Spring AI 对话记忆 + 结构化输出

剪影AS

250人浏览 · 2026-06-30 22:48:15

剪影AS · 2026-06-30 22:48:15 发布

对话记忆

大型语言模型 (LLM) 是无状态的，这意味着它们不会保留先前交互的信息。

<span style="color:#000000"><span style="background-color:#fefef2"><code class="language-java"><span style="color:#015692">@Test</span>
    <span style="color:#015692">public</span> <span style="color:#015692">void</span> <span style="color:#b75501">testChatOptions</span>() {
        <span style="color:#b75501">String</span> <span style="color:#54790d">content</span> <span style="color:#ab5656">=</span> chatClient.prompt()
                .user(<span style="color:#54790d">"我叫小兔子 "</span>)
                .call()
                .content();
        System.out.println(content);
        System.out.println(<span style="color:#54790d">"--------------------------------------------------------------------------"</span>);

       content = chatClient.prompt()
                .user(<span style="color:#54790d">"我叫什么 ？"</span>)
                .call()
                .content();
        System.out.println(content);
    }
</code></span></span>

那我们平常跟一些大模型聊天是怎么记住我们对话的呢？实际上，每次对话都需要将之前的对话消息内置发送给大模型，这种方式称为多轮对话。

SpringAi提供了一个ChatMemory的组件用于存储聊天记录，允许您使用 LLM 跨多个交互存储和检索信息。并且可以为不同用户的多个交互之间维护上下文或状态。
可以在每次对话的时候把当前聊天信息和模型的响应存储到ChatMemory，然后下一次对话把聊天记录取出来再发给大模型。

<span style="color:#000000"><span style="background-color:#fefef2"><code class="language-java">`

<span style="color:#656e77">//输出 名字叫徐庶</span>
</code></span></span>

但是这样做未免太麻烦！能不能简化？思考一下！

用我们之前的Advisor对话拦截是不是就可以不用每次手动去维护了。并且SpringAi早已体贴的为我提供了ChatMemoryAutoConfiguration自动配置类

<span style="color:#000000"><span style="background-color:#fefef2"><code class="language-xml"><<span style="color:#b75501">dependency</span>>
  <<span style="color:#b75501">groupId</span>>org.springframework.ai</<span style="color:#b75501">groupId</span>>
  <<span style="color:#b75501">artifactId</span>>spring-ai-autoconfigure-model-chat-memory</<span style="color:#b75501">artifactId</span>>
</<span style="color:#b75501">dependency</span>>
</code></span></span>

<span style="color:#000000"><span style="background-color:#fefef2"><code class="language-java"><span style="color:#015692">@AutoConfiguration</span>
<span style="color:#015692">@ConditionalOnClass({ ChatMemory.class, ChatMemoryRepository.class })</span>
<span style="color:#015692">public</span> <span style="color:#015692">class</span> <span style="color:#b75501">ChatMemoryAutoConfiguration</span> {

	<span style="color:#015692">@Bean</span>
	<span style="color:#015692">@ConditionalOnMissingBean</span>
	ChatMemoryRepository <span style="color:#b75501">chatMemoryRepository</span>() {
		<span style="color:#015692">return</span> <span style="color:#015692">new</span> <span style="color:#b75501">InMemoryChatMemoryRepository</span>();
	}

	<span style="color:#015692">@Bean</span>
	<span style="color:#015692">@ConditionalOnMissingBean</span>
	ChatMemory <span style="color:#b75501">chatMemory</span>(ChatMemoryRepository chatMemoryRepository) {
		<span style="color:#015692">return</span> MessageWindowChatMemory.builder().chatMemoryRepository(chatMemoryRepository).build();
	}

}
</code></span></span>

所以我们可以这样用：

PromptChatMemoryAdvisor

SpringAi提供了 PromptChatMemoryAdvisor 专门用于对话记忆的拦截

<span style="color:#000000"><span style="background-color:#fefef2"><code class="language-java"><span style="color:#015692">@SpringBootTest</span>
<span style="color:#015692">public</span> <span style="color:#015692">class</span> <span style="color:#b75501">ChatMemoryTest</span> {
    ChatClient chatClient;
    <span style="color:#015692">@BeforeEach</span>
    <span style="color:#015692">public</span>  <span style="color:#015692">void</span> <span style="color:#b75501">init</span>(<span style="color:#015692">@Autowired</span>
                      DeepSeekChatModel chatModel,
                      <span style="color:#015692">@Autowired</span>
                      ChatMemory chatMemory) {
        chatClient = ChatClient
                .builder(chatModel)
                .defaultAdvisors(
                    <span style="color:#656e77">// PromptChatMemoryAdvisor拦截器 就会自动将我们与大模型的历史对话记录下来</span>
                        PromptChatMemoryAdvisor.builder(chatMemory).build()
                )
                .build();
    }
    <span style="color:#015692">@Test</span>
    <span style="color:#015692">public</span> <span style="color:#015692">void</span> <span style="color:#b75501">testChatOptions</span>() {
        <span style="color:#b75501">String</span> <span style="color:#54790d">content</span> <span style="color:#ab5656">=</span> chatClient.prompt()
                .user(<span style="color:#54790d">"我叫徐庶 ？"</span>)
        <span style="color:#656e77">// </span>
                .advisors(<span style="color:#015692">new</span> <span style="color:#b75501">ReReadingAdvisor</span>())
                .call()
                .content();
        System.out.println(content);
        System.out.println(<span style="color:#54790d">"--------------------------------------------------------------------------"</span>);

        content = chatClient.prompt()
                .user(<span style="color:#54790d">"我叫什么 ？"</span>)
                .advisors(<span style="color:#015692">new</span> <span style="color:#b75501">ReReadingAdvisor</span>())
                .call()
                .content();
        System.out.println(content);
    }
}
</code></span></span>

配置聊天记录最大存储数量

你要知道，我们把聊天记录发给大模型，都是算token计数的。

大模型的token是有上限了，如果你发送过多聊天记录，可能就会导致token过长。

如下是大模型存储的 token 历史条数上限。

并且更多的token也意味更多的费用，更久的解析时间. 所以不建议太长
（DEFAULT_MAX_MESSAGES默认20即10次对话）
一旦超出DEFAULT_MAX_MESSAGES只会存最后面N条（可以理解为先进先出），参考MessageWindowChatMemory源码

<span style="color:#000000"><span style="background-color:#fefef2"><code class="language-java"><span style="color:#015692">@Bean</span>
   ChatMemory <span style="color:#b75501">chatMemory</span>(<span style="color:#015692">@Autowired</span> ChatMemoryRepository chatMemoryRepository) {
<span style="color:#656e77">// MessageWindowChatMemory 创建一个历史对话存储的配置，</span>
        <span style="color:#015692">return</span> MessageWindowChatMemory
                .builder()
                .maxMessages(<span style="color:#b75501">10</span>)  <span style="color:#656e77">// 设置最大存储 10 条</span>
                .chatMemoryRepository(chatMemoryRepository).build();
    }
</code></span></span>

配置多用户隔离记忆

如果有多个用户在进行对话，肯定不能将对话记录混在一起，不同的用户的对话记忆需要隔离

<span style="color:#000000"><span style="background-color:#fefef2"><code class="language-java"><span style="color:#015692">@Test</span>
    <span style="color:#015692">public</span> <span style="color:#015692">void</span> <span style="color:#b75501">testChatOptions</span>() {
        <span style="color:#b75501">String</span> <span style="color:#54790d">content</span> <span style="color:#ab5656">=</span> chatClient.prompt()
                .user(<span style="color:#54790d">"我叫徐庶 ？"</span>)
        <span style="color:#656e77">// 注意:这里要先构建一个 ChatMemory的 Bean，和上面类似，这里我们设置历史对话的用户ID</span>
                .advisors(advisorSpec -> advisorSpec.param(ChatMemory.CONVERSATION_ID,<span style="color:#54790d">"1"</span>))
                .call()
                .content();
        System.out.println(content);
        System.out.println(<span style="color:#54790d">"--------------------------------------------------------------------------"</span>);

        content = chatClient.prompt()
                .user(<span style="color:#54790d">"我叫什么 ？"</span>)
                .advisors(advisorSpec -> advisorSpec.param(ChatMemory.CONVERSATION_ID,<span style="color:#54790d">"1"</span>))
                .call()
                .content();
        System.out.println(content);


        System.out.println(<span style="color:#54790d">"--------------------------------------------------------------------------"</span>);

        content = chatClient.prompt()
                .user(<span style="color:#54790d">"我叫什么 ？"</span>)
                .advisors(advisorSpec -> advisorSpec.param(ChatMemory.CONVERSATION_ID,<span style="color:#54790d">"2"</span>))
                .call()
                .content();
        System.out.println(content);
    }
</code></span></span>

会发现，不同的CONVERSATION_ID，会有不同的记忆

原理源码$

主要有前置存储
MessageWindowChatMemory
具体存储实现
ChatMemoryRepository

数据库存储对话记忆

默认情况，对话内容会存在jvm内存会导致：

一直存最终会撑爆JVM导致OOM。
重启就丢了，如果已想存储到第三方存储进行持久化

springAi内置提供了以下几种方式（例如 Cassandra、JDBC 或 Neo4j），这里演示下JDBC方式

添加依赖

<span style="color:#000000"><span style="background-color:#fefef2"><code class="language-xml">        <<span style="color:#b75501">dependency</span>>
            <<span style="color:#b75501">groupId</span>>org.springframework.ai</<span style="color:#b75501">groupId</span>>
            <<span style="color:#b75501">artifactId</span>>spring-ai-starter-model-chat-memory-repository-jdbc</<span style="color:#b75501">artifactId</span>>
        </<span style="color:#b75501">dependency</span>>

        <span style="color:#656e77"><!--jdbc--></span>
        <<span style="color:#b75501">dependency</span>>
            <<span style="color:#b75501">groupId</span>>org.springframework.boot</<span style="color:#b75501">groupId</span>>
            <<span style="color:#b75501">artifactId</span>>spring-boot-starter-jdbc</<span style="color:#b75501">artifactId</span>>
        </<span style="color:#b75501">dependency</span>>


        <span style="color:#656e77"><!--mysql驱动--></span>
        <<span style="color:#b75501">dependency</span>>
            <<span style="color:#b75501">groupId</span>>com.mysql</<span style="color:#b75501">groupId</span>>
            <<span style="color:#b75501">artifactId</span>>mysql-connector-j</<span style="color:#b75501">artifactId</span>>
            <<span style="color:#b75501">scope</span>>runtime</<span style="color:#b75501">scope</span>>
        </<span style="color:#b75501">dependency</span>>
</code></span></span>

添加配置(目前我们的需要创建一个schema-mysql.sql 文件，就是一个 SQL 脚本，在后面有配置)SPRING_AI_CHAT_MEMORY 表存储用户的历史对话，数据库，我们自行定义将该数据表存储到那个数据库中即可。

<span style="color:#000000"><span style="background-color:#fefef2"><code class="language-properties"><span style="color:#015692">spring.ai.chat.memory.repository.jdbc.initialize-schema</span>=<span style="color:#54790d">always</span>
<span style="color:#015692">spring.ai.chat.memory.repository.jdbc.schema</span>=<span style="color:#54790d">classpath:/schema-mysql.sql</span>
</code></span></span>

如下是 MySQL 的配置：

<span style="color:#000000"><span style="background-color:#fefef2"><code class="language-yaml"><span style="color:#015692">spring:</span>
  <span style="color:#015692">datasource:</span>
    <span style="color:#015692">username:</span> <span style="color:#54790d">root</span>
    <span style="color:#015692">password:</span> <span style="color:#b75501">123456</span>
    <span style="color:#015692">url:</span> <span style="color:#54790d">jdbc:mysql://localhost:3306/springai?characterEncoding=utf8&useSSL=false&serverTimezone=UTC&</span>
    <span style="color:#015692">driver-class-name:</span> <span style="color:#54790d">com.mysql.cj.jdbc.Driver</span>
</code></span></span>

配置类

<span style="color:#000000"><span style="background-color:#fefef2"><code class="language-java"><span style="color:#015692">@Configuration</span>
<span style="color:#015692">public</span> <span style="color:#015692">class</span> <span style="color:#b75501">ChatMemoryConfig</span> {


    <span style="color:#015692">@Bean</span>  <span style="color:#656e77">// JdbcChatMemoryRepository 是已经被封装好自动装配好了，就可以使用</span>
    ChatMemory <span style="color:#b75501">chatMemory</span>(<span style="color:#015692">@Autowired</span> JdbcChatMemoryRepository chatMemoryRepository) {
        <span style="color:#015692">return</span> MessageWindowChatMemory
        .builder()
        .maxMessages(<span style="color:#b75501">1</span>)  <span style="color:#656e77">// 设置存储为上面我们传的变量的 jdbc 的存储方式</span>
        .chatMemoryRepository(chatMemoryRepository).build();
    }

}
</code></span></span>

resources/schema-mysql.sql（目前1.0.0版本需要自己定义，没有提供脚本），创建这个SPRING_AI_CHAT_MEMORY 数据表，来存储用户的历史对话

<span style="color:#000000"><span style="background-color:#fefef2"><code class="language-java">CREATE TABLE IF NOT EXISTS <span style="color:#b75501">SPRING_AI_CHAT_MEMORY</span> (
    `conversation_id` VARCHAR(<span style="color:#b75501">36</span>) NOT NULL,
    `content` TEXT NOT NULL,
    `type` VARCHAR(<span style="color:#b75501">10</span>) NOT NULL,
    `timestamp` TIMESTAMP NOT NULL,

    INDEX `SPRING_AI_CHAT_MEMORY_CONVERSATION_ID_TIMESTAMP_IDX` (`conversation_id`, `timestamp`)
    );
</code></span></span>

测试

<span style="color:#000000"><span style="background-color:#fefef2"><code class="language-java"><span style="color:#015692">@SpringBootTest</span>
<span style="color:#015692">public</span> <span style="color:#015692">class</span> <span style="color:#b75501">ChatMemoryTest</span> {


    ChatClient chatClient;
    <span style="color:#015692">@BeforeEach</span>
    <span style="color:#015692">public</span>  <span style="color:#015692">void</span> <span style="color:#b75501">init</span>(<span style="color:#015692">@Autowired</span>
                      DeepSeekChatModel chatModel,
                      <span style="color:#015692">@Autowired</span>
                      ChatMemory chatMemory) {
        chatClient = ChatClient
                .builder(chatModel)
                .defaultAdvisors(
                        PromptChatMemoryAdvisor.builder(chatMemory).build()
                )
                .build();
    }
    <span style="color:#015692">@Test</span>
    <span style="color:#015692">public</span> <span style="color:#015692">void</span> <span style="color:#b75501">testChatOptions</span>() {
        <span style="color:#b75501">String</span> <span style="color:#54790d">content</span> <span style="color:#ab5656">=</span> chatClient.prompt()
                .user(<span style="color:#54790d">"你好，我叫徐庶！"</span>)
                .advisors(<span style="color:#015692">new</span> <span style="color:#b75501">ReReadingAdvisor</span>())
                .advisors(advisorSpec -> advisorSpec.param(ChatMemory.CONVERSATION_ID,<span style="color:#54790d">"1"</span>))
                .call()
                .content();
        System.out.println(content);
        System.out.println(<span style="color:#54790d">"--------------------------------------</span></code></span></span>