refactor(generation-controller): 优化SSE流式生成逻辑，提升稳定性与超时处理

1.  放宽SSE超时时间至5分钟，避免长生成任务被中断
2.  抽离LLM生成逻辑到独立线程池，避免阻塞SSE主线程
3.  添加心跳机制防止连接因空闲超时断开
4.  重构代码结构，提取生成工具方法，优化变量命名
5.  新增任务取消与异常捕获逻辑，完善错误处理

backend/generation-service/src/main/java/com/artedu/generation/controller/GenerationController.java

@@ -28,8 +28,12 @@ import java.util.ArrayList;
 import java.util.HashMap;
 import java.util.List;
 import java.util.Map;
+import java.util.concurrent.CancellationException;
 import java.util.concurrent.ExecutorService;
 import java.util.concurrent.Executors;
+import java.util.concurrent.Future;
+import java.util.concurrent.TimeUnit;
+import java.util.concurrent.TimeoutException;
 import java.util.stream.Collectors;
 
 @Slf4j
@@ -56,6 +60,7 @@ public class GenerationController {
     private QaConfig qaConfig;
 
     private final ExecutorService sseExecutor = Executors.newCachedThreadPool();
+    private final ExecutorService llmExecutor = Executors.newCachedThreadPool();
 
     private static final List<Map<String, String>> STYLES = List.of(
             Map.of("id", "friendly", "name", "亲和风格", "desc", "温暖亲切，像朋友聊天"),
@@ -283,7 +288,8 @@ public class GenerationController {
             @RequestParam(value = "roomid", required = false) String roomid,
             @RequestParam(value = "staffName", required = false) String staffName) {
         log.info("流式生成多风格话术参数:{},{},{},{}", fromUser, toUser, roomid, staffName);
-        SseEmitter emitter = new SseEmitter(120000L);
+        // 生成过程可能超过 30s，SSE 总超时放宽到 5 分钟
+        SseEmitter emitter = new SseEmitter(300000L);
 
         sseExecutor.execute(() -> {
             try {
@@ -292,9 +298,9 @@ public class GenerationController {
                 startEvent.put("styles", STYLES);
                 emitter.send(SseEmitter.event().name("start").data(startEvent));
 
-                List<ArchiveMessage> messages  = buildChatContext(fromUser, toUser, roomid);
+                List<ArchiveMessage> messages = buildChatContext(fromUser, toUser, roomid);
                 // 遍历消息：只取前三条是客户的消息，组装 查询知识库
-                String external_str = ""; //原始问题
+                StringBuilder externalStrBuilder = new StringBuilder(); //原始问题
                 // 历史消息上下文
                 List<ArchiveMessage> top15 = (messages == null || messages.isEmpty())
                         ? new ArrayList<>()
@@ -317,7 +323,7 @@ public class GenerationController {
                     if ("EXTERNAL".equals(messages.get(0).getFromRole())){ //第一条消息是客户的消息 则继续
                         if(count<3){
                             if ("EXTERNAL".equals(msg.getFromRole()) && is_external){ //判断判断前三条是否只包含客户消息
-                                external_str += msg.getContent()+",";
+                                externalStrBuilder.append(msg.getContent()).append(",");
                             }
                             if (!"EXTERNAL".equals(msg.getFromRole())){ //判断前三条是否包含非客户的消息，如果包含，则不需要后续的消息
                                 is_external = false;
@@ -327,33 +333,7 @@ public class GenerationController {
                     count++;
                 }
                 String chatContext = sb.toString();
-                String generated = "";
+                final String external_str = externalStrBuilder.toString();
-                //如果最后几条都是客户消息返回不为空，则直接返回结果，否则走原逻辑
-                if (!StringUtils.isBlank(external_str)){ //不为空 发起知识库的查询
-                    String answer = getAnswer(top15, external_str);
-                    log.info("原问题：{},命中知识库结果:{}", external_str, answer);
-                    if(!StringUtils.isBlank(answer)){
-                        ObjectMapper mapper = new ObjectMapper();
-                        List<String> list = mapper.readValue(answer, new TypeReference<List<String>>(){});
-                        //answer = answer.replaceAll("\"", "").replace("[","").replace("]","");
-                        for (String li : list){
-                            String promptNew = promptEngine.buildAnswerPrompt(li);
-                            String res = qianwenClient.generate(promptNew, 800);
-                            generated = generated + res + "|||";
-                        }
-                        if(StringUtils.isNotBlank(generated)){
-                            String[] parts = generated.split("\\|\\|\\|");
-                            if(parts.length<3){
-                                for (int i = 0; i < 3-parts.length; i++) {
-                                    String systemPrompt = promptEngine.buildSystemPrompt(chatContext, staffName);
-                                    String generatedNew = qianwenClient.generate(systemPrompt, "", 1200);
-                                    generated = generated + generatedNew;
-                                }
-                            }
-                        }
-
-                    }
-                }
 
                 if (chatContext == null || chatContext.isEmpty()) {
                     Map<String, Object> err = new HashMap<>();
@@ -363,15 +343,29 @@ public class GenerationController {
                     emitter.complete();
                     return;
                 }
-                if(StringUtils.isBlank(generated)){
-                    String systemPrompt = promptEngine.buildSystemPrompt(chatContext, staffName);
-                    String userPrompt = promptEngine.buildMultiStyleUserPrompt(staffName);
 
-                    long startTime = System.currentTimeMillis();
+                // LLM 调用在独立线程中同步执行，避免阻塞 SSE 线程
-                    generated = qianwenClient.generate(systemPrompt, userPrompt, 1200);
+                Future<String> generationFuture = llmExecutor.submit(() ->
-                    log.info("多风格话术生成完成, elapsed={}ms", System.currentTimeMillis() - startTime);
+                        generateUtteranceText(top15, chatContext, external_str, staffName));
+                emitter.onCompletion(() -> generationFuture.cancel(true));
+                emitter.onTimeout(() -> generationFuture.cancel(true));
+
+                // 等待生成结果，每 10s 发送一次心跳，防止 nginx/移动网络因 30s 空闲超时而断开
+                while (!generationFuture.isDone()) {
+                    try {
+                        generationFuture.get(10, TimeUnit.SECONDS);
+                        break;
+                    } catch (TimeoutException e) {
+                        try {
+                            emitter.send(SseEmitter.event().comment("keep-alive"));
+                        } catch (IOException ioe) {
+                            log.warn("SSE 心跳发送失败，连接已断开");
+                            generationFuture.cancel(true);
+                            return;
+                        }
+                    }
                 }
-
+                String generated = generationFuture.get();
 
                 List<String> utterances = parseMultiStyle(generated);
                 if (utterances.size() < 3) {
@@ -416,6 +410,8 @@ public class GenerationController {
                 emitter.send(SseEmitter.event().name("done").data(done));
                 emitter.complete();
 
+            } catch (CancellationException e) {
+                log.warn("SSE 流式生成被取消（客户端断开或超时）");
             } catch (Exception e) {
                 log.error("流式生成失败", e);
                 try {
@@ -431,6 +427,50 @@ public class GenerationController {
         return emitter;
     }
 
+    /**
+     * 在后台线程执行 LLM 生成，返回最终话术文本
+     */
+    private String generateUtteranceText(List<ArchiveMessage> top15, String chatContext,
+                                         String external_str, String staffName) throws Exception {
+        String generated = "";
+        //如果最后几条都是客户消息返回不为空，则直接返回结果，否则走原逻辑
+        if (!StringUtils.isBlank(external_str)) { //不为空 发起知识库的查询
+            String answer = getAnswer(top15, external_str);
+            log.info("原问题：{},命中知识库结果:{}", external_str, answer);
+            if (!StringUtils.isBlank(answer)) {
+                ObjectMapper mapper = new ObjectMapper();
+                List<String> list = mapper.readValue(answer, new TypeReference<List<String>>(){});
+                //answer = answer.replaceAll("\"", "").replace("[","").replace("]","");
+                for (String li : list) {
+                    String promptNew = promptEngine.buildAnswerPrompt(li);
+                    String res = qianwenClient.generate(promptNew, 800);
+                    generated = generated + res + "|||";
+                }
+                if (StringUtils.isNotBlank(generated)) {
+                    String[] parts = generated.split("\\|\\|\\|");
+                    if (parts.length < 3) {
+                        for (int i = 0; i < 3 - parts.length; i++) {
+                            String systemPrompt = promptEngine.buildSystemPrompt(chatContext, staffName);
+                            String generatedNew = qianwenClient.generate(systemPrompt, "", 1200);
+                            generated = generated + generatedNew;
+                        }
+                    }
+                }
+
+            }
+        }
+
+        if (StringUtils.isBlank(generated)) {
+            String systemPrompt = promptEngine.buildSystemPrompt(chatContext, staffName);
+            String userPrompt = promptEngine.buildMultiStyleUserPrompt(staffName);
+
+            long startTime = System.currentTimeMillis();
+            generated = qianwenClient.generate(systemPrompt, userPrompt, 1200);
+            log.info("多风格话术生成完成, elapsed={}ms", System.currentTimeMillis() - startTime);
+        }
+        return generated;
+    }
+
     /**
      * 解析多风格话术，按 "|||" 分隔
      */