "
    payload = {
        "message": {"text": message["text"], "files": files_with_mime},
        "history": formatted_history,
        "system_prompt": system_prompt,
    }

    # Send request to backend
    try:
        response = requests.post(settings.BACKEND_URL, json=payload)
        response.raise_for_status()  # Raise exception for HTTP errors

        result = response.json()
        if error := result.get("error"):
            return f"Error: {error}"

        return result.get("response", "No response received from backend")
    except requests.exceptions.RequestException as e:
        return f"Error connecting to backend service: {str(e)}"


if __name__ == "__main__":
    demo = gr.ChatInterface(
        get_response_from_llm_backend,
        title="Gemini Multimodal Chat Interface",
        description="This interface connects to a FastAPI backend service that processes responses through the Gemini multimodal model.",
        type="messages",
        multimodal=True,
        textbox=gr.MultimodalTextbox(file_count="multiple"),
        additional_inputs=[
            gr.Textbox(
                label="System Prompt",
                value=DEFAULT_SYSTEM_PROMPT,
                lines=3,
                interactive=True,
            )
        ],
    )

    demo.launch(
        server_name="0.0.0.0",
        server_port=8080,
    )

執行 frontend.py

 uv run frontend.py

建立 backend.py ，其內容

import base64
from fastapi import FastAPI, Body
from google.genai.types import Content, Part
from google.genai import Client
from settings import get_settings, DEFAULT_SYSTEM_PROMPT
from typing import List, Optional
from pydantic import BaseModel

app = FastAPI(title="Gemini Multimodal Service")

settings = get_settings()
GENAI_CLIENT = Client(
    location=settings.VERTEXAI_LOCATION,
    project=settings.VERTEXAI_PROJECT_ID,
    vertexai=True,
)
GEMINI_MODEL_NAME = "gemini-2.0-flash-001"


class FileData(BaseModel):
    data: str
    mime_type: str

class Message(BaseModel):
    role: str
    content: str | List[FileData]

class LastUserMessage(BaseModel):
    text: str
    files: List[FileData] = []

class ChatRequest(BaseModel):
    message: LastUserMessage
    history: List[Message]
    system_prompt: str = DEFAULT_SYSTEM_PROMPT

class ChatResponse(BaseModel):
    response: str
    error: Optional[str] = None


def handle_multimodal_data(file_data: FileData) -> Part:
    """Converts Multimodal data to a Google Gemini Part object.

    Args:
        file_data: FileData object with base64 data and MIME type.

    Returns:
        Part: A Google Gemini Part object containing the file data.
    """
    data = base64.b64decode(file_data.data)  # decode base64 string to bytes
    return Part.from_bytes(data=data, mime_type=file_data.mime_type)


def format_message_history_to_gemini_standard(
    message_history: List[Message],
) -> List[Content]:
    """Converts message history format to Google Gemini Content format.

    Args:
        message_history: List of message objects from the chat history.
            Each message contains 'role' and 'content' attributes.

    Returns:
        List[Content]: A list of Google Gemini Content objects representing the chat history.

    Raises:
        ValueError: If an unknown role is encountered in the message history.
    """
    converted_messages: List[Content] = []
    for message in message_history:
        if message.role == "assistant":
            converted_messages.append(
                Content(role="model", parts=[Part.from_text(text=message.content)])
            )
        elif message.role == "user":
            # Text-only messages
            if isinstance(message.content, str):
                converted_messages.append(
                    Content(role="user", parts=[Part.from_text(text=message.content)])
                )

            # Messages with files
            elif isinstance(message.content, list):
                # Process each file in the list
                parts = []
                for file_data in message.content:
                    parts.append(handle_multimodal_data(file_data))

                # Add the parts to a Content object
                if parts:
                    converted_messages.append(Content(role="user", parts=parts))

            else:
                raise ValueError(f"Unexpected content format: {type(message.content)}")

        else:
            raise ValueError(f"Unknown role: {message.role}")

    return converted_messages


@app.post("/chat", response_model=ChatResponse)
async def chat(
    request: ChatRequest = Body(...),
) -> ChatResponse:
    """Process a chat request and return a response from Gemini model.

    Args:
        request: The chat request containing message and history.

    Returns:
        ChatResponse: The model's response to the chat request.
    """
    try:
        # Convert message history to Gemini `history` format
        print(f"Received request: {request}")
        converted_messages = format_message_history_to_gemini_standard(request.history)

        # Create chat model
        chat_model = GENAI_CLIENT.chats.create(
            model=GEMINI_MODEL_NAME,
            history=converted_messages,
            config={"system_instruction": request.system_prompt},
        )

        # Prepare multimodal content
        content_parts = []

        # Handle any base64 encoded files in the current message
        if request.message.files:
            for file_data in request.message.files:
                content_parts.append(handle_multimodal_data(file_data))

        # Add text content
        content_parts.append(Part.from_text(text=request.message.text))

        # Send message to Gemini
        response = chat_model.send_message(content_parts)
        print(f"Generated response: {response}")

        return ChatResponse(response=response.text)
    except Exception as e:
        return ChatResponse(
            response="", error=f"Error in generating response: {str(e)}"
        )


if __name__ == "__main__":
    import uvicorn

    uvicorn.run(app, host="0.0.0.0", port=8081)

執行

uv run backend.py

建立 supervisord.conf ，其內容

[supervisord]
nodaemon=true
user=root
logfile=/dev/stdout
logfile_maxbytes=0
pidfile=/var/run/supervisord.pid

[program:backend]
command=uv run backend.py
directory=/app
autostart=true
autorestart=true
stdout_logfile=/dev/stdout
stdout_logfile_maxbytes=0
stderr_logfile=/dev/stderr
stderr_logfile_maxbytes=0
startsecs=10
startretries=3

[program:frontend]
command=uv run frontend.py
directory=/app
autostart=true
autorestart=true
stdout_logfile=/dev/stdout
stdout_logfile_maxbytes=0
stderr_logfile=/dev/stderr
stderr_logfile_maxbytes=0
startsecs=10
startretries=3

建立 Dockerfile ，其內容

FROM python:3.12-slim
COPY --from=ghcr.io/astral-sh/uv:0.6.6 /uv /uvx /bin/

RUN apt-get update && apt-get install -y \
    supervisor curl \
    && rm -rf /var/lib/apt/lists/*

ADD . /app
WORKDIR /app

RUN uv sync --frozen

EXPOSE 8080

# Copy supervisord configuration
COPY supervisord.conf /etc/supervisor/conf.d/supervisord.conf

ENV PYTHONUNBUFFERED=1

ENTRYPOINT ["/usr/bin/supervisord", "-c", "/etc/supervisor/conf.d/supervisord.conf"]

部署至 Cloud Run

gcloud run deploy --source . \
                  --env-vars-file settings.yaml \
                  --port 8080 \
                  --region us-central1

執行結果

總結

起了一個gemini 多模態助理 ，上傳5.3M的gemini-for-google-workspace-prompting-guide-101.pdf
真的可以總結pdf 的重點，還不錯呀

使用Gradio構建介面，還可以下 System Prompt

參考

https://codelabs.developers.google.com/devsite/codelabs/gemini-multimodal-chat-assistant-python?hl=zh_tw