从传统到 AI 原生：Java 微服务集成大语言模型的架构演进与性能突围

@RestController
@RequestMapping("/api/ai")
public class AIController {
    private final ChatClient chatClient;

    public AIController(ChatClient.Builder builder) {
        this.chatClient = builder
                .defaultSystem("你是一个专业的电商助手，请用中文回复")
                .build();
    }

    @GetMapping("/summarize")
    public Flux<String> summarize(@RequestParam String orderId) {
        // 返回 Flux<String> 实现 SSE 流式输出
        return chatClient.prompt()
                .user(u -> u.text("请总结订单 {id} 的购买记录和潜在风险").param("id", orderId))
                .stream()
                .content();
    }
}

# application.yml
spring:
  threads:
    virtual:
      enabled: true

location /api/ai/ {
    proxy_pass http://ai-service;
    proxy_buffering off;          # 关键：禁用缓冲
    proxy_cache off;
    proxy_read_timeout 600s;      # AI 响应慢，超时设大
    chunked_transfer_encoding on;
}

@Configuration
public class AIClientConfig {

    @Bean
    public WebClient aiWebClient() {
        HttpClient httpClient = HttpClient.create()
                .option(ChannelOption.CONNECT_TIMEOUT_MILLIS, 3000) // 3s 连不上就弃
                .responseTimeout(Duration.ofSeconds(60))           // 整体响应 60s 上限
                .doOnConnected(conn -> 
                    conn.addHandlerLast(new ReadTimeoutHandler(60, TimeUnit.SECONDS))
                );

        return WebClient.builder()
                .clientConnector(new ReactorClientHttpConnector(httpClient))
                .codecs(configurer -> configurer.defaultCodecs().maxInMemorySize(1024 * 1024 * 10)) // 10MB
                .build();
    }
}

Flux<String> aiStream = chatClient.prompt().user("...").stream().content();
// 控制每秒最多向下游推送 50 个 Token，防止客户端缓冲区溢出
return aiStream.limitRate(50);

@Configuration
public class EmbeddingCacheConfig {
    
    @Bean
    public Cache<String, float[]> embeddingCache() {
        return Caffeine.newBuilder()
                .maximumSize(10000)
                .expireAfterWrite(Duration.ofHours(1))
                .recordStats()
                .build();
    }

    @Bean
    public EmbeddingClient cachedEmbeddingClient(EmbeddingClient delegate, Cache<String, float[]> cache) {
        return new CachedEmbeddingClient(delegate, cache);
    }
}

@Async
public CompletableFuture<List<Document>> searchSimilarDocs(String query) {
    // 向量检索逻辑
}

// 在 Service 中组合
CompletableFuture<List<Document>> docsFuture = searchSimilarDocs(query);
CompletableFuture<String> answerFuture = chatClient.prompt()
    .user(u -> u.text("基于上下文回答: {context}").param("context", docsFuture.join()))
    .stream().content().collectList().map(list -> String.join("", list));

java -agentlib:native-image-agent=config-output-dir=src/main/resources/META-INF/native-image \
     -jar target/ai-service.jar

<plugin>
    <groupId>org.graalvm.buildtools</groupId>
    <artifactId>native-maven-plugin</artifactId>
    <configuration>
        <buildArgs>
            <buildArg>--no-fallback</buildArg>
            <buildArg>--initialize-at-build-time=org.slf4j</buildArg>
            <buildArg>-H:+ReportExceptionStackTraces</buildArg>
            <buildArg>-march=compatibility</buildArg> <!-- 兼容旧 CPU -->
        </buildArgs>
        <imageName>ai-service</imageName>
    </configuration>
</plugin>

@Component
public class TracingChatClientInterceptor implements ChatClientInterceptor {

    @Override
    public void before(ChatClientRequest request) {
        Span currentSpan = Tracing.currentSpan();
        if (currentSpan != null) {
            currentSpan.setAttribute("ai.prompt.length", request.getUserText().length());
            currentSpan.setAttribute("ai.model", "gpt-4");
        }
    }

    @Override
    public void after(ChatClientResponse response) {
        Span currentSpan = Tracing.currentSpan();
        if (currentSpan != null && response.getMetadata() != null) {
            currentSpan.setAttribute("ai.usage.total_tokens", 
                response.getMetadata().getUsage().getTotalTokens());
        }
    }
}

@Bean
public MeterBinder aiMetrics() {
    return (registry) -> {
        Counter.builder("ai.requests.total")
            .tag("model", "gpt-4")
            .register(registry);
        Gauge.builder("ai.response.time", this::getAvgP99Latency)
            .register(registry);
    };
}

@Bean
public Resilience4JProxyFactory resilience4JProxyFactory() {
    TimeLimiterConfig timeLimiterConfig = TimeLimiterConfig.custom()
        .timeoutDuration(Duration.ofSeconds(30)) // 整体超时 30s
        .build();

    RetryConfig retryConfig = RetryConfig.custom()
        .maxAttempts(3)
        .waitDuration(Duration.ofSeconds(2))
        .retryOnResult(response -> response.contains("RateLimitError"))
        .retryOnException(e -> e instanceof ConnectException)
        .build();

    return new Resilience4JProxyFactory(timeLimiterConfig, retryConfig);
}

@Fallback
public Flux<String> fallbackSummary(String orderId, Throwable ex) {
    logger.warn("AI service degraded for order {}, error: {}", orderId, ex.getMessage());
    return Flux.just("【温馨提示】当前智能总结繁忙，请稍后重试或联系人工客服。");
}