Compartilhar via


Instruções: Criar seu primeiro processo

Aviso

O Framework de Processo de Kernel Semântico é experimental, ainda em desenvolvimento e está sujeito a alterações.

Visão geral

O Semantic Kernel Process Framework é um SDK de orquestração avançado projetado para simplificar o desenvolvimento e a execução de processos integrados à IA. Se você estiver gerenciando fluxos de trabalho simples ou sistemas complexos, essa estrutura permite que você defina uma série de etapas que podem ser executadas de maneira estruturada, aprimorando os recursos do aplicativo com facilidade e flexibilidade.

Criado para extensibilidade, o Process Framework dá suporte a diversos padrões operacionais, como execução sequencial, processamento paralelo, configurações de fan-in e fan-out e até mesmo estratégias de redução de mapa. Essa adaptabilidade o torna adequado para uma variedade de aplicativos do mundo real, particularmente aqueles que exigem fluxos de trabalho inteligentes e de várias etapas.

Introdução

A Estrutura de Processo de Kernel Semântico pode ser usada para infundir a IA em praticamente qualquer processo de negócios que você possa pensar. Como um exemplo ilustrativo para começar, vamos examinar a criação de um processo para gerar documentação para um novo produto.

Antes de começarmos, verifique se você tem os pacotes de Kernel Semântico necessários instalados:

// Install the Semantic Kernel Process Framework Local Runtime package
dotnet add package Microsoft.SemanticKernel.Process.LocalRuntime --version 1.46.0-alpha
// or
// Install the Semantic Kernel Process Framework Dapr Runtime package
dotnet add package Microsoft.SemanticKernel.Process.Runtime.Dapr --version 1.46.0-alpha

pip install semantic-kernel==1.20.0

Exemplo ilustrativo: gerando documentação para um novo produto

Neste exemplo, utilizaremos a Estrutura de Processo de Kernel Semântico para desenvolver um processo automatizado para criar documentação para um novo produto. Esse processo começará simples e evoluirá à medida que avançamos para abranger cenários mais realistas.

Começaremos modelando o processo de documentação com um fluxo muito básico:

  1. GatherProductInfoStep: colete informações sobre o produto.
  2. GenerateDocumentationStep: peça a um LLM para gerar a documentação das informações reunidas na etapa 1.
  3. PublishDocumentationStep: Publique a documentação.

diagrama de fluxo do nosso primeiro processo: A[Documentação do Recurso de Solicitação] --> B[Pedir à LLM para escrever a documentação] --> C[Publicar documentação para o público]

Agora que entendemos nossos processos, vamos compilá-los.

Definir as etapas do processo

Cada etapa de um Processo é definida por uma classe que herda de nossa classe de etapa base. Para esse processo, temos três etapas:

using Microsoft.SemanticKernel.ChatCompletion;
using Microsoft.SemanticKernel;

// A process step to gather information about a product
public class GatherProductInfoStep: KernelProcessStep
{
    [KernelFunction]
    public string GatherProductInformation(string productName)
    {
        Console.WriteLine($"{nameof(GatherProductInfoStep)}:\n\tGathering product information for product named {productName}");

        // For example purposes we just return some fictional information.
        return
            """
            Product Description:
            GlowBrew is a revolutionary AI driven coffee machine with industry leading number of LEDs and programmable light shows. The machine is also capable of brewing coffee and has a built in grinder.

            Product Features:
            1. **Luminous Brew Technology**: Customize your morning ambiance with programmable LED lights that sync with your brewing process.
            2. **AI Taste Assistant**: Learns your taste preferences over time and suggests new brew combinations to explore.
            3. **Gourmet Aroma Diffusion**: Built-in aroma diffusers enhance your coffee's scent profile, energizing your senses before the first sip.

            Troubleshooting:
            - **Issue**: LED Lights Malfunctioning
                - **Solution**: Reset the lighting settings via the app. Ensure the LED connections inside the GlowBrew are secure. Perform a factory reset if necessary.
            """;
    }
}

// A process step to generate documentation for a product
public class GenerateDocumentationStep : KernelProcessStep<GeneratedDocumentationState>
{
    private GeneratedDocumentationState _state = new();

    private string systemPrompt =
            """
            Your job is to write high quality and engaging customer facing documentation for a new product from Contoso. You will be provide with information
            about the product in the form of internal documentation, specs, and troubleshooting guides and you must use this information and
            nothing else to generate the documentation. If suggestions are provided on the documentation you create, take the suggestions into account and
            rewrite the documentation. Make sure the product sounds amazing.
            """;

    // Called by the process runtime when the step instance is activated. Use this to load state that may be persisted from previous activations.
    override public ValueTask ActivateAsync(KernelProcessStepState<GeneratedDocumentationState> state)
    {
        this._state = state.State!;
        this._state.ChatHistory ??= new ChatHistory(systemPrompt);

        return base.ActivateAsync(state);
    }

    [KernelFunction]
    public async Task GenerateDocumentationAsync(Kernel kernel, KernelProcessStepContext context, string productInfo)
    {
        Console.WriteLine($"[{nameof(GenerateDocumentationStep)}]:\tGenerating documentation for provided productInfo...");

        // Add the new product info to the chat history
        this._state.ChatHistory!.AddUserMessage($"Product Info:\n{productInfo.Title} - {productInfo.Content}");

        // Get a response from the LLM
        IChatCompletionService chatCompletionService = kernel.GetRequiredService<IChatCompletionService>();
        var generatedDocumentationResponse = await chatCompletionService.GetChatMessageContentAsync(this._state.ChatHistory!);

        DocumentInfo generatedContent = new()
        {
            Id = Guid.NewGuid().ToString(),
            Title = $"Generated document - {productInfo.Title}",
            Content = generatedDocumentationResponse.Content!,
        };

        this._state!.LastGeneratedDocument = generatedContent;

        await context.EmitEventAsync("DocumentationGenerated", generatedContent);
    }

    public class GeneratedDocumentationState
    {
        public DocumentInfo LastGeneratedDocument { get; set; } = new();
        public ChatHistory? ChatHistory { get; set; }
    }
}

// A process step to publish documentation
public class PublishDocumentationStep : KernelProcessStep
{
    [KernelFunction]
    public DocumentInfo PublishDocumentation(DocumentInfo document)
    {
        // For example purposes we just write the generated docs to the console
        Console.WriteLine($"[{nameof(PublishDocumentationStep)}]:\tPublishing product documentation approved by user: \n{document.Title}\n{document.Content}");
        return document;
    }
}

// Custom classes must be serializable
public class DocumentInfo
{
    public string Id { get; set; } = string.Empty;
    public string Title { get; set; } = string.Empty;
    public string Content { get; set; } = string.Empty;
}

O código acima define as três etapas necessárias para nosso processo. Há alguns pontos a destacar aqui:

  • No Kernel Semântico, um KernelFunction define um bloco de código que pode ser executado por código nativo ou por um LLM. No caso da estrutura de processo, KernelFunctions são os membros invocados de uma etapa e cada etapa requer que pelo menos um KernelFunction seja definido.
  • O Process Framework tem suporte para etapas com estado e sem estado. As etapas com estado verificam automaticamente seu progresso e mantêm o estado em várias invocações. O GenerateDocumentationStep fornece um exemplo disso em que a classe GeneratedDocumentationState é usada para persistir o objeto ChatHistory e LastGeneratedDocument.
  • As etapas podem emitir eventos manualmente chamando EmitEventAsync no objeto KernelProcessStepContext. Para obter uma instância de KernelProcessStepContext basta adicioná-la como um parâmetro em seu KernelFunction e a estrutura a injetará automaticamente.
import asyncio
from typing import ClassVar

from pydantic import BaseModel, Field

from semantic_kernel import Kernel
from semantic_kernel.connectors.ai.chat_completion_client_base import ChatCompletionClientBase
from semantic_kernel.connectors.ai.open_ai import AzureChatCompletion
from semantic_kernel.contents import ChatHistory
from semantic_kernel.functions import kernel_function
from semantic_kernel.processes import ProcessBuilder
from semantic_kernel.processes.kernel_process import KernelProcessStep, KernelProcessStepContext, KernelProcessStepState
from semantic_kernel.processes.local_runtime import KernelProcessEvent, start


# A process step to gather information about a product
class GatherProductInfoStep(KernelProcessStep):
    @kernel_function
    def gather_product_information(self, product_name: str) -> str:
        print(f"{GatherProductInfoStep.__name__}\n\t Gathering product information for Product Name: {product_name}")

        return """
Product Description:

GlowBrew is a revolutionary AI driven coffee machine with industry leading number of LEDs and 
programmable light shows. The machine is also capable of brewing coffee and has a built in grinder.

Product Features:
1. **Luminous Brew Technology**: Customize your morning ambiance with programmable LED lights that sync 
    with your brewing process.
2. **AI Taste Assistant**: Learns your taste preferences over time and suggests new brew combinations 
    to explore.
3. **Gourmet Aroma Diffusion**: Built-in aroma diffusers enhance your coffee's scent profile, energizing 
    your senses before the first sip.

Troubleshooting:
- **Issue**: LED Lights Malfunctioning
    - **Solution**: Reset the lighting settings via the app. Ensure the LED connections inside the 
        GlowBrew are secure. Perform a factory reset if necessary.
        """


# A sample step state model for the GenerateDocumentationStep
class GeneratedDocumentationState(BaseModel):
    """State for the GenerateDocumentationStep."""

    chat_history: ChatHistory | None = None


# A process step to generate documentation for a product
class GenerateDocumentationStep(KernelProcessStep[GeneratedDocumentationState]):
    state: GeneratedDocumentationState = Field(default_factory=GeneratedDocumentationState)

    system_prompt: ClassVar[str] = """
Your job is to write high quality and engaging customer facing documentation for a new product from Contoso. You will 
be provided with information about the product in the form of internal documentation, specs, and troubleshooting guides 
and you must use this information and nothing else to generate the documentation. If suggestions are provided on the 
documentation you create, take the suggestions into account and rewrite the documentation. Make sure the product 
sounds amazing.
"""

    async def activate(self, state: KernelProcessStepState[GeneratedDocumentationState]):
        self.state = state.state
        if self.state.chat_history is None:
            self.state.chat_history = ChatHistory(system_message=self.system_prompt)
        self.state.chat_history

    @kernel_function
    async def generate_documentation(
        self, context: KernelProcessStepContext, product_info: str, kernel: Kernel
    ) -> None:
        print(f"{GenerateDocumentationStep.__name__}\n\t Generating documentation for provided product_info...")

        self.state.chat_history.add_user_message(f"Product Information:\n{product_info}")

        chat_service, settings = kernel.select_ai_service(type=ChatCompletionClientBase)
        assert isinstance(chat_service, ChatCompletionClientBase)  # nosec

        response = await chat_service.get_chat_message_content(chat_history=self.state.chat_history, settings=settings)

        await context.emit_event(process_event="documentation_generated", data=str(response))


# A process step to publish documentation
class PublishDocumentationStep(KernelProcessStep):
    @kernel_function
    async def publish_documentation(self, docs: str) -> None:
        print(f"{PublishDocumentationStep.__name__}\n\t Publishing product documentation:\n\n{docs}")

O código acima define as três etapas necessárias para nosso processo. Há alguns pontos a destacar aqui:

  • No Kernel Semântico, um KernelFunction define um bloco de código que pode ser executado por código nativo ou por um LLM. No caso da estrutura de processo, KernelFunctions são os membros invocados de uma etapa e cada etapa requer que pelo menos um KernelFunction seja definido.
  • O Process Framework tem suporte para etapas com estado e sem estado. As etapas com estado verificam automaticamente seu progresso e mantêm o estado em várias invocações. O GenerateDocumentationStep fornece um exemplo disso em que a classe GeneratedDocumentationState é usada para persistir o objeto ChatHistory.
  • As etapas podem emitir eventos manualmente chamando emit_event no objeto KernelProcessStepContext. Para obter uma instância de KernelProcessStepContext basta adicioná-la como um parâmetro em seu KernelFunction e a estrutura a injetará automaticamente.

Definir o fluxo de processo

// Create the process builder
ProcessBuilder processBuilder = new("DocumentationGeneration");

// Add the steps
var infoGatheringStep = processBuilder.AddStepFromType<GatherProductInfoStep>();
var docsGenerationStep = processBuilder.AddStepFromType<GenerateDocumentationStep>();
var docsPublishStep = processBuilder.AddStepFromType<PublishDocumentationStep>();

// Orchestrate the events
processBuilder
    .OnInputEvent("Start")
    .SendEventTo(new(infoGatheringStep));

infoGatheringStep
    .OnFunctionResult()
    .SendEventTo(new(docsGenerationStep));

docsGenerationStep
    .OnFunctionResult()
    .SendEventTo(new(docsPublishStep));

Há algumas coisas acontecendo aqui, então vamos dividi-lo passo a passo.

  1. Criar o construtor: os processos usam um padrão de construtor para simplificar a configuração de tudo. O construtor fornece métodos para gerenciar as etapas dentro de um processo e para gerenciar o ciclo de vida do processo.

  2. Adicione as etapas: as etapas podem ser adicionadas ao processo chamando o método AddStepFromType do construtor. Isso permite que o Process Framework gerencie o ciclo de vida das etapas instanciando instâncias conforme necessário. Nesse caso, adicionamos três etapas ao processo e criamos uma variável para cada uma delas. Essas variáveis nos dão um identificador para a instância exclusiva de cada etapa, que podemos usar a seguir para definir a orquestração de eventos.

  3. Orquestrar os eventos: é aí que o roteamento de eventos de etapa para etapa é definido. Nesse caso, temos as seguintes rotas:

    • Quando um evento externo com id = Start é enviado ao processo, esse evento e seus dados associados serão enviados para a etapa infoGatheringStep.
    • Quando o infoGatheringStep terminar de ser executado, envie o objeto retornado para a etapa docsGenerationStep.
    • Por fim, quando o docsGenerationStep terminar de ser executado, envie o objeto retornado para a etapa docsPublishStep.

Dica

Roteamento de Eventos no Process Framework: Talvez você esteja se perguntando como os eventos enviados às etapas são roteado para KernelFunctions dentro da etapa. No código acima, cada etapa definiu apenas um único KernelFunction e cada KernelFunction tem apenas um único parâmetro (além de Kernel e o contexto de etapa que são especiais, mais sobre isso posteriormente). Quando o evento que contém a documentação gerada é enviado para o docsPublishStep ele será passado para o parâmetro document do PublishDocumentation KernelFunction da etapa docsGenerationStep porque não há outra opção. No entanto, as etapas podem ter várias Funções de Kernel, e as Funções de Kernel podem ter vários parâmetros. Nesses cenários avançados, é necessário especificar a função e o parâmetro de destino.

# Create the process builder
process_builder = ProcessBuilder(name="DocumentationGeneration")

# Add the steps
info_gathering_step = process_builder.add_step(GatherProductInfoStep)
docs_generation_step = process_builder.add_step(GenerateDocumentationStep)
docs_publish_step = process_builder.add_step(PublishDocumentationStep)

# Orchestrate the events
process_builder.on_input_event("Start").send_event_to(target=info_gathering_step)

info_gathering_step.on_function_result().send_event_to(
    target=docs_generation_step, function_name="generate_documentation", parameter_name="product_info"
)

docs_generation_step.on_event("documentation_generated").send_event_to(target=docs_publish_step)

# Configure the kernel with an AI service and connection details, if necessary
kernel = Kernel()
kernel.add_service(AzureChatCompletion())

# Build the process
kernel_process = process_builder.build()

Há algumas coisas acontecendo aqui, então vamos dividi-lo passo a passo.

  1. Criar o construtor: os processos usam um padrão de construtor para simplificar a configuração de tudo. O construtor fornece métodos para gerenciar as etapas dentro de um processo e para gerenciar o ciclo de vida do processo.

  2. Adicione as etapas: as etapas são incorporadas ao processo ao chamar o método do construtor add_step, o qual adiciona o tipo de etapa ao construtor. Isso permite que o Process Framework gerencie o ciclo de vida das etapas instanciando instâncias conforme necessário. Nesse caso, adicionamos três etapas ao processo e criamos uma variável para cada uma delas. Essas variáveis nos dão um identificador para a instância exclusiva de cada etapa, que podemos usar a seguir para definir a orquestração de eventos.

  3. Orquestrar os eventos: é aí que o roteamento de eventos de etapa para etapa é definido. Nesse caso, temos as seguintes rotas:

    • Quando um evento externo com id = Start é enviado para o processo, esse evento e seus dados associados serão enviados para o info_gathering_step.
    • Quando o info_gathering_step terminar de executar, envie o objeto retornado para o docs_generation_step.
    • Por fim, quando o docs_generation_step terminar de executar, envie o objeto retornado para o docs_publish_step.

Dica

Roteamento de Eventos no Process Framework: Talvez você esteja se perguntando como os eventos enviados às etapas são roteado para KernelFunctions dentro da etapa. No código acima, cada etapa definiu apenas um único KernelFunction e cada KernelFunction tem apenas um único parâmetro (além de Kernel e o contexto de etapa que são especiais, mais sobre isso posteriormente). Quando o evento que contém a documentação gerada for enviado para docs_publish_step, ele será passado para o parâmetro docs da função Kernel publish_documentation do docs_generation_step porque não há outra escolha. No entanto, as etapas podem ter várias Funções de Kernel, e as Funções de Kernel podem ter vários parâmetros. Nesses cenários avançados, é necessário especificar a função e o parâmetro de destino.

Compilar e executar o processo

// Configure the kernel with your LLM connection details
Kernel kernel = Kernel.CreateBuilder()
    .AddAzureOpenAIChatCompletion("myDeployment", "myEndpoint", "myApiKey")
    .Build();

// Build and run the process
var process = processBuilder.Build();
await process.StartAsync(kernel, new KernelProcessEvent { Id = "Start", Data = "Contoso GlowBrew" });

Criamos o processo e chamamos StartAsync para executá-lo. Nosso processo está esperando um evento externo inicial chamado Start para iniciar as coisas e, portanto, fornecemos isso também. A execução desse processo mostra a seguinte saída no Console:

GatherProductInfoStep: Gathering product information for product named Contoso GlowBrew
GenerateDocumentationStep: Generating documentation for provided productInfo
PublishDocumentationStep: Publishing product documentation:

# GlowBrew: Your Ultimate Coffee Experience Awaits!

Welcome to the world of GlowBrew, where coffee brewing meets remarkable technology! At Contoso, we believe that your morning ritual shouldn't just include the perfect cup of coffee but also a stunning visual experience that invigorates your senses. Our revolutionary AI-driven coffee machine is designed to transform your kitchen routine into a delightful ceremony.

## Unleash the Power of GlowBrew

### Key Features

- **Luminous Brew Technology**
  - Elevate your coffee experience with our cutting-edge programmable LED lighting. GlowBrew allows you to customize your morning ambiance, creating a symphony of colors that sync seamlessly with your brewing process. Whether you need a vibrant wake-up call or a soothing glow, you can set the mood for any moment!

- **AI Taste Assistant**
  - Your taste buds deserve the best! With the GlowBrew built-in AI taste assistant, the machine learns your unique preferences over time and curates personalized brew suggestions just for you. Expand your coffee horizons and explore delightful new combinations that fit your palate perfectly.

- **Gourmet Aroma Diffusion**
  - Awaken your senses even before that first sip! The GlowBrew comes equipped with gourmet aroma diffusers that enhance the scent profile of your coffee, diffusing rich aromas that fill your kitchen with the warm, inviting essence of freshly-brewed bliss.

### Not Just Coffee - An Experience

With GlowBrew, it's more than just making coffee-it's about creating an experience that invigorates the mind and pleases the senses. The glow of the lights, the aroma wafting through your space, and the exceptional taste meld into a delightful ritual that prepares you for whatever lies ahead.

## Troubleshooting Made Easy

While GlowBrew is designed to provide a seamless experience, we understand that technology can sometimes be tricky. If you encounter issues with the LED lights, we've got you covered:

- **LED Lights Malfunctioning?**
  - If your LED lights aren't working as expected, don't worry! Follow these steps to restore the glow:
    1. **Reset the Lighting Settings**: Use the GlowBrew app to reset the lighting settings.
    2. **Check Connections**: Ensure that the LED connections inside the GlowBrew are secure.
    3. **Factory Reset**: If you're still facing issues, perform a factory reset to rejuvenate your machine.

With GlowBrew, you not only brew the perfect coffee but do so with an ambiance that excites the senses. Your mornings will never be the same!

## Embrace the Future of Coffee

Join the growing community of GlowBrew enthusiasts today, and redefine how you experience coffee. With stunning visual effects, customized brewing suggestions, and aromatic enhancements, it's time to indulge in the delightful world of GlowBrew-where every cup is an adventure!

### Conclusion

Ready to embark on an extraordinary coffee journey? Discover the perfect blend of technology and flavor with Contoso's GlowBrew. Your coffee awaits!
# Configure the kernel with an AI service and connection details, if necessary
kernel = Kernel()
kernel.add_service(AzureChatCompletion())

# Build the process
kernel_process = process_builder.build()

# Start the process
async with await start(
    process=kernel_process,
    kernel=kernel,
    initial_event=KernelProcessEvent(id="Start", data="Contoso GlowBrew"),
) as process_context:
    _ = await process_context.get_state()

Criamos o processo e chamamos start com o gerenciador de contexto assíncrono para executá-lo. Nosso processo está esperando um evento externo inicial chamado Start para iniciar as coisas e, portanto, fornecemos isso também. A execução desse processo mostra a seguinte saída no Console:

GatherProductInfoStep
         Gathering product information for Product Name: Contoso GlowBrew
GenerateDocumentationStep
         Generating documentation for provided product_info...
PublishDocumentationStep
         Publishing product documentation:

# GlowBrew AI-Driven Coffee Machine: Elevate Your Coffee Experience

Welcome to the future of coffee enjoyment with GlowBrew, the AI-driven coffee machine that not only crafts the perfect cup but does so with a light show that brightens your day. Designed for coffee enthusiasts and tech aficionados alike, GlowBrew combines cutting-edge brewing technology with an immersive lighting experience to start every day on a bright note.

## Unleash the Power of Luminous Brew Technology

With GlowBrew, your mornings will never be dull. The industry-leading number of programmable LEDs offers endless possibilities for customizing your coffee-making ritual. Sync the light show with the brewing process to create a visually stimulating ambiance that transforms your kitchen into a vibrant café each morning.

## Discover New Flavor Dimensions with the AI Taste Assistant

Leave the traditional coffee routines behind and say hello to personalization sophistication. The AI Taste Assistant learns and adapts to your unique preferences over time. Whether you prefer a strong espresso or a light latte, the assistant suggests new brew combinations tailored to your palate, inviting you to explore a world of flavors you never knew existed.

## Heighten Your Senses with Gourmet Aroma Diffusion

The moment you step into the room, let the GlowBrew’s built-in aroma diffusers captivate your senses. This feature is designed to enrich your coffee’s scent profile, ensuring every cup you brew is a multi-sensory delight. Let the burgeoning aroma energize you before the very first sip.

## Troubleshooting Guide: LED Lights Malfunctioning

Occasionally, you might encounter an issue with the LED lights not functioning as intended. Here’s how to resolve it efficiently:

- **Reset Lighting Settings**: Start by using the GlowBrew app to reset the lighting configurations to their default state.
- **Check Connections**: Ensure that all LED connections inside your GlowBrew machine are secure and properly connected.
- **Perform a Factory Reset**: If the problem persists, perform a factory reset on your GlowBrew to restore all settings to their original state.

Experience the art of coffee making like never before with the GlowBrew AI-driven coffee machine. From captivating light shows to aromatic sensations, every feature is engineered to enhance your daily brew. Brew, savor, and glow with GlowBrew.

O que vem a seguir?

Nosso primeiro rascunho do processo de geração de documentação está funcionando, mas deixa muito a desejar. No mínimo, uma versão de produção precisaria:

  • Um agente de leitor de prova que classificará a documentação gerada e verificará se ela atende aos nossos padrões de qualidade e precisão.
  • Um processo de aprovação em que a documentação só é publicada depois que um humano a aprova (intervenção humana).