1、http请求方式
如果elasticsearch服务设置账号密码，则在请求的header中添加 Basic Auth 认证
请求方式：Post
请求地址：/index_name/_search
请求body：json格式
{
"size": 10, //返回条数
"min_score": 0.8, // 设置最低相似分值
"_source": ["file_name", "length", "_es_doc_type"], // 只返回指定字段
"query": {
"script_score": {
"query": {
"match_all": {}
},
"script": {
// _img_vector 为设置的向量索引字段
"source": "cosineSimilarity(params.query_vector, '_img_vector') + 0.0",
"params": {
"query_vector": [-1,1,-0.07559559,-0.007800484,0.11229578,0.064164124,....]
}
}
}
}
}
主要参数说明：
- "from": 0, // 起始位置，0表示第一页
- "size": 10, // 每页返回的记录数
- "min_score": 0.5, //最低相似度，最高1
- "_source": ["image_id", "image_name", "image_vector"], // 返回指定字段
返回结果如下：
{
"took": 3,
"timed_out": false,
"_shards": {
"total": 1,
"successful": 1,
"skipped": 0,
"failed": 0
},
"hits": {
"total": {
"value": 1,
"relation": "eq"
},
"max_score": 0.9014968,
"hits": [
{
"_index": "vedms",
"_type": "_doc",
"_id": "04a40e806be82e87f3c3a2f3877225bd.jpg",
"_score": 0.9014968,
"_source": {
"file_name": "04a40e806be82e87f3c3a2f3877225bd.jpg",
"_es_doc_type": "IMAGE",
"length": 89690
}
}
]
}
}
需要确保传入的query_vector 长度一致性，前面的章节中以设定1024长度。
否则会出现如下错误：
"reason": {
"type": "script_exception",
"reason": "runtime error",
"script_stack": [
"org.elasticsearch.xpack.vectors.query.ScoreScriptUtils$DenseVectorFunction.<init>(ScoreScriptUtils.java:74)",
"org.elasticsearch.xpack.vectors.query.ScoreScriptUtils$CosineSimilarity.<init>(ScoreScriptUtils.java:172)",
"cosineSimilarity(params.query_vector, '_img_vector') + 0.0",
"
],
"script": "cosineSimilarity(params.query_vector, '_img_vector') + 0.0",
"lang": "painless",
"position": {
"offset": 38,
"start": 0,
"end": 58
},
"caused_by": {
"type": "illegal_argument_exception",
"reason": "The query vector has a different number of dimensions [1023] than the document vectors [1024]."
}
}
2、Java调用脚本
SearchRequest 不允许在script设置 _source 属性内容，所以干脆将from、size、score一并拿出，只保留vector数据
_img_vector为前面定义的向量索引字段
public List<Map<String, Object>> search(EsVectorSearchReq req) {
float[] vector = getImgFeature(req);
if (null == vector || vector.length == 0) {
return Collections.emptyList();
}
String queryJson = String.format(VECTOR_FORMAT, vectorToJson(vector));
log.debug("向量检索入参条件={}", queryJson);
Reader input = new StringReader(queryJson);
// 使用查询 DSL 进行搜索
SearchRequest searchRequest = new SearchRequest.Builder()
.index(req.getIndexLib())
.from(req.getFrom())
.size(req.getSize())
.minScore(req.getScore())
.source(SourceConfig.of(src -> src
.filter(SourceFilter.of(i -> i.includes(req.getColumns())))))
.withJson(input)
.build();
// 执行查询
List<Map<String, Object>> result = new ArrayList<>();
try {
SearchResponse<Map> searchResponse = esClient.search(searchRequest, Map.class);
// 输出结果
for (Hit<Map> hit : searchResponse.hits().hits()) {
result.add(hit.source());
}
log.info("成功查询{}条", result.size());
} catch (IOException e) {
e.printStackTrace();
}
return result;
}
private String vectorToJson(float[] vector) {
StringBuilder sb = new StringBuilder("[");
for (int i = 0; i < vector.length; i++) {
sb.append(vector[i]);
if (i < vector.length - 1) {
sb.append(",");
}
}
sb.append("]");
return sb.toString();
}
private static final String VECTOR_FORMAT = "{\n" +
" \"query\": {\n" +
" \"script_score\": {\n" +
" \"query\": {\n" +
" \"match_all\": {}\n" +
" },\n" +
" \"script\": {\n" +
" \"source\": \"cosineSimilarity(params.query_vector, 'img_vector') + 0.0\",\n" +
" \"params\": {\n" +
" \"query_vector\": %s\n" +
" }\n" +
" }\n" +
" }\n" +
" }\n" +
"}";
传入参数格式如下：
{
"query": {
"script_score": {
"query": {
"match_all": {}
},
"script": {
"source": "cosineSimilarity(params.query_vector, '_img_vector') + 0.0",
"params": {
"query_vector": [-0.033....]
}
}
}
}
}
返回结果如下：
{
"_shards": {
"failed": 0.0,
"skipped": 0.0,
"successful": 1.0,
"total": 1.0
},
"hits": {
"hits": [
{
"_id": "04a40e806be82e87f3c3a2f3877225bd.jpg",
"_index": "vedms",
"_score": 1.0,
"_source": "{file_name=04a40e806be82e87f3c3a2f3877225bd.jpg}",
"_type": "_doc"
}
],
"max_score": 1.0,
"total": {
"relation": "eq",
"value": 1
}
},
"timed_out": false,
"took": 46
}