Tác nhân AI bền vững với Gemini và Temporal

temporal server start-dev

pip install temporalio google-genai httpx pydantic python-dotenv

echo "GOOGLE_API_KEY=your-api-key-here" > .env

from temporalio import workflow

with workflow.unsafe.imports_passed_through():
    import pydantic_core  # noqa: F401
    import annotated_types  # noqa: F401

    import httpx
    from pydantic import BaseModel, Field
    from google import genai
    from google.genai import types

SYSTEM_INSTRUCTIONS = """
You are a helpful agent that can use tools to help the user.
You will be given an input from the user and a list of tools to use.
You may or may not need to use the tools to satisfy the user ask.
If no tools are needed, respond in haikus.
"""

import json

NWS_API_BASE = "https://api.weather.gov"
USER_AGENT = "weather-app/1.0"


class GetWeatherAlertsRequest(BaseModel):
    """Request model for getting weather alerts."""

    state: str = Field(description="Two-letter US state code (e.g. CA, NY)")


async def get_weather_alerts(request: GetWeatherAlertsRequest) -> str:
    """Get weather alerts for a US state.

    Args:
        request: The request object containing:
            - state: Two-letter US state code (e.g. CA, NY)
    """
    headers = {"User-Agent": USER_AGENT, "Accept": "application/geo+json"}
    url = f"{NWS_API_BASE}/alerts/active/area/{request.state}"

    async with httpx.AsyncClient() as client:
        response = await client.get(url, headers=headers, timeout=5.0)
        response.raise_for_status()
        return json.dumps(response.json())

class GetLocationRequest(BaseModel):
    """Request model for getting location info from an IP address."""

    ipaddress: str = Field(description="An IP address")


async def get_ip_address() -> str:
    """Get the public IP address of the current machine."""
    async with httpx.AsyncClient() as client:
        response = await client.get("https://icanhazip.com")
        response.raise_for_status()
        return response.text.strip()


async def get_location_info(request: GetLocationRequest) -> str:
    """Get the location information for an IP address including city, state, and country.

    Args:
        request: The request object containing:
            - ipaddress: An IP address to look up
    """
    async with httpx.AsyncClient() as client:
        response = await client.get(f"http://ip-api.com/json/{request.ipaddress}")
        response.raise_for_status()
        result = response.json()
        return f"{result['city']}, {result['regionName']}, {result['country']}"

from typing import Any, Awaitable, Callable

ToolHandler = Callable[..., Awaitable[Any]]


def get_handler(tool_name: str) -> ToolHandler:
    """Get the handler function for a given tool name."""
    if tool_name == "get_location_info":
        return get_location_info
    if tool_name == "get_ip_address":
        return get_ip_address
    if tool_name == "get_weather_alerts":
        return get_weather_alerts
    raise ValueError(f"Unknown tool name: {tool_name}")


def get_tools() -> types.Tool:
    """Get the Tool object containing all available function declarations.

    Uses FunctionDeclaration.from_callable_with_api_option() from the Google GenAI SDK
    to generate tool definitions from the handler functions.
    """
    return types.Tool(
        function_declarations=[
            types.FunctionDeclaration.from_callable_with_api_option(
                callable=get_weather_alerts, api_option="GEMINI_API"
            ),
            types.FunctionDeclaration.from_callable_with_api_option(
                callable=get_location_info, api_option="GEMINI_API"
            ),
            types.FunctionDeclaration.from_callable_with_api_option(
                callable=get_ip_address, api_option="GEMINI_API"
            ),
        ]
    )

import os
from dataclasses import dataclass

from temporalio import activity

@dataclass
class GeminiChatRequest:
    """Request parameters for a Gemini chat completion."""

    model: str
    system_instruction: str
    contents: list[types.Content]
    tools: list[types.Tool]


@dataclass
class GeminiChatResponse:
    """Response from a Gemini chat completion."""

    text: str | None
    function_calls: list[dict[str, Any]]
    raw_parts: list[types.Part]


@activity.defn
async def generate_content(request: GeminiChatRequest) -> GeminiChatResponse:
    """Execute a Gemini chat completion with tool support."""
    api_key = os.environ.get("GOOGLE_API_KEY")
    if not api_key:
        raise ValueError("GOOGLE_API_KEY environment variable is not set")
    client = genai.Client(
        api_key=api_key,
        http_options=types.HttpOptions(
            retry_options=types.HttpRetryOptions(attempts=1),
        ),
    )

    config = types.GenerateContentConfig(
        system_instruction=request.system_instruction,
        tools=request.tools,
        automatic_function_calling=types.AutomaticFunctionCallingConfig(disable=True),
    )

    response = await client.aio.models.generate_content(
        model=request.model,
        contents=request.contents,
        config=config,
    )

    function_calls = []
    raw_parts = []
    text_parts = []

    if response.candidates and response.candidates[0].content:
        for part in response.candidates[0].content.parts:
            raw_parts.append(part)
            if part.function_call:
                function_calls.append(
                    {
                        "name": part.function_call.name,
                        "args": dict(part.function_call.args) if part.function_call.args else {},
                    }
                )
            elif part.text:
                text_parts.append(part.text)

    text = "".join(text_parts) if text_parts and not function_calls else None

    return GeminiChatResponse(
        text=text,
        function_calls=function_calls,
        raw_parts=raw_parts,
    )

import inspect
from collections.abc import Sequence

from temporalio.common import RawValue

@activity.defn(dynamic=True)
async def dynamic_tool_activity(args: Sequence[RawValue]) -> dict:
    """Execute a tool dynamically based on the activity name."""
    tool_name = activity.info().activity_type
    tool_args = activity.payload_converter().from_payload(args[0].payload, dict)
    activity.logger.info(f"Running dynamic tool '{tool_name}' with args: {tool_args}")

    handler = get_handler(tool_name)

    if not inspect.iscoroutinefunction(handler):
        raise TypeError("Tool handler must be async (awaitable).")

    sig = inspect.signature(handler)
    params = list(sig.parameters.values())

    if len(params) == 0:
        result = await handler()
    else:
        param = params[0]
        param_name = param.name
        ann = param.annotation

        if isinstance(ann, type) and issubclass(ann, BaseModel):
            nested_args = tool_args.get(param_name, tool_args)
            result = await handler(ann(**nested_args))
        else:
            result = await handler(**tool_args)

    activity.logger.info(f"Tool '{tool_name}' result: {result}")
    return result

from datetime import timedelta

@workflow.defn
class AgentWorkflow:
    """Agentic loop workflow that uses Gemini for LLM calls and executes tools."""

    @workflow.run
    async def run(self, input: str) -> str:
        contents: list[types.Content] = [
            types.Content(role="user", parts=[types.Part.from_text(text=input)])
        ]

        tools = [get_tools()]

        while True:
            result = await workflow.execute_activity(
                generate_content,
                GeminiChatRequest(
                    model="gemini-3-flash-preview",
                    system_instruction=SYSTEM_INSTRUCTIONS,
                    contents=contents,
                    tools=tools,
                ),
                start_to_close_timeout=timedelta(seconds=60),
            )

            if result.function_calls:
                # Sending the complete raw_parts here ensures Gemini 3 thought
                # signatures are propagated correctly.
                contents.append(types.Content(role="model", parts=result.raw_parts))

                for function_call in result.function_calls:
                    tool_result = await self._handle_function_call(function_call)

                    contents.append(
                        types.Content(
                            role="user",
                            parts=[
                                types.Part.from_function_response(
                                    name=function_call["name"],
                                    response={"result": tool_result},
                                )
                            ],
                        )
                    )
            else:
                return result.text
            # Leave this in place. You will un-comment it during a durability
            # test later on.
            # await asyncio.sleep(10)

    async def _handle_function_call(self, function_call: dict) -> str:
        """Execute a tool via dynamic activity and return the result."""
        tool_name = function_call["name"]
        tool_args = function_call.get("args", {})

        result = await workflow.execute_activity(
            tool_name,
            tool_args,
            start_to_close_timeout=timedelta(seconds=30),
        )

        return result

import asyncio
from concurrent.futures import ThreadPoolExecutor

from dotenv import load_dotenv
from temporalio.client import Client
from temporalio.contrib.pydantic import pydantic_data_converter
from temporalio.envconfig import ClientConfig
from temporalio.worker import Worker

async def main():
    config = ClientConfig.load_client_connect_config()
    config.setdefault("target_host", "localhost:7233")
    client = await Client.connect(
        **config,
        data_converter=pydantic_data_converter,
    )

    worker = Worker(
        client,
        task_queue="gemini-agent-python-task-queue",
        workflows=[
            AgentWorkflow,
        ],
        activities=[
            generate_content,
            dynamic_tool_activity,
        ],
        activity_executor=ThreadPoolExecutor(max_workers=10),
    )
    await worker.run()


if __name__ == "__main__":
    load_dotenv()
    asyncio.run(main())

import asyncio
import sys
import uuid

from temporalio.client import Client
from temporalio.contrib.pydantic import pydantic_data_converter


async def main():
    client = await Client.connect(
        "localhost:7233",
        data_converter=pydantic_data_converter,
    )

    query = sys.argv[1] if len(sys.argv) > 1 else "Tell me about recursion"

    result = await client.execute_workflow(
        "AgentWorkflow",
        query,
        id=f"gemini-agent-id-{uuid.uuid4()}",
        task_queue="gemini-agent-python-task-queue",
    )
    print(f"\nResult:\n{result}")


if __name__ == "__main__":
    asyncio.run(main())

temporal server start-dev

python -m durable_agent_worker

python -m start_workflow "are there any weather alerts for where I am?"

python -m start_workflow "are there any weather alerts for New York?"
python -m start_workflow "where am I?"
python -m start_workflow "what is my ip address?"
python -m start_workflow "tell me a joke"