博主在使用Spring AI的ElasticsearchVectorStore实现ES向量存储的时候，想在metadata中存一个date字段，实体代码如下：

1
2
3
4
5
6
7
8
public class KnowledgeBase {
    private String id;
    private String title;
    private String content;
    private String url;
    private LocalDate date;
    private Map<String, Object> metadata;
}

众所周知，Jackson序列化Java 8时间（Local家族）需要导入jackson-datatype-jsr310模块，博主也是导入了，可是还是报错，尝试了各种方法，都无果后，博主跑去仔细研读了Spring AI源码，总算是发现了罪魁祸首：

1
2
3
4
5
6
7
8
9
protected ElasticsearchVectorStore(Builder builder) {
    super(builder);
    Assert.notNull(builder.restClient, "RestClient must not be null");
    this.initializeSchema = builder.initializeSchema;
    this.options = builder.options;
    this.filterExpressionConverter = builder.filterExpressionConverter;
    String version = Version.VERSION == null ? "Unknown" : Version.VERSION.toString();
    this.elasticsearchClient = (ElasticsearchClient)(new ElasticsearchClient(new RestClientTransport(builder.restClient, new JacksonJsonpMapper((new ObjectMapper()).configure(DeserializationFeature.FAIL_ON_UNKNOWN_PROPERTIES, false))))).withTransportOptions((t) -> t.addHeader("user-agent", "spring-ai elastic-java/" + version));
}

主要看ElasticsearchClient的创建这一部分，Spring AI直接new JacksonJsonpMapper()，导致博主无论怎么配置Bean都无效，那怎么办，Spring AI设计太不灵活，博主就只能”越狱“了

使用Java的”不讲武德“工具——反射！！！

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
return new ElasticsearchVectorStore(ElasticsearchVectorStore
        .builder(elasticsearchRestClient, embeddingModel)
        .initializeSchema(true)
        .options(vectorStoreOptions)
        .batchingStrategy(batchingStrategy)) {
    {
        try {
            // 使用反射注入替换 final 字段
            Field field = ElasticsearchVectorStore.class.getDeclaredField("elasticsearchClient");
            field.setAccessible(true);
            field.set(this, elasticsearchClient);
        } catch (Exception e) {
            throw new RuntimeException("Failed to override elasticsearchClient", e);
        }
    }
};

直接new一个匿名的ElasticsearchVectorStore类，然后在匿名类中声明一个静态代码块强制覆盖ElasticsearchClient，直接美美解决序列化异常问题啦~

就看Spring AI团段后续会不会改成传入ElasticsearchRestClient而不是RestClient，或者其他更为灵活的方法。

目前博主这种方式是最为有效的方式了！