背景 Background

在之前的v1版本中，虽然使用纯SpringBoot原生框架实现了RAG搜索，但仍有以下问题：

• 调用大模型接口使用同步模式，需要大模型回复完才把内容进行前端加载，看不到中间结果+需要等待时间长

• 没有上下文记忆功能

• 不支持markdown格式

• 不支持联网搜索

• 手动维护mysql知识库较繁琐

In the previous v1 version, although the RAG search was implemented using the native SpringBoot framework, there were still the following issues:

• Calling large model APIs used a synchronous mode, which required the model to finish generating a response before loading it on the frontend. This caused long waiting times without any intermediate output.

• No context memory functionality.

• No support for Markdown format.

• No support for online search.

• Manual maintenance of the MySQL knowledge base was cumbersome.

架构优化点 Architecture Optimization Points

本次v2优化，添加了以下功能：

• 引入了SpringAI，集成了知识库维护、上下文记忆等功能，无需手动实现

• 引入了SSE(server-sent events)优化前端，使得模型回答途中的流输出得到前端加载，同时支持markdown格式

• 引入了SearXNG搜索引擎，能支持联网查询，并将相关结果整合到回答中
  

In this v2 upgrade, the following features were added:

• Introduced SpringAI, which integrates features such as knowledge base maintenance and context memory without manual implementation.

• Introduced SSE (Server-Sent Events) to optimize the frontend, enabling streaming output of model responses to be loaded in real-time, with Markdown format support.

• Integrated the SearXNG search engine to support online queries and incorporate relevant results into answers.

引入MCP进行优化 Introducing MCP for optimaztion

进一步优化，通过引入MCP带来以下好处：

• 将查询知识库和联网查询改成MCP功能，由LLM自身决定是否调用（更智能、灵活），而不是手动调用知识库和网络查询

• 提供其他MCP相关支持提高模型能力

We further optimize by introducing MCP to bring more benefits as follows:

• Replace querying the knowledge base and online search with MCP functionality, allowing the LLM to decide whether to invoke it on its own. (smarter & more flexible), rather than manually doing knowledge & Internet searches

• Provide additional MCP-related support to enhance the model's capabilities.

待支持功能 Features to be supported

待支持相关功能：

• 提供更多MCP功能获取网页数据：博客信息、可视化数据、新闻总结等

• 将聊天功能做成弹窗，在每个页面都嵌入支持

• 实现会话隔离
  

Planned features to be supported:

• Provide more MCP functions for website's data like: blog information, visualization data and news summaries.

• Implement the chat feature as a popup window and embed it on every page for support.

• Implement conversation isolation.
  

开发期间挑战 Challenges During Development

• 知识库redis-stack和其他数据使用的redis需要分开，代码单独处理

• tomcat版本问题，需要从直接的9.x版本升级到11.x版本，同时将jdk8升级到jdk21

• SpingBoot+thymeleaf前端传contextPath参数问题，和之前版本不同

• 服务器部署后发送SSE问题，需要对Nginx和后端代码进行调整

• SSE需要维持心跳避免连接被关闭
  

• The Redis stack used for the knowledge base needed to be separated from other Redis data, requiring separate handling in code.

• Tomcat version issues: needed to upgrade from the direct 9.x version to 11.x, and also upgrade JDK from 8 to 21.

• ContextPath parameter passing issue in SpringBoot + Thymeleaf frontend, which differs from previous versions.

• SSE transmission issues after server deployment required adjustments to both Nginx and backend code.

• SSE needs heartbeat to prevent connection from being closed.