前言最近帮公司把AI能力从单体应用迁移到微服务架构踩了无数个坑。这篇文章不是Hello World而是真实生产环境中的血泪总结。如果你也在考虑怎么把ChatGPT接入Spring Cloud建议先看完这篇再动手。一、为什么你的AI项目总是玩具级去年用Spring Boot写了个AI问答Demo本地跑得很爽一上线就崩。问题在哪单体架构的AI应用三大痛点模型调用阻塞主线程- 一个慢请求拖垮整个服务API Key裸奔在代码里- 安全审计直接挂红灯流式响应没法做负载均衡- Nginx都懵圈了直到把架构改成这样才真正敢上生产环境┌─────────────┐ ┌─────────────┐ ┌─────────────────┐ │ Gateway │────▶│ AI-Service │────▶│ LLM Provider │ │ (流式支持) │ │ (熔断/限流) │ │ (OpenAI/Claude) │ └─────────────┘ └─────────────┘ └─────────────────┘ │ ▼ ┌─────────────┐ │ Vector-DB │ │ (RAG检索) │ └─────────────┘核心思路把AI能力拆成独立服务别和业务代码耦合二、环境准备别在版本上浪费时间先给个能直接跑的pom.xml配置这些版本我测过兼容性parentgroupIdorg.springframework.boot/groupIdartifactIdspring-boot-starter-parent/artifactIdversion3.2.5/version/parentpropertiesspring-cloud.version2023.0.1/spring-cloud.versionspring-ai.version0.8.1/spring-ai.version/propertiesdependencies!-- Spring Boot基础 --dependencygroupIdorg.springframework.boot/groupIdartifactIdspring-boot-starter-web/artifactId/dependency!-- Spring Cloud Alibaba Nacos --dependencygroupIdcom.alibaba.cloud/groupIdartifactIdspring-cloud-starter-alibaba-nacos-discovery/artifactId/dependency!-- Spring AI核心 --dependencygroupIdorg.springframework.ai/groupIdartifactIdspring-ai-openai-spring-boot-starter/artifactId/dependency!-- 流式响应必备 --dependencygroupIdorg.springframework.boot/groupIdartifactIdspring-boot-starter-webflux/artifactId/dependency/dependencies避坑点Spring Boot 3.x必须配合JDK 17别问问就是踩过坑Spring AI 0.8才支持函数调用Function Calling老版本别用Nacos 2.2才支持gRPC配置中心用HTTP就行三、核心架构AI-Service设计3.1 配置隔离别让Key泄露生产环境千万别写死在application.yml里用Nacos配置中心加密ConfigurationConfigurationProperties(prefixai.openai)DatapublicclassOpenAiProperties{privateStringapiKey;privateStringbaseUrlhttps://api.openai.com;privateStringmodelgpt-4-turbo-preview;privateDurationtimeoutDuration.ofSeconds(30);// 连接池配置高并发必备privateintmaxConnections100;privateintmaxIdleTime20;}Nacos配置示例记得开启加密插件ai:openai:api-key:${ENC{your-encrypted-key-here}}base-url:https://api.openai.commodel:gpt-4-turbo-preview3.2 服务层流式响应熔断器这是核心代码直接复制改改就能用ServiceSlf4jpublicclassAiChatService{AutowiredprivateOpenAiChatClientchatClient;AutowiredprivateCircuitBreakerRegistrycircuitBreakerRegistry;/** * 同步调用 - 适合简单问答 */CircuitBreaker(nameaiChat,fallbackMethodfallbackChat)publicStringsimpleChat(Stringmessage){log.info(收到请求: {},message);returnchatClient.call(newPrompt(message,OpenAiChatOptions.builder().withTemperature(0.7).withMaxTokens(2000).build())).getResult().getOutput().getContent();}/** * 流式调用 - 生产环境主推用户体验好 */publicFluxStringstreamChat(Stringmessage,StringsessionId){CircuitBreakercbcircuitBreakerRegistry.circuitBreaker(aiStream);returnchatClient.stream(newPrompt(message,OpenAiChatOptions.builder().withStreamUsage(true)// 开启流式用量统计.build())).map(response-response.getResult().getOutput().getContent()).filter(Objects::nonNull).transformDeferred(RxReactiveStreams.toFlux()).transform(cb::executeFlux)// 熔断器包装.doOnError(e-log.error(流式调用失败, session{},sessionId,e)).onErrorResume(e-Flux.just(服务繁忙请稍后重试));}/** * 降级方法 */privateStringfallbackChat(Stringmessage,Exceptionex){log.warn(触发熔断, 异常: {},ex.getMessage());return当前AI服务繁忙请1分钟后再试;}}关键设计说明用CircuitBreaker防止LLM API挂掉时拖垮服务streamChat返回FluxString前端用SSE接收每个请求带sessionId方便链路追踪3.3 控制器SSE流式输出前端要的是打字机效果别用WebSocketSSE更简单RestControllerRequestMapping(/api/ai)RequiredArgsConstructorpublicclassAiChatController{privatefinalAiChatServiceaiChatService;/** * 流式聊天接口 - 前端EventSource直接连 */GetMapping(value/stream,producesMediaType.TEXT_EVENT_STREAM_VALUE)publicFluxServerSentEventStringstreamChat(RequestParamStringmessage,RequestHeader(X-Session-Id)StringsessionId){// 限流检查if(!rateLimiter.tryAcquire()){returnFlux.just(ServerSentEvent.builder(请求过于频繁).build());}returnaiChatService.streamChat(message,sessionId).map(content-ServerSentEvent.Stringbuilder().id(UUID.randomUUID().toString()).event(message).data(content).build()).concatWith(Flux.just(ServerSentEvent.Stringbuilder().event(complete).data([DONE]).build()));}}前端接收示例Vue3consteventSourcenewEventSource(/api/ai/stream?message${encodeURIComponent(msg)});eventSource.onmessage(e){if(e.data[DONE]){eventSource.close();return;}// 逐字显示效果responseText.valuee.data;};四、微服务治理网关熔断限流4.1 Gateway路由配置AI服务单独部署通过Gateway暴露这里配置流式响应支持spring:cloud:gateway:routes:-id:ai-serviceuri:lb://ai-servicepredicates:-Path/api/ai/**filters:-name:Retryargs:retries:3statuses:BAD_GATEWAY-name:RequestRateLimiterargs:redis-rate-limiter.replenishRate:10redis-rate-limiter.burstCapacity:20注意点lb://表示用Nacos服务发现流式响应别加ModifyResponseBody过滤器会缓冲整个流限流用Redis实现集群部署时共享状态4.2 熔断器配置Resilience4j比Hystrix轻量适合AI场景resilience4j:circuitbreaker:configs:default:slidingWindowSize:10minimumNumberOfCalls:5permittedNumberOfCallsInHalfOpenState:3automaticTransitionFromOpenToHalfOpenEnabled:truewaitDurationInOpenState:30sfailureRateThreshold:50eventConsumerBufferSize:10instances:aiChat:baseConfig:defaultwaitDurationInOpenState:60s# AI服务恢复慢多给点时间实战经验LLM API偶尔抽风429错误熔断阈值别设太低半开状态放3个请求测试比默认值1个更稳妥开启自动过渡别手动干预五、RAG增强别让模型瞎编生产环境AI必须接知识库不然回答不靠谱。用Redis Stack做向量库比Pinecone省钱5.1 向量存储配置ConfigurationpublicclassVectorStoreConfig{BeanpublicVectorStorevectorStore(RedisTemplateString,StringredisTemplate,EmbeddingClientembeddingClient){// 使用RedisJSON RediSearchreturnRedisVectorStore.builder(redisTemplate,embeddingClient).withIndexName(kb-index).withPrefix(kb:).withMetadataFields(MetadataField.text(category),MetadataField.numeric(timestamp)).initializeSchema(true).build();}}5.2 RAG检索服务ServicepublicclassRagService{AutowiredprivateVectorStorevectorStore;AutowiredprivateChatClientchatClient;publicStringchatWithKnowledge(Stringquestion,Stringcategory){// 1. 向量化检索SearchRequestsearchRequestSearchRequest.query(question).withTopK(5).withSimilarityThreshold(0.7).withFilterExpression(category category);ListDocumentrelevantDocsvectorStore.similaritySearch(searchRequest);// 2. 构造PromptStringcontextrelevantDocs.stream().map(Document::getContent).collect(Collectors.joining(\n---\n));Stringprompt 基于以下参考资料回答问题 %s 问题%s 要求 1. 如果资料不足以回答明确告知根据现有资料无法确定 2. 不要编造信息 3. 引用资料来源 .formatted(context,question);// 3. 调用模型returnchatClient.call(prompt);}}数据导入脚本Python辅助# 把公司文档导入向量库fromredisimportRedisfromredis.commands.search.fieldimportVectorField,TextFielddefindex_documents(docs):redis_clientRedis(hostlocalhost,port6379,decode_responsesTrue)fori,docinenumerate(docs):# 用Spring AI的Embedding API生成向量embeddingget_embedding(doc[content])redis_client.hset(fkb:{i},mapping{content:doc[content],category:doc[category],embedding:np.array(embedding).tobytes()})六、生产环境 checklist上线前逐项检查检查项工具/方法合格标准API Key安全Vault/Nacos加密代码中无明文Key流式响应超时Gateway配置5分钟不断连模型降级策略多模型路由OpenAI挂了转ClaudeToken消耗监控Micrometer Prometheus按用户统计用量敏感词过滤本地敏感词库政治/暴力内容拦截并发压力测试JMeter100并发响应3s多模型降级代码ComponentpublicclassModelRouter{privatefinalListChatClientclientsnewArrayList();privatefinalAtomicIntegercounternewAtomicInteger(0);publicChatClientgetAvailableClient(){// 简单轮询生产环境用更复杂的健康检查intindexcounter.getAndIncrement()%clients.size();returnclients.get(index);}}七、性能优化实录最后分享几个压测后的优化点1. HTTP连接池必须调BeanpublicClientHttpRequestFactoryrequestFactory(){HttpComponentsClientHttpRequestFactoryfactorynewHttpComponentsClientHttpRequestFactory();factory.setConnectTimeout(5000);factory.setReadTimeout(30000);// 关键连接复用否则高并发会炸factory.setHttpClient(HttpClientBuilder.create().setMaxConnTotal(200).setMaxConnPerRoute(50).build());returnfactory;}2. 流式响应别用Jackson默认Jackson会缓冲整个响应改用StreamingResponseBody或者WebFlux。3. 上下文压缩长对话历史别全传用摘要算法压缩publicStringcompressHistory(ListMessagehistory){if(history.size()10){// 早期消息用LLM生成摘要StringearlyContextsummarize(history.subList(0,5));returnearlyContextrecentMessages;}returnformat(history);}心得总结Spring Boot Spring Cloud AI 不是简单对接API而是完整的工程化问题。核心就三点服务隔离- AI能力独立部署别和业务耦合流式优先- 用户体验好资源占用反而更低防御编程- 熔断、降级、限流一个不能少代码已经脱敏可以直接拿去改改用。有问题评论区见看到都会回也可以私信关注。转载请注明出处商业使用请联系授权。