目录#

• 项目简介
• 核心功能
• 技术架构
• 项目结构
• 快速开始
• 核心机制解析
• API 文档
• 测试
• 项目亮点
• Roadmap
• 免责声明

🩺 智能在线问诊#

小智作为训练有素的医疗顾问，针对患者描述的健康问题，基于当前临床实践与研究提供：

• 病因分析与诊断方向建议
• 治疗方案与预防措施参考（含药品名称、剂量、疗程说明）
• 就医时机判断，明确指示是否需要进一步检查或现场就诊

🧭 AI 分导诊#

根据患者描述的病情和就医需求，智能推荐最合适的科室，帮助患者快速找到正确的就医入口，减少科室间反复奔波。

📅 智能预约挂号#

患者通过自然语言即可完成挂号全流程：

1. 实时查询目标科室、指定日期与时段（上午 / 下午）的号源情况
2. 自动收集姓名、身份证号、科室、日期、时段、医生等预约信息，确认无误后才执行预约
3. 防重复预约校验：相同科室 + 时段已有预约时自动拦截并提示

❌ 取消预约#

通过对话即可快速取消已有预约，系统自动查询并删除对应记录，全程无需人工干预。

💬 多轮对话与记忆管理#

• 借助 MongoDB 持久化聊天记忆，同一会话内上下文连贯
• 每个会话窗口最多保留最近 20 条消息
• 不同用户（memoryId）之间的记忆相互隔离，互不干扰

架构总览#

技术栈#

一次对话的完整链路#

以”挂号”为例：

环境要求#

1. 初始化 MySQL#

在 MySQL 中执行：

1
CREATE DATABASE IF NOT EXISTS guiguxiaozhi DEFAULT CHARACTER SET utf8mb4;
2
USE guiguxiaozhi;
3

4
CREATE TABLE IF NOT EXISTS appointment (
5
    id          BIGINT AUTO_INCREMENT PRIMARY KEY COMMENT '主键',
6
    username    VARCHAR(64)  NOT NULL COMMENT '患者姓名',
7
    id_card     VARCHAR(32)  NOT NULL COMMENT '身份证号',
8
    department  VARCHAR(64)  NOT NULL COMMENT '预约科室',
9
    date        VARCHAR(16)  NOT NULL COMMENT '预约日期，如 2025-04-14',
10
    time        VARCHAR(8)   NOT NULL COMMENT '时段：上午 / 下午',
11
    doctor_name VARCHAR(64)  DEFAULT NULL COMMENT '医生姓名'
12
) ENGINE = InnoDB DEFAULT CHARSET = utf8mb4 COMMENT ='预约挂号记录表';

如需修改数据库连接，编辑 guiguXiaozhi_backend/src/main/resources/application.properties 中的 spring.datasource.* 配置。

2. 配置环境变量#

也可以在 application.properties 中直接填入密钥（注意勿提交到公开仓库）。

3. 初始化知识库（RAG）#

1. 准备医院知识文档（Markdown / PDF 均可，项目已引入 PDFBox 解析器）
2. 打开测试类 EmbeddingTest，将 testUploadKnowledgeLibrary() 中的文档路径替换为你的知识库文件路径（如 医院信息.md、科室信息.md、神经内科.md）
3. 运行该测试方法：文档将被切分、向量化并写入 Pinecone 索引

4. 启动后端#

1
cd guiguXiaozhi_backend
2
mvn spring-boot:run

服务启动于 http://localhost:8080。

5. 启动前端#

1
cd guiguXiaozhi_frontend
2
npm install
3
npm run dev

访问 http://localhost:5173 即可开始与小智对话。Vite 已将 /api 请求代理至后端 8080 端口。

RAG 检索增强生成#

• 医院信息、科室介绍等知识文档经 text-embedding-v3 向量化后存入 Pinecone
• 每次对话由 EmbeddingStoreContentRetriever 检索相关知识片段（Top-1，相似度 ≥ 0.8）注入大模型上下文，提升回答的专业性与准确性
• 相关配置：EmbeddingStoreConfig（向量存储）、XiaozhiAgentConfig（ContentRetriever）

Function Calling 工具调用#

大模型通过 LangChain4j 的 @Tool 注解自动识别调用意图，将自然语言映射为数据库操作（AppointmentTools）：

System Prompt 中约定了严格的执行规则：预约前必须确认患者姓名、身份证号、科室、日期、时段（医生可选），信息不全时主动向用户追问，杜绝”信息不全直接下单”。

流式输出#

• 后端 POST /xiaozhi/chat 返回 Flux<String>（produces = "text/stream;charset=utf-8"），基于 LangChain4j Reactor 适配器逐 token 下发
• 前端通过 axios onDownloadProgress 增量读取响应体并实时渲染，实现类 ChatGPT 的打字机效果，显著改善等待体验

对话记忆（MongoDB 持久化）#

• MongoChatMemoryStore 实现 LangChain4j 的 ChatMemoryStore 接口，将对话消息序列化为 JSON 存入 MongoDB 的 chat_messages 集合
• MessageWindowChatMemory 按 memoryId 隔离，每会话保留最近 20 条消息（滑动窗口）
• 前端在浏览器 localStorage 中保存会话标识，刷新页面后对话上下文依然连续

Table of Contents#

• Project Overview
• Key Features
• Technical Architecture
• Project Structure
• Quick Start
• Core Mechanisms
• API Documentation
• Tests
• Highlights
• Roadmap
• Disclaimer

Project Overview#

Guigu Xiaozhi is an intelligent AI medical companion assistant built on the LangChain4j framework, designed for Peking Union Medical College Hospital. Instead of navigating complex hospital procedures, patients simply describe their symptoms and needs in conversation, and Xiaozhi handles the rest:

• AI Medical Consultation — evidence-based cause analysis, treatment suggestions and guidance on when to seek care
• Smart Department Triage — recommends the most suitable department based on symptoms
• Appointment Booking — queries availability, confirms details, and completes or cancels bookings through conversation

The project fully integrates Large Language Models (LLM), Retrieval-Augmented Generation (RAG), Function Calling, streaming output and multi-turn chat memory — an end-to-end showcase of LangChain4j’s core features. A single conversation covers the complete journey from consultation and triage to appointment booking, with no manual database operations required.

This is a hands-on AI application project. Some parts, such as slot querying, are demo implementations (see Roadmap). It serves as a reference implementation for a LangChain4j + Spring Boot + Vue full-stack AI application.

Key Features#

Technical Architecture#

Project Structure#

1
guiguXiaozhi/
2
├── guiguXiaozhi_backend/                    # Backend service (Spring Boot 3.2)
3
│   ├── pom.xml                              # Maven dependencies (Spring Boot / LangChain4j BOM)
4
│   └── src/
5
│       ├── main/java/com/atguigu/java/ai/langchain4j/
6
│       │   ├── XiaozhiApp.java              # Application entry
7
│       │   ├── assistant/                   # AI Service interface definitions
8
│       │   │   ├── XiaozhiAgent.java        # ★ Main agent (streaming + tools + RAG + memory)
9
│       │   │   ├── MemoryChatAssistant.java # Memory feature example
10
│       │   │   └── SeparateChatAssistant.java # Per-user isolation example
11
│       │   ├── config/                      # Bean wiring
12
│       │   │   ├── XiaozhiAgentConfig.java  # ChatMemoryProvider + RAG ContentRetriever
13
│       │   │   └── EmbeddingStoreConfig.java# Pinecone embedding store
14
│       │   ├── controller/XiaozhiController.java  # POST /xiaozhi/chat
15
│       │   ├── tools/                       # Function Calling tools
16
│       │   │   ├── AppointmentTools.java    # ★ Book / cancel / query slots
17
│       │   │   └── CalculatorTools.java     # Calculator example
18
│       │   ├── store/MongoChatMemoryStore.java  # ★ MongoDB memory persistence
19
│       │   ├── entity/ mapper/ service/     # Booking business layer (MyBatis-Plus)
20
│       │   └── bean/                        # ChatForm / ChatMessages
21
│       └── main/resources/
22
│           ├── application.properties       # Model / DB / vector store config
23
│           ├── xhaozhi-prompt-template.txt  # ★ System Prompt (role & rules)
24
│           └── mapper/AppointmentMapper.xml
25
└── guiguXiaozhi_frontend/                   # Frontend (Vue 3 + Vite)
26
    └── src/
27
        ├── App.vue                          # App entry
28
        ├── components/ChatWindow.vue        # ★ Core chat component (streaming render / session)
29
        └── main.js                          # Element Plus registration

Quick Start#

1
CREATE DATABASE IF NOT EXISTS guiguxiaozhi DEFAULT CHARACTER SET utf8mb4;
2
USE guiguxiaozhi;
3

4
CREATE TABLE IF NOT EXISTS appointment (
5
    id          BIGINT AUTO_INCREMENT PRIMARY KEY COMMENT 'Primary key',
6
    username    VARCHAR(64)  NOT NULL COMMENT 'Patient name',
7
    id_card     VARCHAR(32)  NOT NULL COMMENT 'ID card number',
8
    department  VARCHAR(64)  NOT NULL COMMENT 'Department',
9
    date        VARCHAR(16)  NOT NULL COMMENT 'Appointment date, e.g. 2025-04-14',
10
    time        VARCHAR(8)   NOT NULL COMMENT 'Period: morning / afternoon',
11
    doctor_name VARCHAR(64)  DEFAULT NULL COMMENT 'Doctor name'
12
) ENGINE = InnoDB DEFAULT CHARSET = utf8mb4 COMMENT ='Appointment records';

1
cd guiguXiaozhi_backend
2
mvn spring-boot:run

1
cd guiguXiaozhi_frontend
2
npm install
3
npm run dev

Core Mechanisms#

API Documentation#

Knife4j (OpenAPI 3) is integrated. After startup, visit:

• Knife4j UI: http://localhost:8080/doc.html
• OpenAPI JSON: http://localhost:8080/v3/api-docs

Core endpoint:

Tests#

The backend ships a set of unit tests (src/test/java) covering every stage of the AI pipeline:

Highlights#

• Full-stack AI integration: RAG + Function Calling + streaming output + multi-turn memory — one project covering LangChain4j’s core features
• Multi-model support: Alibaba Cloud Qwen, DeepSeek-V3 (DashScope compatible mode) and local Ollama are all supported and switchable
• Deep business-AI integration: natural language directly drives database CRUD with no extra middle layer
• Production-grade persistence: MongoDB for chat history, MySQL for business data — clear separation of concerns
• Strict prompt engineering: the System Prompt defines role, response format and boundary rules, effectively reducing hallucinations and out-of-scope behavior

Roadmap#

• Integrate a real doctor scheduling system for live slot queries and waiting queue management
• Support image input (multimodal) for remote preliminary diagnosis in departments like dermatology
• Introduce user authentication and link patients’ historical visit records for personalized consultations
• Connect to the hospital HIS system to complete the data chain of reports, prescriptions, etc.
• Refactor into microservices for horizontal scaling under high concurrency

Disclaimer#

This is an AI technology demo and educational project. AI-generated content is for reference only and does not constitute a medical diagnosis. If you have health concerns, please visit a qualified medical institution in time.

Last Updated: 2026.03.20