LLM服务架构重构文档

概述

本文档说明Prompt Optimizer的LLM服务架构重构,从单体Service转变为Provider-Adapter-Registry三层架构,实现了更高的模块化、可扩展性和可维护性。

重构目标

1. Provider抽象层: 统一各LLM提供商的元数据定义(能力、参数等)
2. Adapter统一接口: 规范不同SDK的调用方式,隔离SDK细节
3. 配置自包含: 配置包含完整元数据,不依赖运行时查找
4. 向后兼容: 传统配置自动转换,保持API签名不变

架构设计

三层架构图

graph TB
    subgraph "应用层"
        UI[UI组件]
        LLMService[LLMService]
    end

    subgraph "配置层"
        TextModelConfig[TextModelConfig<br/>自包含配置]
        ModelManager[ModelManager<br/>配置管理]
    end

    subgraph "Provider层"
        TextProvider[TextProvider<br/>提供商元数据]
        TextModel[TextModel<br/>模型元数据]
    end

    subgraph "Adapter层"
        Registry[TextAdapterRegistry<br/>适配器注册表]
        OpenAIAdapter[OpenAIAdapter]
        GeminiAdapter[GeminiAdapter]
        AnthropicAdapter[AnthropicAdapter]
    end

    subgraph "SDK层"
        OpenAISDK[OpenAI SDK]
        GeminiSDK[Google Generative AI]
        AnthropicSDK[Anthropic SDK]
    end

    UI --> LLMService
    LLMService --> ModelManager
    LLMService --> Registry

    ModelManager --> TextModelConfig
    TextModelConfig --> TextProvider
    TextModelConfig --> TextModel

    Registry --> OpenAIAdapter
    Registry --> GeminiAdapter
    Registry --> AnthropicAdapter

    OpenAIAdapter --> OpenAISDK
    GeminiAdapter --> GeminiSDK
    AnthropicAdapter --> AnthropicSDK

核心组件

1. TextProvider (Provider元数据)

interface TextProvider {
  id: string;                    // 'openai' | 'gemini' | 'anthropic'
  name: string;                  // 'OpenAI' | 'Google Gemini' | 'Anthropic'
  description: string;
  defaultBaseURL?: string;
  connectionSchema: ConnectionSchema;  // 连接参数定义
}

职责: 定义Provider的基本信息和连接要求

2. TextModel (Model元数据)

interface TextModel {
  id: string;                    // 'gpt-4o-mini' | 'gemini-2.0-flash-exp'
  name: string;
  description?: string;
  providerId: string;            // 归属Provider
  capabilities: {
    supportsStreaming: boolean;
    supportsTools: boolean;
    supportsReasoning: boolean;
    maxContextLength: number;
  };
  parameterDefinitions: ParameterDefinition[];
  defaultParameterValues: Record<string, any>;
}

职责: 定义Model的能力和参数

3. TextModelConfig (自包含配置)

interface TextModelConfig {
  id: string;
  name: string;
  enabled: boolean;
  providerMeta: TextProvider;     // 嵌入Provider元数据
  modelMeta: TextModel;            // 嵌入Model元数据
  connectionConfig: ConnectionConfig;  // 连接配置(apiKey, baseURL)
  paramOverrides: Record<string, any>; // 参数覆盖
}

特点:

• 自包含: 包含运行时所需的全部信息
• 类型安全: 完整TypeScript类型定义
• 元数据嵌入: 无需运行时查找Provider/Model

4. ITextProviderAdapter (Adapter接口)

interface ITextProviderAdapter {
  getProvider(): TextProvider;
  getModels(): TextModel[];
  getModelsAsync?(config: TextModelConfig): Promise<TextModel[]>;
  buildDefaultModel(modelId: string): TextModel;

  sendMessage(messages: Message[], config: TextModelConfig): Promise<LLMResponse>;
  sendMessageStream(messages: Message[], config: TextModelConfig, handlers: StreamHandlers): Promise<void>;
  sendMessageStreamWithTools?(messages: Message[], config: TextModelConfig, tools: ToolDefinition[], handlers: StreamHandlers): Promise<void>;
}

职责:

• 提供Provider和Model元数据
• 封装SDK调用逻辑
• 处理消息格式转换
• 保留错误堆栈

5. TextAdapterRegistry (注册表)

class TextAdapterRegistry implements ITextAdapterRegistry {
  private adapters: Map<string, ITextProviderAdapter>;
  private staticModelsCache: Map<string, TextModel[]>;

  getAdapter(providerId: string): ITextProviderAdapter;
  getAllProviders(): TextProvider[];
  getStaticModels(providerId: string): TextModel[];
  getDynamicModels(providerId: string, config: TextModelConfig): Promise<TextModel[]>;
  getModels(providerId: string, config?: TextModelConfig): Promise<TextModel[]>;
}

职责:

• 注册和管理所有Adapter实例
• 提供统一的Adapter查找接口
• 缓存静态模型列表
• 支持动态模型获取(OpenAI)

配置迁移流程

传统配置 → 新配置

flowchart LR
    A[传统ModelConfig] --> B{检测格式}
    B -->|Legacy| C[convertLegacyToTextModelConfigWithRegistry]
    B -->|New| D[直接使用]

    C --> E[获取Registry]
    E --> F[getAdapter<br/>providerId]
    F --> G[获取Provider元数据]
    F --> H[获取Model元数据]
    H -->|找到| I[使用静态模型]
    H -->|未找到| J[buildDefaultModel]
    I --> K[构建TextModelConfig]
    J --> K
    K --> L[保存到Storage]
    D --> L

转换逻辑 (converter.ts)

export async function convertLegacyToTextModelConfigWithRegistry(
  key: string,
  legacy: ModelConfig,
  registry: ITextAdapterRegistry
): Promise<TextModelConfig> {
  // 1. Provider映射
  const providerId = mapProviderToAdapterId(legacy.provider);

  // 2. 获取Adapter
  const adapter = registry.getAdapter(providerId);

  // 3. 获取Provider元数据
  const providerMeta = adapter.getProvider();

  // 4. 获取Model元数据
  let modelMeta = adapter.getModels().find(m => m.id === legacy.defaultModel);
  if (!modelMeta) {
    modelMeta = adapter.buildDefaultModel(legacy.defaultModel);
  }

  // 5. 构建TextModelConfig
  return {
    id: key,
    name: legacy.name,
    enabled: legacy.enabled,
    providerMeta,
    modelMeta,
    connectionConfig: {
      apiKey: legacy.apiKey,
      baseURL: legacy.baseURL
    },
    paramOverrides: legacy.llmParams || {}
  };
}

Provider映射规则:

• gemini → gemini (GeminiAdapter)
• anthropic → anthropic (AnthropicAdapter)
• openai | deepseek | zhipu | siliconflow | custom → openai (OpenAIAdapter)

自动转换时机

在ModelManager.init()初始化时:

async init(): Promise<void> {
  const existingModels = await this.getModelsFromStorage();

  for (const [key, existingModel] of Object.entries(existingModels)) {
    if (isLegacyConfig(existingModel)) {
      try {
        // 优先使用Registry转换
        const registry = await this.getRegistry();
        const convertedModel = await convertLegacyToTextModelConfigWithRegistry(
          key,
          existingModel,
          registry
        );
        updatedModels[key] = convertedModel;
        hasUpdates = true;
      } catch (error) {
        // Fallback到硬编码转换
        const convertedModel = convertLegacyToTextModelConfig(key, existingModel);
        updatedModels[key] = convertedModel;
      }
    }
  }

  // 保存转换后的配置
  if (hasUpdates) {
    await this.saveModelsToStorage(updatedModels);
  }
}

Service层集成

LLMService使用Registry

export class LLMService implements ILLMService {
  constructor(
    private modelManager: ModelManager,
    private registry: ITextAdapterRegistry
  ) {}

  async sendMessage(messages: Message[], provider: string): Promise<string> {
    // 1. 获取配置
    const config = await this.modelManager.getModel(provider) as TextModelConfig;

    // 2. 获取Adapter
    const adapter = this.registry.getAdapter(config.providerMeta.id);

    // 3. 调用Adapter
    const response = await adapter.sendMessage(messages, config);

    return response.content;
  }
}

关键特性:

• 通过config.providerMeta.id获取正确的Adapter
• 无需switch/case Provider类型
• SDK调用完全由Adapter封装
• 错误堆栈保留

工厂函数

export function createLLMService(modelManager: ModelManager): ILLMService {
  if (isRunningInElectron()) {
    return new ElectronLLMProxy();
  }

  // 创建Registry实例
  const registry = new TextAdapterRegistry();

  // 注入Registry到Service
  return new LLMService(modelManager, registry);
}

开发者指南

如何添加新Provider

1. 创建Adapter实现

// packages/core/src/services/llm/adapters/example-adapter.ts
import { AbstractTextProviderAdapter } from './abstract-adapter';
import type { TextProvider, TextModel, TextModelConfig, LLMResponse, Message, StreamHandlers } from '../types';

export class ExampleAdapter extends AbstractTextProviderAdapter {
  getProvider(): TextProvider {
    return {
      id: 'example',
      name: 'Example Provider',
      description: 'Example LLM Provider',
      defaultBaseURL: 'https://api.example.com/v1',
      connectionSchema: {
        required: ['apiKey'],
        optional: ['baseURL'],
        fieldTypes: {
          apiKey: 'string',
          baseURL: 'url'
        }
      }
    };
  }

  getModels(): TextModel[] {
    return [
      {
        id: 'example-model-v1',
        name: 'Example Model V1',
        description: 'Fast and efficient model',
        providerId: 'example',
        capabilities: {
          supportsStreaming: true,
          supportsTools: false,
          supportsReasoning: false,
          maxContextLength: 8000
        },
        parameterDefinitions: [
          {
            name: 'temperature',
            type: 'number',
            description: 'Sampling temperature',
            min: 0,
            max: 2,
            default: 0.7
          }
        ],
        defaultParameterValues: {
          temperature: 0.7
        }
      }
    ];
  }

  protected async doSendMessage(
    messages: Message[],
    config: TextModelConfig
  ): Promise<LLMResponse> {
    // 实现SDK调用逻辑
    const client = new ExampleSDK({
      apiKey: config.connectionConfig.apiKey,
      baseURL: config.connectionConfig.baseURL || this.getProvider().defaultBaseURL
    });

    try {
      const response = await client.chat.completions.create({
        model: config.modelMeta.id,
        messages: messages,
        ...config.paramOverrides
      });

      return {
        content: response.choices[0].message.content || '',
        reasoning: undefined,
        metadata: {
          model: config.modelMeta.id,
          usage: response.usage
        }
      };
    } catch (error: any) {
      // 保留原始错误堆栈
      throw error;
    }
  }

  protected async doSendMessageStream(
    messages: Message[],
    config: TextModelConfig,
    handlers: StreamHandlers
  ): Promise<void> {
    // 实现流式调用逻辑
    const client = new ExampleSDK({
      apiKey: config.connectionConfig.apiKey,
      baseURL: config.connectionConfig.baseURL
    });

    try {
      const stream = await client.chat.completions.create({
        model: config.modelMeta.id,
        messages: messages,
        stream: true,
        ...config.paramOverrides
      });

      for await (const chunk of stream) {
        const content = chunk.choices[0]?.delta?.content || '';
        if (content && handlers.onToken) {
          handlers.onToken(content);
        }
      }

      if (handlers.onComplete) {
        handlers.onComplete({ content: '', metadata: {} });
      }
    } catch (error: any) {
      if (handlers.onError) {
        handlers.onError(error);
      }
      throw error;
    }
  }
}

2. 注册到Registry

// packages/core/src/services/llm/adapters/registry.ts
import { ExampleAdapter } from './example-adapter';

export class TextAdapterRegistry implements ITextAdapterRegistry {
  private adapters: Map<string, ITextProviderAdapter>;

  constructor() {
    this.adapters = new Map();
    this.staticModelsCache = new Map();

    // 注册所有Adapter
    this.adapters.set('openai', new OpenAIAdapter());
    this.adapters.set('gemini', new GeminiAdapter());
    this.adapters.set('anthropic', new AnthropicAdapter());
    this.adapters.set('example', new ExampleAdapter());  // 新增
  }
}

3. 更新配置转换逻辑 (如需要)

// packages/core/src/services/model/converter.ts
function mapProviderToAdapterId(provider: string): string {
  switch (provider) {
    case 'gemini':
      return 'gemini';
    case 'anthropic':
      return 'anthropic';
    case 'example':  // 新增
      return 'example';
    case 'openai':
    case 'deepseek':
    case 'zhipu':
    case 'siliconflow':
    case 'custom':
    default:
      return 'openai';
  }
}

常见问题 (FAQ)

Q1: 为什么要重构为Adapter模式?

A:

1. 解耦SDK: 不同SDK的调用逻辑分离,易于维护和测试
2. 统一接口: 所有Provider遵循相同接口,简化Service层逻辑
3. 扩展性: 添加新Provider只需实现Adapter,无需修改Service
4. 可测试性: Adapter可独立Mock和测试

Q2: 传统配置如何迁移?

A: 自动迁移,无需手动操作:

1. ModelManager初始化时检测传统配置
2. 自动调用convertLegacyToTextModelConfigWithRegistry()
3. 转换后保存到Storage
4. 下次加载直接识别为新格式

Q3: TextModelConfig为什么要嵌入元数据?

A:

1. 自包含: 运行时无需查找Provider/Model元数据
2. 类型安全: 完整类型定义,IDE智能提示
3. 性能: 避免运行时查找,直接访问
4. 可追溯: 配置包含完整历史信息

Q4: 如何处理SDK错误?

A: Adapter必须保留原始错误堆栈:

try {
  const response = await sdk.call();
} catch (error: any) {
  // 直接throw,不要包装,保留原始堆栈
  throw error;
}

Q5: 如何支持动态模型获取?

A: 实现getModelsAsync()方法:

async getModelsAsync(config: TextModelConfig): Promise<TextModel[]> {
  const client = new OpenAI({
    apiKey: config.connectionConfig.apiKey,
    baseURL: config.connectionConfig.baseURL
  });

  const response = await client.models.list();

  return response.data.map(model => ({
    id: model.id,
    name: model.id,
    description: '',
    providerId: 'openai',
    capabilities: { /*...*/ },
    parameterDefinitions: [],
    defaultParameterValues: {}
  }));
}

Registry会自动fallback到静态模型。

Q6: Provider映射规则是什么?

A:

• Gemini: gemini → GeminiAdapter (Google SDK)
• Anthropic: anthropic → AnthropicAdapter (Anthropic SDK)
• OpenAI及兼容:
  • openai → OpenAIAdapter
  • deepseek → OpenAIAdapter (OpenAI兼容)
  • zhipu → OpenAIAdapter (OpenAI兼容)
  • siliconflow → OpenAIAdapter (OpenAI兼容)
  • custom → OpenAIAdapter (OpenAI兼容)

测试策略

单元测试

• Adapter测试: Mock SDK,测试数据转换和错误处理
• Registry测试: 测试Adapter注册、查找、缓存逻辑
• 转换测试: 测试传统配置 → 新配置映射正确性

集成测试

• Adapter集成: 使用真实API密钥测试SDK调用
• 迁移集成: 测试配置自动转换和持久化
• 回归测试: 验证API签名和行为不变

相关链接

• 需求文档
• 设计文档
• 任务文档
• LLM Service源码
• Adapter目录

Source: docs/architecture/llm-refactor.md