feat(ai): 优化 AI 陪练问答流程

Administrator

2025-08-18 a46c4d2db30c2f534400a6179cd82f7beb07a29d

 src/main/java/cc/mrbird/febs/ai/service/impl/AiServiceImpl.java

@@ -1,13 +1,20 @@
package cc.mrbird.febs.ai.service.impl;

import cc.mrbird.febs.ai.controller.enumerates.AiTypeEnum;
import cc.mrbird.febs.ai.entity.AiProductRole;
import cc.mrbird.febs.ai.req.ai.AiMessage;
import cc.mrbird.febs.ai.req.ai.AiRequest;
import cc.mrbird.febs.ai.res.ai.AiResponse;
import cc.mrbird.febs.ai.res.ai.RadarDataItem;
import cc.mrbird.febs.ai.res.ai.Report;
import cc.mrbird.febs.ai.service.AiProductRoleService;
import cc.mrbird.febs.ai.service.AiService;
import com.volcengine.ark.runtime.model.completion.chat.ChatCompletionChoice;
import com.volcengine.ark.runtime.model.completion.chat.ChatCompletionRequest;
import com.volcengine.ark.runtime.model.completion.chat.ChatMessage;
import com.volcengine.ark.runtime.model.completion.chat.ChatMessageRole;
import cn.hutool.core.collection.CollUtil;
import cn.hutool.json.JSONUtil;
import com.fasterxml.jackson.core.JsonProcessingException;
import com.fasterxml.jackson.databind.JsonNode;
import com.fasterxml.jackson.databind.ObjectMapper;
import com.volcengine.ark.runtime.model.completion.chat.*;
import com.volcengine.ark.runtime.service.ArkService;
import lombok.RequiredArgsConstructor;
import lombok.extern.slf4j.Slf4j;
@@ -22,6 +29,9 @@
import java.util.ArrayList;
import java.util.List;
import java.util.concurrent.TimeUnit;
import java.util.function.Consumer;
import java.util.regex.Matcher;
import java.util.regex.Pattern;
import java.util.stream.Collectors;

/**
@@ -33,10 +43,28 @@
public class AiServiceImpl implements AiService {

    private static final String CODE_SUCCESS = "200";
    private static final String CODE_GOING_ON = "199";
    private static final String CODE_NOT_FOUND = "201";
    private static final String CODE_ERROR = "500";

    private static final String SCHEMA_JSON = "{\n" +
            "     \"radar_data\": {\n" +
            "       \"problem_understanding\": \"object\",\n" +
            "       \"fluency\": \"object\",\n" +
            "       \"principle_adherence\": \"object\",\n" +
            "       \"logicality\": \"object\",\n" +
            "       \"knowledge_mastery\": \"object\"\n" +
            "     },\n" +
            "     \"evaluation\": {\n" +
            "       \"highlight\": \"object\",\n" +
            "       \"suggestion\": \"object\",\n" +
            "       \"reference_answer\": \"object\",\n" +
            "       \"key_knowledge\": \"object\"\n" +
            "     }\n" +
            "   }";

    private final AiProductRoleService aiProductRoleService;
    private final ObjectMapper objectMapper;

    @Value("${ai.service.ak}")
    private String ak;
@@ -51,8 +79,13 @@

    @PostConstruct
    public void init() {
        ConnectionPool connectionPool = new ConnectionPool(10, 30, TimeUnit.SECONDS);
        // 增加连接池大小和存活时间
        ConnectionPool connectionPool = new ConnectionPool(32, 60, TimeUnit.SECONDS);
        Dispatcher dispatcher = new Dispatcher();
        // 增加并发请求数量
        dispatcher.setMaxRequests(128);
        dispatcher.setMaxRequestsPerHost(32);

        this.service = ArkService.builder()
                .dispatcher(dispatcher)
                .connectionPool(connectionPool)
@@ -70,7 +103,7 @@
    }

    @Override
    public AiResponse start(String productRoleId, String content) {
    public AiResponse start(List<AiMessage> aiMessageDtoList,Integer type,String productRoleId, String content, String question) {
        if (!StringUtils.hasText(productRoleId)) {
            log.warn("productRoleId 不能为空");
            return buildErrorResponse(CODE_NOT_FOUND, "AI陪练不存在");
@@ -82,22 +115,137 @@
            return buildErrorResponse(CODE_NOT_FOUND, "AI陪练不存在");
        }

        String promptTemplate = aiProductRole.getPromptTemplate();
        String linkId = aiProductRole.getLinkId();

        if (!StringUtils.hasText(promptTemplate) || !StringUtils.hasText(linkId)) {
            log.warn("角色配置不完整，promptTemplate 或 linkId 为空，productRoleId: {}", productRoleId);
        String promptTemplate = "作为一个智能助手，请回答我提出的问题。";
        if (AiTypeEnum.QUESTION.getCode() ==  type){
            promptTemplate = aiProductRole.getPromptHead();
        }
        if (AiTypeEnum.ANSWER.getCode() ==  type){
            promptTemplate = aiProductRole.getPromptTemplate()+question;
        }
        log.info("promptTemplate: {}", promptTemplate);
        String linkId = aiProductRole.getLinkId();
        String jsonTemplate = aiProductRole.getJsonTemplate();

        if (
                !StringUtils.hasText(promptTemplate)
                || !StringUtils.hasText(linkId)
                || !StringUtils.hasText(jsonTemplate)
        ) {
            log.warn("角色配置不完整，promptTemplate 或 linkId 或 jsonTemplate为空，productRoleId: {}", productRoleId);
            return buildErrorResponse(CODE_ERROR, "角色配置不完整");
        }

        return question(promptTemplate, linkId, content);
        AiRequest aiRequest = new AiRequest();
        aiRequest.setPromptTemplate(promptTemplate);
        aiRequest.setJsonTemplate(jsonTemplate);
        aiRequest.setLinkId(linkId);
        aiRequest.setContent(content);
        if (CollUtil.isNotEmpty(aiMessageDtoList)){
            aiRequest.setAiMessageDtoList(aiMessageDtoList);
        }
        return this.question(aiRequest);
    }

    @Override
    public AiResponse question(String promptTemplate, String linkId, String content) {
    public AiResponse question(AiRequest aiRequest) {
        String promptTemplate = aiRequest.getPromptTemplate();
        String linkId = aiRequest.getLinkId();
        String content = aiRequest.getContent();
        String jsonTemplate = aiRequest.getJsonTemplate();
        if (
                !StringUtils.hasText(promptTemplate)
                        || !StringUtils.hasText(linkId)
                        || !StringUtils.hasText(content)
                        || !StringUtils.hasText(jsonTemplate)
        ) {
            log.warn("请求参数不完整，promptTemplate: {}, linkId: {}, content: {}, jsonTemplate: {}", promptTemplate, linkId, content, jsonTemplate);
            return buildErrorResponse(CODE_ERROR, "请求参数不完整");
        }

        List<ChatMessage> messages = new ArrayList<>();
        ChatMessage systemMessage = ChatMessage.builder().role(ChatMessageRole.SYSTEM).content(promptTemplate).build();
        ChatMessage userMessage = ChatMessage.builder().role(ChatMessageRole.USER).content(content).build();
        messages.add(systemMessage);
        if (CollUtil.isNotEmpty(aiRequest.getAiMessageDtoList())){
            aiRequest.getAiMessageDtoList().forEach(aiMessageDto -> {
                ChatMessage message = ChatMessage.builder()
                        .role(aiMessageDto.getRole())
                        .content(aiMessageDto.getContent())
                        .build();
                messages.add(message);
            });
        }
        messages.add(userMessage);

        try {
            JsonNode schemaNode = objectMapper.readTree(jsonTemplate);
            ChatCompletionRequest.ChatCompletionRequestResponseFormat responseFormat = new ChatCompletionRequest.ChatCompletionRequestResponseFormat(
                    "json_schema",
                    new ResponseFormatJSONSchemaJSONSchemaParam(
                            "ai_response",
                            "json数据响应",
                            schemaNode,
                            true
                    )
            );
            ChatCompletionRequest chatCompletionRequest = ChatCompletionRequest.builder()
                    .model(linkId)
                    .messages(messages)
                    .stream(false)
                    .responseFormat(responseFormat)
                    .temperature(0.7) // 降低温度参数，提高确定性，可能提升速度
                    .topP(0.9)        // 调整topP参数
                    .maxTokens(2048)  // 减少最大token数
                    .frequencyPenalty(0.0)
                    .build();

            List<ChatCompletionChoice> choices = service.createChatCompletion(chatCompletionRequest).getChoices();
            String result = choices.stream()
                    .map(choice -> choice.getMessage().getContent())
                    .filter(contentObj -> contentObj != null)
                    .map(Object::toString)
                    .collect(Collectors.joining());
            Report report = this.extractReportData(result);
            return buildSuccessResponse(report, result);
        } catch (JsonProcessingException e) {
            log.error("初始化AI服务失败，JSON格式化输出初始化失败", e);
            return buildErrorResponse(CODE_ERROR, "AI服务调用失败");
        } catch (Exception e) {
            log.error("调用AI服务失败，modelId: {}", linkId, e);
            return buildErrorResponse(CODE_ERROR, "AI服务调用失败");
        }
    }

    public static void main(String[] args) {
        Report report = new Report();
        List<RadarDataItem> radarDataItems = new ArrayList<>();

        RadarDataItem item1 = new RadarDataItem();
        item1.setName("A");
        item1.setCode("A");
        item1.setScore("80");
        radarDataItems.add(item1);

        RadarDataItem item2 = new RadarDataItem();
        item2.setName("A");
        item2.setCode("A");
        item2.setScore("80");
        radarDataItems.add(item2);
        report.setRadarDataItems(radarDataItems);

        System.out.println(JSONUtil.parse( report));

    }

    @Override
    public void streamQuestion(AiRequest aiRequest, Consumer<AiResponse> callback) {

        String promptTemplate = aiRequest.getPromptTemplate();
        String linkId = aiRequest.getLinkId();
        String content = aiRequest.getContent();
        if (!StringUtils.hasText(promptTemplate) || !StringUtils.hasText(linkId) || !StringUtils.hasText(content)) {
            log.warn("请求参数不完整，promptTemplate: {}, linkId: {}, content: {}", promptTemplate, linkId, content);
            return buildErrorResponse(CODE_ERROR, "请求参数不完整");
        }

        final List<ChatMessage> messages = new ArrayList<>();
@@ -106,28 +254,124 @@
        messages.add(systemMessage);
        messages.add(userMessage);

        ChatCompletionRequest chatCompletionRequest = ChatCompletionRequest.builder()
                .model(linkId)
                .messages(messages)
                .temperature(1.0)
                .topP(0.7)
                .maxTokens(4096)
                .frequencyPenalty(0.0)
                .build();
        try {
            JsonNode schemaNode = objectMapper.readTree(SCHEMA_JSON);
            ChatCompletionRequest.ChatCompletionRequestResponseFormat responseFormat = new ChatCompletionRequest.ChatCompletionRequestResponseFormat(
                    "json_schema",
                    new ResponseFormatJSONSchemaJSONSchemaParam(
                            "ai_response",
                            "json数据响应",
                            schemaNode,
                            true
                    )
            );
            ChatCompletionRequest chatCompletionRequest = ChatCompletionRequest.builder()
                    .model(linkId)
                    .messages(messages)
                    .stream(true) // 启用流式响应
                    .responseFormat(responseFormat)
                    .temperature(0.7)
                    .topP(0.9)
                    .maxTokens(2048)
                    .build();

            service.streamChatCompletion(chatCompletionRequest)
                    .doOnError(Throwable::printStackTrace) // 处理错误
                    .blockingForEach(response -> {
                        AiResponse partialResponse = new AiResponse();
                        if (response.getChoices() != null && !response.getChoices().isEmpty()) {
                            String responseStr = String.valueOf(response.getChoices().get(0).getMessage().getContent());
                            if (responseStr != null) {
                                // 构造部分响应并回调
                                partialResponse = buildGOINGONResponse(responseStr);
                            }
                        }else{
                            partialResponse = buildPartialResponse("成功");
                        }
                        callback.accept(partialResponse);
                    });
//            service.streamChatCompletion(chatCompletionRequest)
//                    .doOnError(throwable -> {
//                        log.error("流式调用AI服务失败", throwable);
//                        callback.accept(buildErrorResponse(CODE_ERROR, "AI服务调用失败"));
//                    })
//                    .subscribe(chatCompletionChunk -> {
//                        // 处理每个数据块
//                        Object chunkContent = chatCompletionChunk.getChoices().get(0).getMessage().getContent();
//                        // 构造部分响应并回调
//                        AiResponse partialResponse = buildGOINGONResponse(chunkContent);
//                        callback.accept(partialResponse);
//                    });
        } catch (Exception e) {
            log.error("调用AI服务失败", e);
            callback.accept(buildErrorResponse(CODE_ERROR, "AI服务调用失败"));
        }
    }

    private AiResponse buildGOINGONResponse(Object chunkContent) {
        AiResponse response = new AiResponse();
        response.setCode(CODE_GOING_ON);
        response.setDescription("成功");
        response.setResContext(chunkContent.toString());
        return response;
    }

    private AiResponse buildPartialResponse(Object chunkContent) {
        AiResponse response = new AiResponse();
        response.setCode(CODE_SUCCESS);
        response.setDescription("成功");
        response.setResContext(chunkContent.toString());
        return response;
    }


    private static final Pattern JSON_PATTERN = Pattern.compile(
            "<\\|FunctionCallBegin\\|>(.*?)<\\|FunctionCallEnd\\|>",
            Pattern.DOTALL
    );

    @Override
    public Report extractReportData(String modelOutput) {
        Matcher matcher = JSON_PATTERN.matcher(modelOutput);
        if (!matcher.find()) {
            log.warn("未匹配到FunctionCall内容，原始输出长度: {}", modelOutput);
            return null;
        }

        String jsonContent = matcher.group(1);
        log.debug("提取到的JSON内容长度: {}", jsonContent.length());

        try {
            List<ChatCompletionChoice> choices = service.createChatCompletion(chatCompletionRequest).getChoices();
            String result = choices.stream()
                    .map(choice -> choice.getMessage().getContent())
                    .filter(contentObj -> contentObj != null)
                    .map(Object::toString)
                    .collect(Collectors.joining());

            return buildSuccessResponse(result);
        } catch (Exception e) {
            log.error("调用AI服务失败，modelId: {}, content: {}", linkId, content, e);
            return buildErrorResponse(CODE_ERROR, "AI服务调用失败");
            return objectMapper.readValue(jsonContent, Report.class);
        } catch (JsonProcessingException e) {
            log.error("JSON解析失败，原始内容长度: {}", jsonContent.length(), e);
            Report repairedReport = tryRepairTruncatedJson(jsonContent);
            if (repairedReport != null) {
                log.info("成功修复截断的JSON");
                return repairedReport;
            }
            return null;
        }
    }

    private Report tryRepairTruncatedJson(String truncatedJson) {
        String[] repairAttempts = {
                truncatedJson + "\"}}}",
                truncatedJson + "}}}",
                truncatedJson + "}}"
        };

        for (String attempt : repairAttempts) {
            try {
                return objectMapper.readValue(attempt, Report.class);
            } catch (JsonProcessingException e) {
                log.debug("修复尝试失败: {}", attempt);
                continue;
            }
        }

        log.warn("无法修复截断的JSON，原始内容长度: {}", truncatedJson.length());
        return null;
    }

    private AiResponse buildErrorResponse(String code, String description) {
@@ -144,4 +388,13 @@
        response.setResContext(result);
        return response;
    }

    private AiResponse buildSuccessResponse(Report report, String result) {
        AiResponse response = new AiResponse();
        response.setCode(CODE_SUCCESS);
        response.setDescription("成功");
        response.setResContext(result);
        response.setReport(report);
        return response;
    }
}