LLM大模型智能引擎实战–SpringAI+RAG+MCP+实时搜索（已完结）

[用户请求]
      |
      v
+-----------------------------------------------------------------+
|                   SpringAI 应用 (智能网关)                      |
|                                                                 |
|  +-------------------------+                                    |
|  |        Prompt           |                                    |
|  |                         |                                    |
|  | +---------------------+ |                                    |
|  | |      User Query     | |                                    |
|  | +---------------------+ |                                    |
|  | | Context from RAG    |<---+                                 |
|  | | Context from Search |<---+                                 |
|  | | Tool Calls (MCP)    |<---+                                 |
|  | +---------------------+ | |                                 |
|  +-------------------------+ |                                 |
|                              |                                 |
|  +-----------------------+   |                                 |
|  |    ChatClient         |   |                                 |
|  | (OpenAI, Anthropic...)|   |                                 |
|  +-----------------------+   |                                 |
+-----------------------------------------------------------------+
      |              |                |                 |
      |              |                |                 |
      v              v                v                 v
+----------+   +----------------+   +-------------+   +-----------------+
| LLM      |   | 向量数据库      |   | 实时搜索API   |   |  MCP Servers    |
| (GPT-4,  |   | (PgVector,     |   | (Google,    |   | (Calculator,    |
| Claude...)|   | RedisStack,   |   | Serper...)  |   | SQL, Shell...)  |
+----------+   | Chroma...)     |   +-------------+   +-----------------+
               +----------------+

<dependency>
    <groupId>org.springframework.ai</groupId>
    <artifactId>spring-ai-openai-spring-boot-starter</artifactId>
</dependency>
<dependency>
    <groupId>org.springframework.ai</groupId>
    <artifactId>spring-ai-pgvector-store-spring-boot-starter</artifactId>
</dependency>
<!-- 其他必要依赖，如Web -->

spring:
  ai:
    openai:
      api-key: ${OPENAI_API_KEY}
    vectorstore:
      pgvector:
        enabled: true
        # ... 数据库配置

@Service
@RequiredArgsConstructor
public class RagService {

    private final VectorStore vectorStore;
    private final EmbeddingClient embeddingClient;

    public List<Document> retrieveContext(String query) {
        // 将查询转换为嵌入向量
        List<Double> queryEmbedding = embeddingClient.embed(query);
        
        // 在向量库中搜索相似文档
        SearchRequest request = SearchRequest.query(query).withTopK(5);
        return vectorStore.similaritySearch(request);
    }

    // 初始化向量库的方法（文档入库）
    public void loadDocuments(List<Document> docs) { ... }
}

@Service
public class RealTimeSearchService {

    @Value("${serper.api.key}")
    private String serperApiKey;

    public String searchWeb(String query) {
        RestTemplate restTemplate = new RestTemplate();
        HttpHeaders headers = new HttpHeaders();
        headers.set("X-API-KEY", serperApiKey);
        headers.setContentType(MediaType.APPLICATION_JSON);

        Map<String, String> requestMap = Map.of("q", query, "num", "5");
        HttpEntity<Map<String, String>> request = new HttpEntity<>(requestMap, headers);

        ResponseEntity<Map> response = restTemplate.postForEntity(
            "https://api.serper.dev/search", request, Map.class);
        
        // 处理响应，提取搜索摘要...
        return extractSnippetsFromResponse(response.getBody());
    }
}

@Service
public class McpService {

    public String executeTool(String toolName, String input) {
        // 简化示例：实际应使用正式的MCP客户端库
        ProcessBuilder processBuilder = new ProcessBuilder("node", "/path/to/mcp/server", input);
        try {
            Process process = processBuilder.start();
            // 读取进程输出...
            String result = ...;
            return result;
        } catch (IOException e) {
            return "Error executing tool: " + e.getMessage();
        }
    }
}

@Service
@RequiredArgsConstructor
public class IntelligentEngineService {

    private final ChatClient chatClient;
    private final RagService ragService;
    private final RealTimeSearchService searchService;
    private final McpService mcpService;

    public String processQuery(String userQuery) {

        // 1. 并行获取RAG上下文和实时信息
        List<Document> ragContext = ragService.retrieveContext(userQuery);
        String realTimeInfo = searchService.searchWeb(userQuery);

        // 2. 构建系统提示，注入上下文、工具描述和指令
        String systemPrompt = """
                你是一个强大的AI助手。请根据以下上下文和实时信息回答问题。
                如果用户要求你执行计算、操作文件等，你可以调用可用工具。

                相关文档上下文：
                %s

                实时网络信息：
                %s

                可用工具：
                - get_current_time: 获取当前时间
                - read_file: 读取文件内容
                ...其他工具...

                请谨慎判断是否需要使用工具。
                """.formatted(ragContext, realTimeInfo);

        // 3. 构建用户消息
        Prompt prompt = new Prompt(
                List.of(
                        new SystemPromptTemplate(systemPrompt).createMessage(),
                        new UserMessage(userQuery)
                )
        );

        // 4. 调用LLM
        ChatResponse response = chatClient.call(prompt);

        // 5. (后续)处理LLM响应，解析其中的工具调用请求，并通过McpService执行
        // 这是一个复杂步骤，可能需要多轮对话

        return response.getResult().getOutput().getText();
    }
}