一、示例背景
假设我们需要针对视频进行向量检索，挑战是每条视频抽帧后的图片数量不一样，此时我们可以通过 nested 嵌套字段来实现。
举例：一条视频信息有 id、title、content、image 字段，其中 image 保存视频抽帧后的图片向量数据，每个视频有 n 张图，n 值不固定。
二、创建索引
创建 mappings 时指定 image 为 nested 嵌套字段：
//id、title、content 为文本字段//image中的num为图片编号，比如第几帧（可不要）//image中的emb为图片embedding之后的数据，image[image1_emb、image2_emb、image3_emb…imagen_emb]PUT /image_embeddings{ "mappings": { "properties": { "id": { "type": "keyword" }, "title": { "type": "text" }, "content": { "type": "text" }, "image": { "type": "nested", "properties": { "num": {"type": "keyword"}, "emb": { "type": "dense_vector", "dims": 5, "index_options": { "type": "int8_hnsw" }, "similarity": "cosine" } } } } }}
三、写入数据
写入格式用数组形式将多组向量包起来，格式如下：
POST /image_embeddings/_doc/1{ "id": "book_001", "title": "晴天", "content": "刮风这天，我试着握着你手", "image": [ {"num": "0", "emb": [0.1,0.2,0.3,0.4,0.5]}, {"num": "1", "emb": [0.6,0.7,0.8,0.9,1.0]}, {"num": "2", "emb": [0.2,0.3,0.4,0.5,0.6]} ]}POST /image_embeddings/_doc/2{ "id": "book_002", "title": "一路向北", "content": "一路向北，我试着握着你手", "image": [ {"num": "0", "emb": [0.1,0.2,0.3,0.4,0.5]}, {"num": "1", "emb": [0.6,0.7,0.8,0.9,1.0]} ]}POST /image_embeddings/_doc/3{ "id": "book_003", "title": " 双截棍", "content": "哼哼哈嘿", "image": [ {"num": "0", "emb": [0.1,0.2,0.3,0.4,0.5]}, {"num": "1", "emb": [0.6,0.7,0.8,0.9,1.0]}, {"num": "2", "emb": [0.1,0.2,0.3,0.4,0.5]}, {"num": "3", "emb": [0.6,0.7,0.8,0.9,1.0]}, {"num": "4", "emb": [0.1,0.2,0.3,0.4,0.5]}, {"num": "5", "emb": [0.6,0.7,0.8,0.9,1.0]} ]}
四、执行向量搜索
查询数据，写法与前述混合检索算法一致，nested 嵌套字段的评分将取 max 作为最终评分。
GET book-index/_search{ "retriever": { "rrf": { "retrievers": [ { "retriever": { "knn": { "field": "image.emb", "query_vector": [0.1, 0.2, 0.3,0.4,0.5], "k": 5, "num_candidates": 50 } }, "weight": 0.8 }, { "retriever": { "standard": { "query": { "match": { "title": "晴天" } } } }, "weight": 0.2 } ], "rank_window_size": 50, "rank_constant": 20 } }}