提示词完善

src/main/java/com/rj/common/AudioAnalysisSceneType.java

@@ -15,8 +15,8 @@ public enum AudioAnalysisSceneType {
             "qwen-plus"
     ),
     SCENARIO_SOFT_SALE(
-            "prompts/audio_text_analysis_furniture_system.txt",
+            "prompts/audio_text_analysis_soft_sale_system_prompts.txt",
-            "prompts/audio_text_analysis_furniture_user.txt",
+            "prompts/audio_text_analysis_soft_sale_user_prompts.txt",
             "qwen-plus"
     ),
 
@@ -24,9 +24,9 @@ public enum AudioAnalysisSceneType {
      * 会议纪要/会议分析场景
      * 提示词文件和模型可根据实际需要进行调整。
      */
-    SCENARIO_COMMON_SALE(
+    SCENARIO_SUMMARY(
-            "prompts/audio_text_analysis_meeting_system.txt",
+            "prompts/audio_text_analysis_summary_system_prompts.txt",
-            "prompts/audio_text_analysis_meeting_user.txt",
+            "prompts/audio_text_analysis_summary_user_prompts.txt",
             "qwen-plus"
     ),
 

src/main/java/com/rj/controller/AudioManagementController.java

@@ -49,7 +49,7 @@ public class AudioManagementController {
     private static final String SCENARIO_FURNITURE_SALE = "FURNITURE";
     private static final String SCENARIO_MEETING_SUMMARY = "MEETING_SUMMARY";
     private static final String SCENARIO_CAR_SALE = "CAR_SALE";
-    private static final String SCENARIO_COMMON_SALE = "COMMON_SALE";
+    private static final String SCENARIO_SUMMARY = "SUMMARY";
     private static final String SCENARIO_SPEAKING_TRAINING= "SPEAKING_TRAINING";  //租赁模式
 
 
@@ -510,6 +510,7 @@ public class AudioManagementController {
                         customerToUpdate.setDetailedAddress(audioManagement.getRemarks());
                         customerToUpdate.setSalesPhone(audioManagement.getSalesPhone());
                         customerToUpdate.setSalesName(audioManagement.getSalesName());
+                        customerToUpdate.setRecordingCount(customerToUpdate.getRecordingCount()+1);
                         customerToUpdate.setUpdateTime(LocalDateTime.now());
                         customerManagementService.updateById(customerToUpdate);
                         log.info("同步更新客户信息成功，客户ID: {}", customerToUpdate.getId());
@@ -728,6 +729,15 @@ public class AudioManagementController {
                     audioManagement.getCustomerName(),
                     audioManagement.getCustomerPhone()
             );
+//            AudioTextAnalysisFurniture furniture = audioTextAnalysisLlmService.generateSummaryAndSave_V2(
+//                    sceneType,
+//                    recordingText,
+//                    audioManagement.getId(),
+//                    audioManagement.getSalesName(),
+//                    audioManagement.getSalesPhone(),
+//                    audioManagement.getCustomerName(),
+//                    audioManagement.getCustomerPhone()
+//            );
             log.info("llm return  furniture Analysis   : ",furniture.toString() );
             if (recordingText == null || recordingText.trim().isEmpty()) {
                 result.put("success", false);
@@ -762,8 +772,8 @@ public class AudioManagementController {
         if (SCENARIO_FURNITURE_SALE.equals(s)) {
             return AudioAnalysisSceneType.SCENARIO_FURNITURE_SALE;
         }
-        if (SCENARIO_COMMON_SALE.equals(s)) {
+        if (SCENARIO_SUMMARY.equals(s)) {
-            return AudioAnalysisSceneType.SCENARIO_COMMON_SALE;
+            return AudioAnalysisSceneType.SCENARIO_SUMMARY;
         }
         if (SCENARIO_CAR_SALE.equals(s)) {
             return AudioAnalysisSceneType.SCENARIO_CAR_SALE;

src/main/java/com/rj/controller/SalesManagementController.java

@@ -197,7 +197,7 @@ public class SalesManagementController {
         Map<String, Object> result = new HashMap<>();
         try {
             LambdaQueryWrapper<SalesManagement> queryWrapper = new LambdaQueryWrapper<>();
-            queryWrapper.eq(SalesManagement::getLoginAccount, loginAccount);
+            queryWrapper.like(SalesManagement::getLoginAccount, loginAccount);
             List<SalesManagement> salesList = salesManagementService.list(queryWrapper);
             result.put("success", true);
             result.put("message", "查询成功");

src/main/java/com/rj/service/IAudioTextAnalysisLlmService.java

@@ -21,6 +21,17 @@ public interface IAudioTextAnalysisLlmService {
      */
     String generateSummaryByLLM(AudioAnalysisSceneType sceneType, String recordingText);
 
+    /**
+     * 与 {@link #generateSummaryByLLM} 相同的提示词与场景逻辑，但通过本地 OpenAI 兼容接口调用大模型
+     * （配置项与客户问答本地模型一致：{@code ai.qa.local.chat-url}、{@code ai.qa.local.default-model}、{@code ai.qa.local.max-tokens}；
+     * 另支持 {@code ai.qa.local.context-length}、{@code ai.qa.local.token-margin}、{@code ai.qa.local.min-output-tokens}、{@code ai.qa.local.input-tokens-per-char} 以避免超出本地模型上下文。）
+     *
+     * @param sceneType     业务场景类型
+     * @param recordingText 录音转写文本
+     * @return 大模型返回的原始内容
+     */
+    String generateSummaryByLocalLLM(AudioAnalysisSceneType sceneType, String recordingText);
+
     /**
      * 处理完整的业务逻辑：调用大模型生成总结，解析JSON，保存数据
      * 此方法用于家具场景的完整业务处理

src/main/java/com/rj/service/impl/AudioTextAnalysisLlmServiceImpl.java

@@ -21,10 +21,16 @@ import com.rj.service.IAudioTextAnalysisFurnitureService;
 import com.rj.service.IAudioTextAnalysisLlmService;
 import com.rj.service.IAudioTextAnalysisSopService;
 import com.rj.service.ITodoItemService;
+import dev.langchain4j.model.chat.response.ChatResponse;
+import dev.langchain4j.model.chat.response.StreamingChatResponseHandler;
+import dev.langchain4j.model.openai.OpenAiStreamingChatModel;
 import lombok.extern.slf4j.Slf4j;
 import org.springframework.beans.factory.annotation.Autowired;
+import org.springframework.beans.factory.annotation.Value;
 import org.springframework.stereotype.Service;
 
+import jakarta.annotation.PostConstruct;
+
 import java.io.InputStream;
 import java.nio.charset.StandardCharsets;
 import java.time.LocalDateTime;
@@ -33,6 +39,9 @@ import java.util.Map;
 import java.util.Scanner;
 import java.util.UUID;
 import java.util.concurrent.ConcurrentHashMap;
+import java.util.concurrent.CountDownLatch;
+import java.util.concurrent.TimeUnit;
+import java.util.concurrent.atomic.AtomicReference;
 
 /**
  * 音频文本分析 - 大模型通用服务实现
@@ -59,6 +68,54 @@ public class AudioTextAnalysisLlmServiceImpl implements IAudioTextAnalysisLlmSer
     @Autowired
     private IAudioTextAnalysisSopService sopService;
 
+    @Value("${ai.qa.local.chat-url:http://192.168.1.44:8000/v1/chat/completions}")
+    private String chatCompletionsUrl;
+
+    @Value("${ai.qa.local.default-model:Qwen2.5-7B-Instruct}")
+    private String localDefaultModel;
+
+    /** 单次补全上限（OpenAI max_tokens）；与「模型总上下文 context-length」不是同一概念 */
+    @Value("${ai.qa.local.max-tokens:512}")
+    private int localMaxTokens;
+
+    /** 本地模型上下文长度（与推理侧 max context 一致，用于避免 input+max_tokens 超过上限；默认 16384） */
+    @Value("${ai.qa.local.context-length:16384}")
+    private int localContextLength;
+
+    /** 预留余量，避免服务端计数与本地估算不一致 */
+    @Value("${ai.qa.local.token-margin:80}")
+    private int localTokenMargin;
+
+    /**
+     * 截断转写时希望保留的「最小输出预算」启发值（须显著小于 context-length，且不得大于 max-tokens，
+     * 因为实际 effMax 恒为 min(max-tokens, 剩余上下文)）。
+     */
+    @Value("${ai.qa.local.min-output-tokens:128}")
+    private int configuredLocalMinOutputTokens;
+
+    /** {@link #configuredLocalMinOutputTokens} 经与 max-tokens 对齐后的实际取值 */
+    private int effectiveLocalMinOutputTokens;
+
+    /**
+     * 对完整 prompt 字符串做输入 token 的保守上界估算（偏大会多截断、偏少仍可能 400）。
+     * 中文为主可保持默认；英文偏多时可酌情调低（如 0.35）。
+     */
+    @Value("${ai.qa.local.input-tokens-per-char:1.25}")
+    private double localInputTokensPerChar;
+
+    @PostConstruct
+    void clampLocalMinOutputTokens() {
+        effectiveLocalMinOutputTokens = Math.max(1, Math.min(configuredLocalMinOutputTokens, localMaxTokens));
+        if (effectiveLocalMinOutputTokens < configuredLocalMinOutputTokens) {
+            log.warn(
+                    "ai.qa.local.min-output-tokens={} 大于 ai.qa.local.max-tokens={}：有效输出 effMax 恒 ≤ max-tokens，已钳制为 {}。"
+                            + " 切勿将 min-output-tokens 配成与 context-length 相同。",
+                    configuredLocalMinOutputTokens,
+                    localMaxTokens,
+                    effectiveLocalMinOutputTokens);
+        }
+    }
+
     @Override
     public String generateSummaryByLLM(AudioAnalysisSceneType sceneType, String recordingText) {
         long startTime = System.currentTimeMillis();
@@ -112,6 +169,195 @@ public class AudioTextAnalysisLlmServiceImpl implements IAudioTextAnalysisLlmSer
         }
     }
 
+    @Override
+    public String generateSummaryByLocalLLM(AudioAnalysisSceneType sceneType, String recordingText) {
+        long startTime = System.currentTimeMillis();
+        log.info("开始调用本地大模型生成总结，场景: {}", sceneType);
+
+        String systemPrompt = getSystemPrompt(sceneType);
+        String userPromptTemplate = getUserPromptTemplate(sceneType);
+        String recordingForPrompt = fitRecordingTextToLocalContext(systemPrompt, userPromptTemplate, recordingText);
+        String userPrompt = buildPrompt(userPromptTemplate, recordingForPrompt);
+
+        try {
+            int effectiveMaxTokens = computeEffectiveLocalMaxTokens(systemPrompt, userPrompt);
+            String rawContent = callLocalOpenAiChatCompletionsByLangChain4j(
+                    systemPrompt,
+                    userPrompt,
+                    localDefaultModel,
+                    effectiveMaxTokens);
+            log.info("本地大模型生成总结完成，场景: {}, 结果长度: {}",
+                    sceneType, rawContent != null ? rawContent.length() : 0);
+            return rawContent;
+        } catch (Exception e) {
+            log.error("调用本地大模型生成总结失败, 场景: {}", sceneType, e);
+            throw new RuntimeException("调用本地大模型失败: " + e.getMessage(), e);
+        } finally {
+            long endTime = System.currentTimeMillis();
+            long durationMillis = endTime - startTime;
+            double durationSeconds = durationMillis / 1000.0;
+            double durationMinutes = durationSeconds / 60.0;
+            log.info("调用本地大模型耗时: {} 毫秒, {} 秒, {} 分钟, 场景: {}",
+                    durationMillis,
+                    String.format("%.2f", durationSeconds),
+                    String.format("%.2f", durationMinutes),
+                    sceneType);
+        }
+    }
+
+    /**
+     * 与 {@link AiQaCustomerAskServiceImpl} 中非流式问答路径一致：LangChain4j + OpenAI 兼容接口，流式聚合为完整文本。
+     */
+    private String callLocalOpenAiChatCompletionsByLangChain4j(
+            String systemPrompt, String userContent, String model, int maxTokens) {
+        log.info("音频分析本地 LLM：model={}, chatCompletionsUrl={}", model, chatCompletionsUrl);
+        String openAiBaseUrl = normalizeOpenAiBaseUrl(chatCompletionsUrl);
+        OpenAiStreamingChatModel chatModel = OpenAiStreamingChatModel.builder()
+                .baseUrl(openAiBaseUrl)
+                .apiKey("not-used")
+                .modelName(model)
+                .maxTokens(maxTokens)
+                .build();
+
+        String prompt = buildFullLocalLlmPrompt(systemPrompt, userContent);
+        StringBuilder answerBuffer = new StringBuilder();
+        CountDownLatch done = new CountDownLatch(1);
+        AtomicReference<Throwable> errorRef = new AtomicReference<>();
+
+        chatModel.chat(prompt, new StreamingChatResponseHandler() {
+            @Override
+            public void onPartialResponse(String partialResponse) {
+                if (partialResponse != null && !partialResponse.isEmpty()) {
+                    answerBuffer.append(partialResponse);
+                }
+            }
+
+            @Override
+            public void onCompleteResponse(ChatResponse completeResponse) {
+                done.countDown();
+            }
+
+            @Override
+            public void onError(Throwable error) {
+                errorRef.set(error);
+                done.countDown();
+            }
+        });
+
+        try {
+            boolean finished = done.await(120, TimeUnit.SECONDS);
+            if (!finished) {
+                throw new IllegalStateException("本地大模型流式调用超时（120s）");
+            }
+        } catch (InterruptedException e) {
+            Thread.currentThread().interrupt();
+            throw new IllegalStateException("本地大模型流式调用被中断", e);
+        }
+
+        if (errorRef.get() != null) {
+            throw new IllegalStateException("本地大模型流式调用失败: " + errorRef.get().getMessage(), errorRef.get());
+        }
+
+        String answer = answerBuffer.toString();
+        if (answer.isEmpty()) {
+            throw new IllegalStateException("本地大模型流式调用返回空响应");
+        }
+        return answer;
+    }
+
+    private String normalizeOpenAiBaseUrl(String rawUrl) {
+        if (rawUrl == null || rawUrl.trim().isEmpty()) {
+            throw new IllegalArgumentException("ai.qa.local.chat-url 不能为空");
+        }
+        String normalized = rawUrl.trim();
+        while (normalized.endsWith("/")) {
+            normalized = normalized.substring(0, normalized.length() - 1);
+        }
+        String suffix = "/chat/completions";
+        if (normalized.endsWith(suffix)) {
+            normalized = normalized.substring(0, normalized.length() - suffix.length());
+        }
+        return normalized;
+    }
+
+    private String buildFullLocalLlmPrompt(String systemPrompt, String userContent) {
+        return "系统指令：" + systemPrompt + "\n\n用户输入：" + userContent;
+    }
+
+    private int estimateLocalPromptInputTokens(String fullPrompt) {
+        if (fullPrompt == null || fullPrompt.isEmpty()) {
+            return 0;
+        }
+        return (int) Math.ceil(fullPrompt.length() * localInputTokensPerChar);
+    }
+
+    private int computeEffectiveLocalMaxTokens(String systemPrompt, String userPrompt) {
+        String full = buildFullLocalLlmPrompt(systemPrompt, userPrompt);
+        int estIn = estimateLocalPromptInputTokens(full);
+        int capByContext = localContextLength - localTokenMargin - estIn;
+        int effective = Math.min(localMaxTokens, capByContext);
+        return Math.max(1, effective);
+    }
+
+    /**
+     * 在不超过本地模型上下文的前提下，尽量保留更多转写文本；必要时从尾部截断。
+     */
+    private String fitRecordingTextToLocalContext(
+            String systemPrompt, String userPromptTemplate, String recordingText) {
+        String recording = recordingText != null ? recordingText : "";
+        if (recording.isEmpty()) {
+            return recording;
+        }
+        String userFull = buildPrompt(userPromptTemplate, recording);
+        String full = buildFullLocalLlmPrompt(systemPrompt, userFull);
+        int est = estimateLocalPromptInputTokens(full);
+        int effMax = Math.min(localMaxTokens, localContextLength - localTokenMargin - est);
+        if (effMax >= effectiveLocalMinOutputTokens) {
+            return recording;
+        }
+        int lo = 0;
+        int hi = recording.length();
+        int best = 0;
+        while (lo <= hi) {
+            int mid = (lo + hi) >>> 1;
+            String sub = recording.substring(0, mid);
+            userFull = buildPrompt(userPromptTemplate, sub);
+            full = buildFullLocalLlmPrompt(systemPrompt, userFull);
+            est = estimateLocalPromptInputTokens(full);
+            effMax = Math.min(localMaxTokens, localContextLength - localTokenMargin - est);
+            if (effMax >= effectiveLocalMinOutputTokens) {
+                best = mid;
+                lo = mid + 1;
+            } else {
+                hi = mid - 1;
+            }
+        }
+        if (best >= recording.length()) {
+            return recording;
+        }
+        if (best <= 0) {
+            userFull = buildPrompt(userPromptTemplate, "");
+            full = buildFullLocalLlmPrompt(systemPrompt, userFull);
+            int estEmpty = estimateLocalPromptInputTokens(full);
+            int effMaxEmpty = Math.min(localMaxTokens, localContextLength - localTokenMargin - estEmpty);
+            log.warn(
+                    "本地 LLM 上下文过紧：无转写时估算剩余输出预算 effMax≈{}（max-tokens={}），仍小于 effective min-output-tokens={}"
+                            + "（配置 min-output-tokens={}）。请减小提示词、提高 context-length，或降低 min-output-tokens / 提高 max-tokens。将仍尝试调用。",
+                    effMaxEmpty,
+                    localMaxTokens,
+                    effectiveLocalMinOutputTokens,
+                    configuredLocalMinOutputTokens);
+            return "";
+        }
+        log.warn(
+                "录音转写过长，已按本地模型上下文截断：原长度 {} 字符，保留 {} 字符（context-length={}，max-tokens={}）",
+                recording.length(),
+                best,
+                localContextLength,
+                localMaxTokens);
+        return recording.substring(0, best);
+    }
+
     private String buildPrompt(String template, String recordingText) {
         if (template == null) {
             return recordingText != null ? recordingText : "";
@@ -245,10 +491,18 @@ public class AudioTextAnalysisLlmServiceImpl implements IAudioTextAnalysisLlmSer
             String ownerPhone,
             String customerName,
             String customerPhone) {
-
+        // 1. 调用大模型生成总结
+        String rawContent = generateSummaryByLocalLLM(sceneType, recordingText);
+        if (rawContent == null || rawContent.trim().isEmpty()) {
+            log.warn("大模型返回内容为空");
+            return null;
+        }
         AudioTextAnalysisFurniture furniture = new AudioTextAnalysisFurniture();
-        furniture.setRecordingText(recordingText);
+        // 4. 更新AudioManagement的summary字段
-
+        AudioManagement audioManagement = new AudioManagement();
+        audioManagement.setId(parentId);
+        audioManagement.setSummary(rawContent);
+        audioManagementService.updateById(audioManagement);
 
         return furniture;
     }

src/main/resources/application.yml

@@ -16,7 +16,11 @@ ai:
     local:
       chat-url: http://192.168.1.44:8000/v1/chat/completions
       default-model: Qwen2.5-7B-Instruct
-      max-tokens: 512
+      # 模型总上下文（与推理服务一致）；勿与 max-tokens、min-output-tokens 混用同一数值
+      context-length: 16384
+      # 单次补全上限。过小（如 512）会导致课堂纪要等 JSON 总结被截断、字段残缺或不准；建议 2048～4096（须保证 input+max_tokens 不超过 context-length）
+      max-tokens: 2048
+      min-output-tokens: 128
 
 langchain4j:
   open-ai:

src/main/resources/prompts/audio_text_analysis_summary_user_prompts.txt

@@ -1,4 +1,4 @@
-请阅读以下录音文本，从中提取客户信息并生成 JSON。必须覆盖所有字段，缺失信息请基于语境合理推断或标注"暂无信息"。
+请阅读以下录音文本，从中提取重要关键信息并生成 JSON。必须覆盖所有字段，缺失信息请基于语境合理推断或标注"暂无信息"。
 
 字段要求：
 1. style_type：主题类型，仅可填「会议纪要」「课堂纪要」「面试纪要」。