跳至主要內容

Spring AI Alibaba

HZY2026/4/3大约 4 分钟Javajava

基础概念

快速入门

so eary

相关概念

  • token
  • 模型参数:
    • 温度: 控制文本生成的随机性，值越低越准确，
    • 核采样
    • 流式输出
    • message
      • system
      • user
      • assiant
      • tool
    • prompt

SpringAIAlibaba

  • 核心升级: 加入了Graph

基础组件

Message

    @GetMapping("/message")
    public String message(@RequestParam(name = "query") String query){
        SystemMessage systemMessage = new SystemMessage("你是一个智能机器人");
        UserMessage userMessage = new UserMessage(query);

        String call = chatModel.call(systemMessage, userMessage );

        return call;
    }

Prompt

    @GetMapping("/chatOptions")
    public ChatResponse chatOptions(@RequestParam(name = "query") String query){
        SystemMessage systemMessage = new SystemMessage("你是一个智能机器人");
        UserMessage userMessage = new UserMessage(query);
        var options = new ZhiPuAiChatOptions.Builder()
                .model("glm-4.5")
                .temperature(0.0)
                .maxTokens(15536)
                .build();
                // 这里的配置会被application.yaml覆盖，目前不知道为啥，学到后面再说
        return chatModel.call(new Prompt(List.of(systemMessage,
                userMessage),options));

        // ChatResponse chatResponse = chatModel.call(new Prompt(List.of(systemMessage,
        // userMessage),options));

        // var result = chatResponse.getResult().getOutput().getText();
        // return result;
    }

ChatModel

流式

    // 注意：一定要配置字符集
    @GetMapping("/flux")
    public Flux<String> flux(@RequestParam(name = "query") String query){
        SystemMessage systemMessage = new SystemMessage("你是一个智能机器人");
        UserMessage userMessage = new UserMessage(query);
        var call = chatModel.stream(systemMessage, userMessage );
        return call;
    }

ChatClient


     // constructor 注入
    private final ChatClient chatClient;

    public ZhiPuAIClientController(ChatClient.Builder builder  ) {
        this.chatClient = builder.build();
    }

    @GetMapping("/simple")
    public String simple(@RequestParam(name = "query") String query){
        var options = new ZhiPuAiChatOptions.Builder()
                .model("glm-4.5")
                .temperature(0.0)
                .maxTokens(15536)
                .build();
//        不优雅
//        SystemMessage systemMessage = new SystemMessage("你是一个智能机器人");
//        UserMessage userMessage = new UserMessage(query);
//        var result = chatClient.prompt(new Prompt(List.of(systemMessage,
//                userMessage),options))
//                .call().content();
        var result = chatClient.prompt()
                .system("你是一个ai助手")
                .user(query)
                .options(options)
                .call().content();
                //当然这里也可以返回 ChatResponse
        return result;
    }

Advisors

  • 对话记忆
  • 敏感词过滤
  • RAG检索
介绍

其实就是一个请求前后增强器

  • getOrder值越小越早请求
  • 这是一个调用链
  • CallAdvisor 和 StreamAdvisor
  • BaseAdvisor 是上面俩合起来的封装
记忆增强
  • 如何区分不同用户
    • session_ID
// 简单写一下，很多功能没有打磨
public class SimpleMessageChatAdvisor implements BaseAdvisor {

    public static List<Advisor> SimpleMessageChatAdvisor;
    private static Map<String, List<Message>> chatMemory = new HashMap<String, List<Message>>();

    @Override
    public ChatClientRequest before(ChatClientRequest chatClientRequest, AdvisorChain advisorChain) {
        //

        //
        List<Message> messages = chatMemory.get("haiziyao");
        if (messages == null) {
            messages = new ArrayList<Message>();
            chatMemory.put("haiziyao", messages);
        }
        List<Message> instructions = chatClientRequest.prompt().getInstructions();
        messages.addAll(instructions);
        //
        return chatClientRequest.mutate()
                .prompt(chatClientRequest.prompt().mutate().messages(messages).build())
                .build();
    }

    @Override
    public ChatClientResponse after(ChatClientResponse chatClientResponse, AdvisorChain advisorChain) {

        // 找到会话记录
        List<Message> messages = chatMemory.get("haiziyao");
        if (messages == null) {
            messages = new ArrayList<>();
        }
        // 获取response中ai的消息
        if(Objects.isNull(chatClientResponse)){
            return chatClientResponse;
        }
        // 存入
        AssistantMessage output = chatClientResponse.chatResponse()
                .getResult()
                .getOutput();
        messages.add(output);

        return chatClientResponse;
    }

    @Override
    public int getOrder() {
        return 0;
    }
}

    @GetMapping("simplemessageclient")
    public String demo(@RequestParam(name ="query")String query){
        return chatClient.prompt()
                .user(query)
                .system("你是个“疯狂星期四文案”生成器，你需要做的就是生成vivo50的文案，无论用户说啥，你都用文案回复")
                .advisors(advisorSpec -> advisorSpec.param("conID","haiziyao"))
                .advisors(new SimpleMessageChatAdvisor())
                .call()
                .content();
    }
使用官方Advisor

MessageCharMemoryAdvisor


    @RestController
    @RequestMapping("/chatclientplus")
    public class ZhiPuCharMemoryController {

        private final ChatClient chatClient;


        public ZhiPuCharMemoryController(ChatClient.Builder builder) {

            MessageWindowChatMemory memory = MessageWindowChatMemory.builder()
                    .maxMessages(20)
                    .build();

            MessageChatMemoryAdvisor advisor = MessageChatMemoryAdvisor.builder(memory).build();


            this.chatClient = builder
                    .defaultAdvisors(advisor)
                    .build();
        }

        @GetMapping("simplemessageclient")
        public String demo(@RequestParam(name ="query")String query,
                        @RequestParam(name = "conversationId")String conversationId){
            return chatClient.prompt()
                    .user(query)
                    .advisors(advisorSpec -> advisorSpec.param(ChatMemory.CONVERSATION_ID,conversationId))
                    .call()
                    .content();
        }
    }
  • 这里其实几句话说不完
  • 如果想自定义，就去实现ChatMemoryRepository接口，实现所有方法
  • MessageWindowMemory的局限性或特性
    • 对话记忆全量保存，只是做了max限制而已

Prompt Template

  • 提示词拼接不够优雅，所以封装起来
  • 不就是format吗
  • 暂时不给太多例子，因为这个也不好用。
  • 我更倾向于那种read from file

RAG

  • Retrieval-Augmented Generation 检索增强生成

  • 文本转向量

  • Embedding嵌入

  • EmbeddingModle嵌入模型

  • 向量数据库库

  • 向量维度

quick-start

构建向量知识库
@GetMapping("/import")
    public String importCoffee() {
        try {
            ClassPathResource resource = new ClassPathResource("QA.csv");
            InputStreamReader isr = new InputStreamReader(resource.getInputStream());
            CSVParser csvParser = CSVFormat.DEFAULT.builder()
                    .setHeader()
                    .setSkipHeaderRecord(true)
                    .build()
                    .parse(isr);
            List<Document> documents = new ArrayList<Document>();
            for (CSVRecord record : csvParser) {
                String question = record.get("问题");
                String answer = record.get("回答");
                String content = "问题:" + question + "\n回答"+ answer;
                Document document = new Document(content);
                documents.add(document);
            }
            csvParser.close();
            vectorStore.add(documents);
            return "成功导入 " + documents.size()+ "条数据到向量数据库";
        }catch (Exception e) {
            e.printStackTrace();
            return "导入失败"+e.getMessage();
        }
    }

ToolCalling

    @Tool(description = "通过时区id获取当前时间")
    public String getTimeByZoneId(@ToolParam(description = "时区id, 比如 Asia/Shanghai")
                                  String zoneId) {
        ZoneId zid = ZoneId.of(zoneId);
        ZonedDateTime zonedDateTime = ZonedDateTime.now(zid);
        DateTimeFormatter formatter = DateTimeFormatter.ofPattern("yyyy-MM-dd HH:mm:ss z");
        return zonedDateTime.format(formatter);
    }
  • 这个怎么测都测不出来，烦死我了

MCP

  • MCP host
  • MCP client
  • MCP server

调包侠进阶

Graph

  • 为什么需要Graph

    • 实际的工作是一个工作流
    • 需要步骤
    • 需要各种行为

  • 核心概念

    • 节点
    • 边（条件边）

KeyStrategyFactory


KeyStrategyFactory keyStrategyFactory = new KeyStrategyFactoryBuilder()
        .addStrategy("input1", KeyStrategy.REPLACE)
        .addStrategy("input2",KeyStrategy.MERGE)
        .addStrategy("input3",KeyStrategy.APPEND)
        .build();
  • 替换
  • 合并
  • 追加

NodeAction & AsyncNodeAction


stateGraph.addNode("node1", AsyncNodeAction.node_async(node ->{
    return Map.of("input1",1);
}));

StateGraph


    @Bean("simpleGraph")
    public CompiledGraph simpleGraph() throws GraphStateException {

        KeyStrategyFactory keyStrategyFactory = new KeyStrategyFactoryBuilder()
                .addStrategy("sentence",KeyStrategy.REPLACE)
                .addStrategy("word",KeyStrategy.REPLACE)
                .build();

        StateGraph stateGraph = new StateGraph("simpleGraph", keyStrategyFactory);

        stateGraph.addNode("SentenceConstructionNode",AsyncNodeAction.node_async(new SentenceConstrucrionNode(chatClient)));
        stateGraph.addNode("translateNode",AsyncNodeAction.node_async(new TranslateNode(chatClient)));


        stateGraph.addEdge(StateGraph.START,"SentenceConstructionNode");
        stateGraph.addEdge("SentenceConstructionNode","translateNode");
        stateGraph.addEdge("translateNode",StateGraph.END);
        return stateGraph.compile();
    }

循环分支控制语句

条件边

stateGraph.addConditionalEdges("评估笑话", AsyncEdgeAction.edge_async(
                        state -> state.value("result","优秀")),
                Map.of("优秀",StateGraph.END,
                        "不够优秀","优化笑话"));
循环

就是基于条件实现的，只不过在状态中存储了一个用来检验循环次数的值而已

状态存储

对于一些信息或者数据，我们可以使用追加，就是定义keystrateFactory的时候

也可以使用我们的状态存储，可以做对话区分

默认使用的都是ThreadLocal

可视化

项目

亥子曜工厂 · HZY-Lab
Copyright © 2026 HZY