Spring AI Alibaba Graph：中断！人类反馈介入，流程丝滑走完～

<properties>
    <spring-ai-alibaba.version>1.0.0.3-SNAPSHOT</spring-ai-alibaba.version>
</properties>

<dependencies>

    <dependency>
        <groupId>org.springframework.ai</groupId>
        <artifactId>spring-ai-autoconfigure-model-openai</artifactId>
    </dependency>

    <dependency>
        <groupId>org.springframework.ai</groupId>
        <artifactId>spring-ai-autoconfigure-model-chat-client</artifactId>
    </dependency>

    <dependency>
        <groupId>com.alibaba.cloud.ai</groupId>
        <artifactId>spring-ai-alibaba-graph-core</artifactId>
        <version>${spring-ai-alibaba.version}</version>
    </dependency>

    <dependency>
        <groupId>org.springframework.boot</groupId>
        <artifactId>spring-boot-starter-web</artifactId>
    </dependency>
</dependencies>

application.yml

server:
  port: 8080
spring:
  application:
    name: human-node
  ai:
    openai:
      api-key: ${AIDASHSCOPEAPIKEY}
      base-url: https://dashscope.aliyuncs.com/compatible-mode
      chat:
        options:
          model: qwen-max

config

OverAllState 中存储的字段

query：用户的问题
expandernumber：扩展的数量
expandercontent：扩展的内容
feedback：人类反馈的内容
humannextnode：人类反馈后的下一个节点
translatelanguage：翻译的目标语言，默认为英文
translatecontent：翻译的内容

定义 ExpanderNode，边的连接为：

START -> expander -> humanfeedback
humanfeedback -> translate
humanfeedback -> END
translate -> END

package com.spring.ai.tutorial.graph.human.config;

import com.alibaba.cloud.ai.graph.GraphRepresentation;
import com.alibaba.cloud.ai.graph.KeyStrategy;
import com.alibaba.cloud.ai.graph.KeyStrategyFactory;
import com.alibaba.cloud.ai.graph.StateGraph;
import com.alibaba.cloud.ai.graph.action.AsyncEdgeAction;
import com.alibaba.cloud.ai.graph.exception.GraphStateException;
import com.alibaba.cloud.ai.graph.state.strategy.ReplaceStrategy;
import com.spring.ai.tutorial.graph.human.dispatcher.HumanFeedbackDispatcher;
import com.spring.ai.tutorial.graph.human.node.ExpanderNode;
import com.spring.ai.tutorial.graph.human.node.HumanFeedbackNode;
import com.spring.ai.tutorial.graph.human.node.TranslateNode;
import org.slf4j.Logger;
import org.slf4j.LoggerFactory;
import org.springframework.ai.chat.client.ChatClient;
import org.springframework.context.annotation.Bean;
import org.springframework.context.annotation.Configuration;

import java.util.HashMap;
import java.util.Map;

import static com.alibaba.cloud.ai.graph.action.AsyncNodeAction.nodeasync;

/**
 * @author yingzi
 * @since 2025/6/13
 */
@Configuration
public class GraphHumanConfiguration {

    private static final Logger logger = LoggerFactory.getLogger(GraphHumanConfiguration.class);

    @Bean
    public StateGraph humanGraph(ChatClient.Builder chatClientBuilder) throws GraphStateException {
        KeyStrategyFactory keyStrategyFactory = () -> {
            HashMap<String, KeyStrategy> keyStrategyHashMap = new HashMap<>();
            // 用户输入
            keyStrategyHashMap.put("query", new ReplaceStrategy());
            keyStrategyHashMap.put("threadid", new ReplaceStrategy());

            keyStrategyHashMap.put("expandernumber", new ReplaceStrategy());
            keyStrategyHashMap.put("expandercontent", new ReplaceStrategy());

            // 人类反馈
            keyStrategyHashMap.put("feedback", new ReplaceStrategy());
            keyStrategyHashMap.put("humannextnode", new ReplaceStrategy());

            // 是否需要翻译
            keyStrategyHashMap.put("translatelanguage", new ReplaceStrategy());
            keyStrategyHashMap.put("translatecontent", new ReplaceStrategy());
            return keyStrategyHashMap;
        };

        StateGraph stateGraph = new StateGraph(keyStrategyFactory)
                .addNode("expander", nodeasync(new ExpanderNode(chatClientBuilder)))
                .addNode("translate", nodeasync(new TranslateNode(chatClientBuilder)))
                .addNode("humanfeedback", nodeasync(new HumanFeedbackNode()))

                .addEdge(StateGraph.START, "expander")
                .addEdge("expander", "humanfeedback")
                .addConditionalEdges("humanfeedback", AsyncEdgeAction.edgeasync((new HumanFeedbackDispatcher())), Map.of(
                        "translate", "translate", StateGraph.END, StateGraph.END))
                .addEdge("translate", StateGraph.END);

        // 添加 PlantUML 打印
        GraphRepresentation representation = stateGraph.getGraph(GraphRepresentation.Type.PLANTUML,
                "human flow");
        logger.info("\n=== expander UML Flow ===");
        logger.info(representation.content());
        logger.info("==================================\n");

        return stateGraph;
    }
}

node

ExpanderNode

package com.spring.ai.tutorial.graph.human.node;

import com.alibaba.cloud.ai.graph.NodeOutput;
import com.alibaba.cloud.ai.graph.OverAllState;
import com.alibaba.cloud.ai.graph.action.NodeAction;
import com.alibaba.cloud.ai.graph.async.AsyncGenerator;
import com.alibaba.cloud.ai.graph.streaming.StreamingChatGenerator;
import org.slf4j.Logger;
import org.slf4j.LoggerFactory;
import org.springframework.ai.chat.client.ChatClient;
import org.springframework.ai.chat.model.ChatResponse;
import org.springframework.ai.chat.prompt.PromptTemplate;
import reactor.core.publisher.Flux;

import java.util.Arrays;
import java.util.HashMap;
import java.util.List;
import java.util.Map;

/**
 * @author yingzi
 * @since 2025/6/13
 */

public class ExpanderNode implements NodeAction {

    private static final Logger logger = LoggerFactory.getLogger(ExpanderNode.class);

    private static final PromptTemplate DEFAULTPROMPTTEMPLATE = new PromptTemplate("You are an expert at information retrieval and search optimization.\nYour task is to generate {number} different versions of the given query.\n\nEach variant must cover different perspectives or aspects of the topic,\nwhile maintaining the core intent of the original query. The goal is to\nexpand the search space and improve the chances of finding relevant information.\n\nDo not explain your choices or add any other text.\nProvide the query variants separated by newlines.\n\nOriginal query: {query}\n\nQuery variants:\n");

    private final ChatClient chatClient;

    private final Integer NUMBER = 3;

    public ExpanderNode(ChatClient.Builder chatClientBuilder) {
        this.chatClient = chatClientBuilder.build();
    }

    @Override
    public Map<String, Object> apply(OverAllState state) {
        logger.info("expander node is running.");

        String query = state.value("query", "");
        Integer expanderNumber = state.value("expandernumber", this.NUMBER);

        Flux<ChatResponse> chatResponseFlux = this.chatClient.prompt().user((user) -> user.text(DEFAULTPROMPTTEMPLATE.getTemplate()).param("number", expanderNumber).param("query", query)).stream().chatResponse();

        AsyncGenerator<? extends NodeOutput> generator = StreamingChatGenerator.builder()
                .startingNode("expanderllmstream")
                .startingState(state)
                .mapResult(response -> {
                    String text = response.getResult().getOutput().getText();
                    List<String> queryVariants = Arrays.asList(text.split("\n"));
                    return Map.of("expandercontent", queryVariants);
                }).build(chatResponseFlux);
        return Map.of("expandercontent", generator);
    }

}

HumanFeedbackNode

package com.spring.ai.tutorial.graph.human.node;

import com.alibaba.cloud.ai.graph.OverAllState;
import com.alibaba.cloud.ai.graph.StateGraph;
import com.alibaba.cloud.ai.graph.action.NodeAction;
import org.slf4j.Logger;
import org.slf4j.LoggerFactory;

import java.util.HashMap;
import java.util.Map;

/**
 * @author yingzi
 * @since 2025/6/19
 */

public class HumanFeedbackNode implements NodeAction {

    private static final Logger logger = LoggerFactory.getLogger(HumanFeedbackNode.class);

    @Override
    public Map<String, Object> apply(OverAllState state) {
        logger.info("humanfeedback node is running.");
        HashMap<String, Object> resultMap = new HashMap<>();
        String nextStep = StateGraph.END;

        Map<String, Object> feedBackData = state.humanFeedback().data();
        boolean feedback = (boolean) feedBackData.getOrDefault("feedback", true);
        if (feedback) {
            nextStep = "translate";
        }

        resultMap.put("humannextnode", nextStep);
        logger.info("humanfeedback node -> {} node", nextStep);
        return resultMap;
    }
}

TranslateNode

package com.spring.ai.tutorial.graph.human.node;

import com.alibaba.cloud.ai.graph.NodeOutput;
import com.alibaba.cloud.ai.graph.OverAllState;
import com.alibaba.cloud.ai.graph.action.NodeAction;
import com.alibaba.cloud.ai.graph.async.AsyncGenerator;
import com.alibaba.cloud.ai.graph.streaming.StreamingChatGenerator;
import org.slf4j.Logger;
import org.slf4j.LoggerFactory;
import org.springframework.ai.chat.client.ChatClient;
import org.springframework.ai.chat.model.ChatResponse;
import org.springframework.ai.chat.prompt.PromptTemplate;
import reactor.core.publisher.Flux;

import java.util.Arrays;
import java.util.HashMap;
import java.util.List;
import java.util.Map;

/**
 * @author yingzi
 * @since 2025/6/13
 */

public class TranslateNode implements NodeAction {

    private static final Logger logger = LoggerFactory.getLogger(ExpanderNode.class);

    private static final PromptTemplate DEFAULTPROMPTTEMPLATE = new PromptTemplate("Given a user query, translate it to {targetLanguage}.\nIf the query is already in {targetLanguage}, return it unchanged.\nIf you don't know the language of the query, return it unchanged.\nDo not add explanations nor any other text.\n\nOriginal query: {query}\n\nTranslated query:\n");

    private final ChatClient chatClient;

    private final String  TARGETLANGUAGE= "English";

    public TranslateNode(ChatClient.Builder chatClientBuilder) {
        this.chatClient = chatClientBuilder.build();
    }

    @Override
    public Map<String, Object> apply(OverAllState state) {
        logger.info("translate node is running.");

        String query = state.value("query", "");
        String targetLanguage = state.value("translatelanguage", TARGETLANGUAGE);

        Flux<ChatResponse> chatResponseFlux = this.chatClient.prompt().user((user) -> user.text(DEFAULTPROMPTTEMPLATE.getTemplate()).param("targetLanguage", targetLanguage).param("query", query)).stream().chatResponse();
        AsyncGenerator<? extends NodeOutput> generator = StreamingChatGenerator.builder()
                .startingNode("translatellmstream")
                .startingState(state)
                .mapResult(response -> {
                    String text = response.getResult().getOutput().getText();
                    List<String> queryVariants = Arrays.asList(text.split("\n"));
                    return Map.of("translatecontent", queryVariants);
                }).build(chatResponseFlux);
        return Map.of("translatecontent", generator);
    }
}

edge

人类节点的下一个边是条件边，由 HumanFeedbackDispatcher 控制下一步跳转到哪一个节点

package com.spring.ai.tutorial.graph.human.dispatcher;

import com.alibaba.cloud.ai.graph.OverAllState;
import com.alibaba.cloud.ai.graph.StateGraph;
import com.alibaba.cloud.ai.graph.action.EdgeAction;

/**
 * @author yingzi
 * @since 2025/6/19
 */

public class HumanFeedbackDispatcher implements EdgeAction {
    @Override
    public String apply(OverAllState state) throws Exception {
        return (String) state.value("humannextnode", StateGraph.END);
    }
}

controller

GraphHumanController

CompileConfig.builder().saverConfig(saverConfig).interruptBefore(“humanfeedback”）：在人类反馈节点前断流
Sinks.Many<ServerSentEvent> sink：接收 Stream 数据

package com.spring.ai.tutorial.graph.human.controller;

import com.alibaba.cloud.ai.graph.CompileConfig;
import com.alibaba.cloud.ai.graph.CompiledGraph;
import com.alibaba.cloud.ai.graph.NodeOutput;
import com.alibaba.cloud.ai.graph.OverAllState;
import com.alibaba.cloud.ai.graph.RunnableConfig;
import com.alibaba.cloud.ai.graph.StateGraph;
import com.alibaba.cloud.ai.graph.async.AsyncGenerator;
import com.alibaba.cloud.ai.graph.checkpoint.config.SaverConfig;
import com.alibaba.cloud.ai.graph.checkpoint.constant.SaverConstant;
import com.alibaba.cloud.ai.graph.checkpoint.savers.MemorySaver;
import com.alibaba.cloud.ai.graph.exception.GraphRunnerException;
import com.alibaba.cloud.ai.graph.exception.GraphStateException;
import com.alibaba.cloud.ai.graph.state.StateSnapshot;
import com.spring.ai.tutorial.graph.human.controller.GraphProcess.GraphProcess;
import org.slf4j.Logger;
import org.slf4j.LoggerFactory;
import org.springframework.beans.factory.annotation.Autowired;
import org.springframework.beans.factory.annotation.Qualifier;
import org.springframework.http.MediaType;
import org.springframework.http.codec.ServerSentEvent;
import org.springframework.web.bind.annotation.GetMapping;
import org.springframework.web.bind.annotation.RequestMapping;
import org.springframework.web.bind.annotation.RequestParam;
import org.springframework.web.bind.annotation.RestController;
import reactor.core.publisher.Flux;
import reactor.core.publisher.Sinks;

import java.util.HashMap;
import java.util.Map;

/**
 * @author yingzi
 * @since 2025/6/13
 */
@RestController
@RequestMapping("/graph/human")
public class GraphHumanController {

    private static final Logger logger = LoggerFactory.getLogger(GraphHumanController.class);

    private final CompiledGraph compiledGraph;

    @Autowired
    public GraphHumanController(@Qualifier("humanGraph") StateGraph stateGraph) throws GraphStateException {
        SaverConfig saverConfig = SaverConfig.builder().register(SaverConstant.MEMORY, new MemorySaver()).build();
        this.compiledGraph = stateGraph
                .compile(CompileConfig.builder().saverConfig(saverConfig).interruptBefore("humanfeedback").build());    }

    @GetMapping(value = "/expand", produces = MediaType.TEXTEVENTSTREAMVALUE)
    public Flux<ServerSentEvent<String>> expand(@RequestParam(value = "query", defaultValue = "你好，很高兴认识你，能简单介绍一下自己吗？", required = false) String query,
                                                @RequestParam(value = "expandernumber", defaultValue = "3", required = false) Integer expanderNumber,
                                                @RequestParam(value = "threadid", defaultValue = "yingzi", required = false) String threadId) throws GraphRunnerException {
        RunnableConfig runnableConfig = RunnableConfig.builder().threadId(threadId).build();
        Map<String, Object> objectMap = new HashMap<>();
        objectMap.put("query", query);
        objectMap.put("expandernumber", expanderNumber);

        GraphProcess graphProcess = new GraphProcess(this.compiledGraph);
        Sinks.Many<ServerSentEvent<String>> sink = Sinks.many().unicast().onBackpressureBuffer();
        AsyncGenerator<NodeOutput> resultFuture = compiledGraph.stream(objectMap, runnableConfig);
        graphProcess.processStream(resultFuture, sink);

        return sink.asFlux()
                .doOnCancel(() -> logger.info("Client disconnected from stream"))
                .doOnError(e -> logger.error("Error occurred during streaming", e));
    }

    @GetMapping(value = "/resume", produces = MediaType.TEXTEVENTSTREAMVALUE)
    public Flux<ServerSentEvent<String>> resume(@RequestParam(value = "threadid", defaultValue = "yingzi", required = false) String threadId,
                                      @RequestParam(value = "feedback", defaultValue = "true", required = false) boolean feedBack) throws GraphRunnerException {
        RunnableConfig runnableConfig = RunnableConfig.builder().threadId(threadId).build();
        StateSnapshot stateSnapshot = this.compiledGraph.getState(runnableConfig);
        OverAllState state = stateSnapshot.state();
        state.withResume();

        Map<String, Object> objectMap = new HashMap<>();
        objectMap.put("feedback", feedBack);

        state.withHumanFeedback(new OverAllState.HumanFeedback(objectMap, ""));

        // Create a unicast sink to emit ServerSentEvents
        Sinks.Many<ServerSentEvent<String>> sink = Sinks.many().unicast().onBackpressureBuffer();
        GraphProcess graphProcess = new GraphProcess(this.compiledGraph);
        AsyncGenerator<NodeOutput> resultFuture = compiledGraph.streamFromInitialNode(state, runnableConfig);
        graphProcess.processStream(resultFuture, sink);

        return sink.asFlux()
                .doOnCancel(() -> logger.info("Client disconnected from stream"))
                .doOnError(e -> logger.error("Error occurred during streaming", e));    }
}

GraphProcess

ExecutorService executor：配置线程池，获取 stream 流

将结果写入到 sink 中

package com.spring.ai.tutorial.graph.stream.controller.GraphProcess;

import com.alibaba.cloud.ai.graph.CompiledGraph;
import com.alibaba.cloud.ai.graph.NodeOutput;
import com.alibaba.cloud.ai.graph.streaming.StreamingOutput;
import com.alibaba.fastjson.JSON;
import com.alibaba.fastjson.JSONObject;
import org.bsc.async.AsyncGenerator;
import org.slf4j.Logger;
import org.slf4j.LoggerFactory;
import org.springframework.http.codec.ServerSentEvent;
import reactor.core.publisher.Sinks;

import java.util.Map;
import java.util.concurrent.CompletionException;
import java.util.concurrent.ExecutorService;
import java.util.concurrent.Executors;

public class GraphProcess {

    private static final Logger logger = LoggerFactory.getLogger(GraphProcess.class);

    private final ExecutorService executor = Executors.newSingleThreadExecutor();

    private CompiledGraph compiledGraph;

    public GraphProcess(CompiledGraph compiledGraph) {
        this.compiledGraph = compiledGraph;
    }

    public void processStream(AsyncGenerator<NodeOutput> generator, Sinks.Many<ServerSentEvent<String>> sink) {
        executor.submit(() -> {
            generator.forEachAsync(output -> {
                try {
                    logger.info("output = {}", output);
                    String nodeName = output.node();
                    String content;
                    if (output instanceof StreamingOutput streamingOutput) {
                        content = JSON.toJSONString(Map.of(nodeName, streamingOutput.chunk()));
                    } else {
                        JSONObject nodeOutput = new JSONObject();
                        nodeOutput.put("data", output.state().data());
                        nodeOutput.put("node", nodeName);
                        content = JSON.toJSONString(nodeOutput);
                    }
                    sink.tryEmitNext(ServerSentEvent.builder(content).build());
                } catch (Exception e) {
                    throw new CompletionException(e);
                }
            }).thenAccept(v -> {
                // 正常完成
                sink.tryEmitComplete();
            }).exceptionally(e -> {
                sink.tryEmitError(e);
                return null;
            });
        });
    }
}

效果

调用 expand 接口，流式输出 && 断流得到最终结果

再调用 resume 接口，状态恢复续上流，接着走后续逻辑

学习交流圈

你好，我是影子，曾先后在🐻、新能源、老铁就职，现在是一名AI研发工程师，同时作为Spring AI Alibaba开源社区的Committer。目前新建了一个交流群，一个人走得快，一群人走得远，关注公众号后可获得个人微信，添加微信后备注“交流”入群。另外，本人长期维护一套飞书云文档笔记，涵盖后端、大数据系统化的面试资料，可私信免费获取