open search 近似knn搜索支持

face-search-client/src/main/java/com/visual/face/search/handle/CollectHandler.java

@@ -48,7 +48,8 @@ public class CollectHandler extends BaseHandler<CollectHandler>{
                 .setFaceColumns(collect.getFaceColumns())
                 .setShardsNum(collect.getShardsNum())
                 .setStorageFaceInfo(collect.getStorageFaceInfo())
-                .setStorageEngine(collect.getStorageEngine());
+                .setStorageEngine(collect.getStorageEngine())
+                .setApproximateKnn(collect.isApproximateKnn());
         return HttpClient.post(Api.getUrl(this.serverHost, Api.collect_create), collectReq);
     }
 

face-search-client/src/main/java/com/visual/face/search/handle/SearchHandler.java

@@ -45,7 +45,8 @@ public class SearchHandler extends BaseHandler<SearchHandler>{
                 .setMaxFaceNum(search.getMaxFaceNum())
                 .setLimit(search.getLimit())
                 .setConfidenceThreshold(search.getConfidenceThreshold())
-                .setFaceScoreThreshold(search.getFaceScoreThreshold());
+                .setFaceScoreThreshold(search.getFaceScoreThreshold())
+                .setApproximateKnn(search.isApproximateKnn());
         return HttpClient.post(Api.getUrl(this.serverHost, Api.visual_search), searchReq, new TypeReference<Response<List<SearchRep>>>(){});
     }
 }

face-search-client/src/main/java/com/visual/face/search/model/Collect.java

@@ -23,6 +23,9 @@ public class Collect<ExtendsVo extends Collect<ExtendsVo>> implements Serializab
     /**保留图片及人脸信息的存储组件**/
     private StorageEngine storageEngine;
 
+    /**是否启用近似knn搜索**/
+    private boolean approximateKnn;
+
     /**
      * 构建集合对象
      * @return
@@ -51,6 +54,15 @@ public class Collect<ExtendsVo extends Collect<ExtendsVo>> implements Serializab
         return (ExtendsVo) this;
     }
 
+    public boolean isApproximateKnn(){
+        return this.approximateKnn;
+    }
+
+    public ExtendsVo setApproximateKnn(boolean approximateKnn){
+        this.approximateKnn = approximateKnn;
+        return (ExtendsVo) this;
+    }
+
     public Integer getShardsNum() {
         return shardsNum;
     }

face-search-client/src/main/java/com/visual/face/search/model/Search.java

@@ -15,6 +15,8 @@ public class Search<ExtendsVo extends Search<ExtendsVo>> implements Serializable
     /**对输入图像中多少个人脸进行检索比对：默认5**/
     private Integer maxFaceNum=5;
 
+    /**是否启用近似knn搜索**/
+    private boolean approximateKnn = false;
     /**
      * 构建检索对象
      * @param imageBase64       待检索的图片
@@ -33,6 +35,15 @@ public class Search<ExtendsVo extends Search<ExtendsVo>> implements Serializable
         return (ExtendsVo) this;
     }
 
+    public boolean isApproximateKnn() {
+        return approximateKnn;
+    }
+
+    public ExtendsVo setApproximateKnn(boolean approximateKnn) {
+        this.approximateKnn = approximateKnn;
+        return (ExtendsVo)this;
+    }
+
     public Float getFaceScoreThreshold() {
         return faceScoreThreshold;
     }

face-search-engine/src/main/java/com/visual/face/search/engine/api/SearchEngine.java

@@ -14,7 +14,11 @@ public interface SearchEngine {
 
     public boolean dropCollection(String collectionName);
 
-    public boolean createCollection(String collectionName, MapParam param);
+    default public boolean createCollection(String collectionName, MapParam param){
+        return createCollection(collectionName,param,false);
+    };
+
+    public boolean createCollection(String collectionName, MapParam param,boolean approximateKnn);
 
     public boolean insertVector(String collectionName, String sampleId, String faceId, float[] vectors);
 
@@ -22,7 +26,11 @@ public interface SearchEngine {
 
     public boolean deleteVectorByKey(String collectionName, List<String> faceIds);
 
-    public SearchResponse search(String collectionName, float[][] features, String algorithm, int topK);
+    default public SearchResponse search(String collectionName, float[][] features, String algorithm, int topK){
+        return search(collectionName,features,algorithm,topK,false);
+    };
+
+    public SearchResponse search(String collectionName, float[][] features, String algorithm, int topK, boolean approximateKnn);
 
     public float searchMinScoreBySampleId(String collectionName, String sampleId,float[] feature, String algorithm);
 

face-search-engine/src/main/java/com/visual/face/search/engine/conf/Constant.java

@@ -4,6 +4,7 @@ public class Constant {
 
     public final static String IndexShardsNum = "shardsNum";
     public final static String IndexReplicasNum = "replicasNum";
+    public final static String IndexAlgoParamEfSearch = "algoParamEfSearch";
 
     public final static String ColumnNameFaceId = "face_id";
     public final static String ColumnNameSampleId = "sample_id";

face-search-engine/src/main/java/com/visual/face/search/engine/impl/OpenSearchEngine.java

@@ -3,6 +3,7 @@ package com.visual.face.search.engine.impl;
 import com.visual.face.search.engine.api.SearchEngine;
 import com.visual.face.search.engine.conf.Constant;
 import com.visual.face.search.engine.exps.SearchEngineException;
+import com.visual.face.search.engine.impl.query.ApproximateKnnQueryBuilder;
 import com.visual.face.search.engine.model.*;
 import org.apache.commons.collections4.MapUtils;
 import org.opensearch.action.DocWriteResponse;
@@ -71,17 +72,32 @@ public class OpenSearchEngine implements SearchEngine {
     }
 
     @Override
-    public boolean createCollection(String collectionName, MapParam param) {
+    public boolean createCollection(String collectionName, MapParam param,boolean approximateKnn) {
         try {
             //构建请求
             CreateIndexRequest createIndexRequest = new CreateIndexRequest(collectionName);
-            createIndexRequest.settings(Settings.builder()
+            Settings.Builder builder = Settings.builder()
-                .put("index.number_of_shards", param.getIndexShardsNum())
+                    .put("index.number_of_shards", param.getIndexShardsNum())
-                .put("index.number_of_replicas", param.getIndexReplicasNum())
+                    .put("index.number_of_replicas", param.getIndexReplicasNum());
-            );
+            if(approximateKnn){
+                //启用open search近似knn搜索支持
+                builder.put("index.knn",true);
+                builder.put("index.knn.algo_param.ef_search",param.getIndexAlgoParamEfSearch());
+            }
+
+            createIndexRequest.settings(builder);
             HashMap<String, Object> properties = new HashMap<>();
             properties.put(Constant.ColumnNameSampleId, Map.of("type", "keyword"));
-            properties.put(Constant.ColumnNameFaceVector, Map.of("type", "knn_vector", "dimension", "512"));
+            if(approximateKnn){
+                //启用open search近似knn搜索支持
+                properties.put(Constant.ColumnNameFaceVector, Map.of("type", "knn_vector", "dimension", "512",
+                        "method",Map.of("engine","nmslib",
+                                "space_type","cosinesimil",
+                                "name","hnsw",
+                                "parameters",Map.of())));
+            }else {
+                properties.put(Constant.ColumnNameFaceVector, Map.of("type", "knn_vector", "dimension", "512"));
+            }
             createIndexRequest.mapping(Map.of("properties", properties));
             //创建集合
             CreateIndexResponse createIndexResponse = client.indices().create(createIndexRequest, RequestOptions.DEFAULT);
@@ -135,23 +151,37 @@ public class OpenSearchEngine implements SearchEngine {
     }
 
     @Override
-    public SearchResponse search(String collectionName, float[][] features, String algorithm, int topK) {
+    public SearchResponse search(String collectionName, float[][] features, String algorithm, int topK,boolean approximateKnn) {
         try {
             //构建搜索请求
             MultiSearchRequest multiSearchRequest = new MultiSearchRequest();
             for(float[] feature : features){
-                QueryBuilder queryBuilder = new MatchAllQueryBuilder();
+                if(approximateKnn){
-                Map<String, Object> params = new HashMap<>();
+                    //近似knn搜索
-                params.put("field", Constant.ColumnNameFaceVector);
+                    Map<String, Object> params = new HashMap<>();
-                params.put("space_type", algorithm);
+                    params.put("vector",feature);
-                params.put("query_value", feature);
+                    params.put("k",topK);
-                Script script = new Script(Script.DEFAULT_SCRIPT_TYPE, "knn", "knn_score", params);
+                    ApproximateKnnQueryBuilder approximateKnnQueryBuilder = new ApproximateKnnQueryBuilder(params);
-                ScriptScoreQueryBuilder scriptScoreQueryBuilder = new ScriptScoreQueryBuilder(queryBuilder, script);
+                    SearchSourceBuilder searchSourceBuilder =new SearchSourceBuilder()
-                SearchSourceBuilder searchSourceBuilder = new SearchSourceBuilder()
+                            .query(approximateKnnQueryBuilder).size(topK)
-                    .query(scriptScoreQueryBuilder).size(topK)
+                            .fetchSource(null,Constant.ColumnNameFaceVector);
-                    .fetchSource(null, Constant.ColumnNameFaceVector); //是否需要向量字段
+                    SearchRequest searchRequest = new SearchRequest(collectionName).source(searchSourceBuilder);
-                SearchRequest searchRequest = new SearchRequest(collectionName).source(searchSourceBuilder);
+                    multiSearchRequest.add(searchRequest);
-                multiSearchRequest.add(searchRequest);
+                }else {
+                    //常规搜索
+                    QueryBuilder queryBuilder = new MatchAllQueryBuilder();
+                    Map<String, Object> params = new HashMap<>();
+                    params.put("field", Constant.ColumnNameFaceVector);
+                    params.put("space_type", algorithm);
+                    params.put("query_value", feature);
+                    Script script = new Script(Script.DEFAULT_SCRIPT_TYPE, "knn", "knn_score", params);
+                    ScriptScoreQueryBuilder scriptScoreQueryBuilder = new ScriptScoreQueryBuilder(queryBuilder, script);
+                    SearchSourceBuilder searchSourceBuilder = new SearchSourceBuilder()
+                            .query(scriptScoreQueryBuilder).size(topK)
+                            .fetchSource(null, Constant.ColumnNameFaceVector); //是否需要向量字段
+                    SearchRequest searchRequest = new SearchRequest(collectionName).source(searchSourceBuilder);
+                    multiSearchRequest.add(searchRequest);
+                }
             }
             //查询索引
             MultiSearchResponse response = this.client.msearch(multiSearchRequest, RequestOptions.DEFAULT);
@@ -167,7 +197,7 @@ public class OpenSearchEngine implements SearchEngine {
                 if(searchHits != null){
                     for(SearchHit searchHit : searchHits){
                         String faceId = searchHit.getId();
-                        float score = searchHit.getScore()-1;
+                        float score = approximateKnn? searchHit.getScore() : (searchHit.getScore()-1);
                         Map<String, Object> sourceMap = searchHit.getSourceAsMap();
                         String sampleId = MapUtils.getString(sourceMap, Constant.ColumnNameSampleId);
                         Object faceVector = MapUtils.getObject(sourceMap, Constant.ColumnNameFaceVector);

face-search-engine/src/main/java/com/visual/face/search/engine/impl/query/ApproximateKnnQueryBuilder.java

@@ -0,0 +1,58 @@
+package com.visual.face.search.engine.impl.query;
+
+import com.visual.face.search.engine.conf.Constant;
+import org.apache.lucene.search.Query;
+import org.opensearch.common.io.stream.StreamOutput;
+import org.opensearch.core.xcontent.XContentBuilder;
+import org.opensearch.index.query.AbstractQueryBuilder;
+import org.opensearch.index.query.QueryShardContext;
+
+import java.io.IOException;
+import java.util.Map;
+import java.util.Objects;
+
+/**
+ * 近似KNN搜索
+ * @Author Foy Lian
+ * @Date 2023/9/13 14:26
+ **/
+
+public class ApproximateKnnQueryBuilder extends AbstractQueryBuilder<ApproximateKnnQueryBuilder> {
+
+    private Map<String, Object> mParams;
+
+    public ApproximateKnnQueryBuilder(Map<String, Object> params){
+        this.mParams = params;
+    }
+
+    @Override
+    protected void doWriteTo(StreamOutput streamOutput) throws IOException {
+    }
+
+    @Override
+    protected void doXContent(XContentBuilder xContentBuilder, Params params) throws IOException {
+        xContentBuilder.startObject("knn");
+        xContentBuilder.field(Constant.ColumnNameFaceVector,mParams);
+        xContentBuilder.endObject();
+    }
+
+    @Override
+    protected Query doToQuery(QueryShardContext queryShardContext) throws IOException {
+        return null;
+    }
+
+    @Override
+    protected boolean doEquals(ApproximateKnnQueryBuilder approximateKnnQueryBuilder) {
+        return Objects.equals(this.mParams, approximateKnnQueryBuilder.mParams);
+    }
+
+    @Override
+    protected int doHashCode() {
+        return Objects.hash(new Object[]{this.mParams});
+    }
+
+    @Override
+    public String getWriteableName() {
+        return "knn";
+    }
+}

face-search-engine/src/main/java/com/visual/face/search/engine/model/MapParam.java

@@ -71,4 +71,9 @@ public class MapParam extends ConcurrentHashMap<String, Object> {
         return shardsNum;
     }
     
+    public Integer getIndexAlgoParamEfSearch(){
+        Integer algoParamEfSearch = this.getInteger(Constant.IndexAlgoParamEfSearch, 512);
+        algoParamEfSearch = (null == algoParamEfSearch || algoParamEfSearch <= 0) ? 512  : algoParamEfSearch;
+        return algoParamEfSearch;
+    }
 }

face-search-server/src/main/java/com/visual/face/search/server/domain/base/CollectVo.java

@@ -46,6 +46,11 @@ public class CollectVo<ExtendsVo extends CollectVo<ExtendsVo>> extends BaseVo {
     @ApiModelProperty(value="保留图片及人脸信息的存储组件", position = 9,required = false)
     private StorageEngine storageEngine;
 
+    /**是否启用近似knn搜索**/
+    @ApiModelProperty(value="是否启用近似knn搜索", position = 10,required = false)
+    private boolean approximateKnn;
+
+
     /**
      * 构建集合对象
      * @param namespace         命名空间
@@ -56,6 +61,15 @@ public class CollectVo<ExtendsVo extends CollectVo<ExtendsVo>> extends BaseVo {
         return new CollectVo().setNamespace(namespace).setCollectionName(collectionName);
     }
 
+    public boolean isApproximateKnn() {
+        return approximateKnn;
+    }
+
+    public ExtendsVo setApproximateKnn(boolean approximateKnn){
+        this.approximateKnn = approximateKnn;
+        return (ExtendsVo) this;
+    }
+
     public String getNamespace() {
         return namespace;
     }
@@ -144,4 +158,5 @@ public class CollectVo<ExtendsVo extends CollectVo<ExtendsVo>> extends BaseVo {
         }
         return (ExtendsVo) this;
     }
+
 }

face-search-server/src/main/java/com/visual/face/search/server/domain/request/FaceSearchReqVo.java

@@ -42,6 +42,12 @@ public class FaceSearchReqVo extends BaseVo {
     @ApiModelProperty(value="对输入图像中多少个人脸进行检索比对：默认5", position = 7, required = false)
     private Integer maxFaceNum;
 
+    /**是否使用近似knn搜索**/
+    @ApiModelProperty(value="是否使用近似knn搜索", position = 8, required = false)
+    private boolean approximateKnn;
+
+
+
     /**
      * 构建检索对象
      * @param namespace         命名空间
@@ -130,4 +136,13 @@ public class FaceSearchReqVo extends BaseVo {
         this.maxFaceNum = maxFaceNum;
         return this;
     }
+    
+    public boolean isApproximateKnn() {
+        return approximateKnn;
+    }
+
+    public FaceSearchReqVo setApproximateKnn(boolean approximateKnn) {
+        this.approximateKnn = approximateKnn;
+        return this;
+    }
 }

face-search-server/src/main/java/com/visual/face/search/server/service/impl/CollectServiceImpl.java

@@ -90,7 +90,7 @@ public class CollectServiceImpl extends BaseService implements CollectService {
             MapParam param = MapParam.build()
                     .put(Constant.IndexShardsNum, collect.getShardsNum())
                     .put(Constant.IndexReplicasNum, collect.getReplicasNum());
-            boolean createVectorFlag = searchEngine.createCollection(vectorTableName, param);
+            boolean createVectorFlag = searchEngine.createCollection(vectorTableName, param,collect.isApproximateKnn());
             if(!createVectorFlag){
                 throw new RuntimeException("create vector table error");
             }

face-search-server/src/main/java/com/visual/face/search/server/service/impl/FaceSearchServiceImpl.java

@@ -83,7 +83,7 @@ public class FaceSearchServiceImpl extends BaseService implements FaceSearchServ
         }
         //特征搜索
         int topK = (null == search.getLimit() || search.getLimit()  <= 0) ? 5 : search.getLimit();
-        SearchResponse searchResponse =searchEngine.search(collection.getVectorTable(), vectors, search.getAlgorithm().algorithm(), topK);
+        SearchResponse searchResponse =searchEngine.search(collection.getVectorTable(), vectors, search.getAlgorithm().algorithm(), topK,search.isApproximateKnn());
         if(!searchResponse.getStatus().ok()){
             throw new RuntimeException(searchResponse.getStatus().getReason());
         }

face-search-test/src/main/java/com/visual/face/search/valid/exps/FaceSearchExample.java

@@ -26,6 +26,12 @@ public class FaceSearchExample {
     public static String collectionName = "collect_20211201_v11";
     public static FaceSearch faceSearch = FaceSearch.build(serverHost, namespace, collectionName);
 
+    //是否启用近似knn,建议底库集比较大时启用.
+    public static boolean approximateKnn = false;
+
+    //底库集比较大时，建议调大，如:32个分片
+    public static int shardsNum = 4;
+
     /**集合创建*/
     public static void collect(){
         //样本属性字段
@@ -44,7 +50,12 @@ public class FaceSearchExample {
                 //是否保存人脸及图片数据信息
                 .setStorageFaceInfo(true)
                 //目前只实现了数据库存储，对其他类型存储实现StorageImageService接口即可
-                .setStorageEngine(StorageEngine.CURR_DB);
+                .setStorageEngine(StorageEngine.CURR_DB)
+                //设置分片大小
+                .setShardsNum(shardsNum)
+                //开启关闭近似knn搜索
+                .setApproximateKnn(approximateKnn);
+
         //删除集合
         Response<Boolean> deleteCollect =  faceSearch.collect().deleteCollect();
         System.out.println(deleteCollect);
@@ -97,6 +108,9 @@ public class FaceSearchExample {
                 .search(Search.build(imageBase64)
                 .setConfidenceThreshold(50f)        //最小置信分：50
                 .setMaxFaceNum(10).setLimit(1)
+                //这里其实即使开了近似knn搜索，这里设置为false也可以使用精确knn搜索。
+                //这里数据量足够大时，就能发现近似knn返回较快
+                .setApproximateKnn(approximateKnn)
             );
             Long e = System.currentTimeMillis();
             System.out.println("search cost:" + (e-s)+"ms");