已部署上线：https://ppxdada.online
开发了一个基于 Spring Boot、Redis、ChatGLM、RxJava、SSE、Vue3、Arco Design 和 Pinia 的 AI 答题应用平台。用户可以通过 AI 快速生成题目并制作应用，管理员审核后，用户可以在线答题，并根据多种评分算法或 AI 得到回答总结。管理员还可以集中管理全站内容并进行统计分析。
An AI quiz application platform was developed using Spring Boot, Redis, ChatGLM, RxJava, SSE, Vue3, Arco Design, and Pinia. Users can quickly generate questions and create applications through AI. After administrator approval, users can take quizzes online and receive summarized answers based on various scoring algorithms or AI. Administrators can centrally manage site content and perform statistical analysis.
根据业务需求设计了用户、应用、题目、评分结果、用户答题等表结构。题目表采用 JSON 格式存储复杂嵌套题目和选项，便于维护和扩展，通过添加 appId 索引提高检索性能。
采用策略模式实现了多种用户回答评分算法（如统计得分、AI 评分等），全局执行器会扫描策略类上的自定义注解并选取相应策略，提升系统的可扩展性。
封装了基于 ChatGLM 的通用 AI 服务，通过配置类自动读取密钥配置，初始化 AI 客户端 Bean，便于全局使用。
封装了基于腾讯云 COS 的通用文件管理服务，通过配置类自动读取密钥配置，初始化 COS 客户端 Bean，便于全局使用。
通过优化 Prompt（如系统预设、少样本学习、任务拆解等）使 AI 返回 JSON 格式的题目内容，便于后端处理。由于 AI 生成题目较慢，选用 ChatGLM 的流式 API，并通过 SSE 实时推送单道题目给前端，提高用户体验。
基于 RxJava 的操作符链式调用处理 AI 异步数据流，通过 map 获取并处理字符串、filter 过滤空值、flatMap 映射字符串为单个字符，再通过括号平衡算法拼接出单道题目，使逻辑清晰简洁。
采用雪花算法为每次答题分配唯一 id，并通过数据库主键实现幂等设计，防止用户多次提交重复答案，避免生成重复数据。
使用 Caffeine 本地缓存答案 Hash 对应的 AI 评分结果，提高评分性能（从 10s 到 5ms），并通过 Redisson 分布式锁解决缓存击穿问题。
为应对用户答题数增长，基于 Sharding JDBC 配置取模分片算法，根据应用 id 对用户答题记录进行分表，提高单表查询性能和可扩展性。
由于服务器资源有限，为会员创建核心线程数更大的隔离线程池，确保普通用户不会影响会员使用 AI 的体验。
后端基于 MyBatis 注解自定义 SQL 实现热门应用和回答分布的统计，前端通过 ECharts 实现数据可视化。
使用 Knife4j 和 Swagger 自动生成后端接口文档，并通过编写 ApiOperation 等注解补充接口说明，避免了手动编写和维护文档的麻烦。
Designed table structures for users, applications, questions, scoring results, and user responses based on business requirements. The questions table uses JSON format to store complex nested questions and options, facilitating maintenance and expansion. An appId index was added to improve retrieval performance.
Implemented various user response scoring algorithms (e.g., statistical scores, AI scoring) using the strategy pattern. A global executor scans custom annotations on strategy classes and selects the appropriate strategy, enhancing system scalability.
Encapsulated a general AI service based on ChatGLM. The service reads key configurations through configuration classes and initializes AI client beans for global use.
Encapsulated a general file management service based on Tencent Cloud COS. The service reads key configurations through configuration classes and initializes COS client beans for global use.
Optimized prompts (e.g., system presets, few-shot learning, task decomposition) to make AI return question content in JSON format for backend processing. Due to the slow AI question generation, ChatGLM's streaming API was selected, and questions were pushed to the frontend in real-time using SSE to improve user experience.
Processed AI asynchronous data streams using RxJava's operator chain calls. Used map to retrieve and process strings, filter to remove null values, and flatMap to map strings to individual characters. Then, concatenated single questions using a bracket balancing algorithm, resulting in clear and concise logic.
Used a snowflake algorithm to assign a unique ID to each quiz attempt and implemented idempotent design through database primary keys to prevent users from submitting duplicate answers, avoiding duplicate data generation.
Used Caffeine for local caching of AI scoring results corresponding to answer hashes, improving scoring performance (from 10 seconds to 5 milliseconds). Solved cache penetration issues using Redisson distributed locks.
To handle the growth in user quiz attempts, configured modulo sharding algorithms based on Sharding JDBC to shard user quiz records by application ID, improving single-table query performance and scalability.
Due to limited server resources, created an isolated thread pool with a larger core thread count for members to ensure that ordinary users do not affect members' AI usage experience.
Performed statistical analysis of popular applications and answer distributions using custom SQL with MyBatis annotations on the backend. Visualized data on the frontend using ECharts.
Automatically generated backend API documentation using Knife4j and Swagger. Supplemented interface descriptions by writing ApiOperation and other annotations, avoiding the hassle of manually writing and maintaining documentation.