MongoDB 搜索

Doubao 中英对照 MongoDB Search

MongoDB 让用户能够使用专用的搜索索引进行关键词或词法搜索，以及向量搜索数据。

向量搜索

MongoDB 向量搜索使用 $vectorSearch 聚合阶段，对专用索引运行查询。关于 vectorSearch 索引的要求与限制，请参阅 MongoDB 文档了解更多信息。

管理向量索引

由 MongoTemplate 实现的 SearchIndexOperationsProvider 是 SearchIndexOperations 的入口点，后者提供了各种用于管理向量索引的方法。

以下代码片段展示了如何为集合创建向量索引

Java

VectorIndex index = new VectorIndex("vector_index")
  .addVector("plotEmbedding", vector -> vector.dimensions(1536).similarity(COSINE)) 1
  .addFilter("year"); 2

mongoTemplate.searchIndexOps(Movie.class) 3
  .createIndex(index);

• A vector index may cover multiple vector embeddings that can be added via the addVector method.

• Vector indexes can contain additional fields to narrow down search results when running queries.

• Obtain SearchIndexOperations bound to the Movie type which is used for field name mapping.

Mongo Shell

db.movie.createSearchIndex("movie", "vector_index",
  {
    "fields": [
      {
        "type": "vector",
        "numDimensions": 1536,
        "path": "plot_embedding", // <1>
        "similarity": "cosine"
      },
      {
        "type": "filter",
        "path": "year"
      }
    ]
  }
)

• Field name plotEmbedding got mapped to plot_embedding considering a @Field(name = "…​") annotation.

向量索引创建后并不会立即可用，尽管 exists 检查返回 true。搜索索引的实际状态可通过 SearchIndexOperations#status(…) 获取。READY 状态表示索引已准备好接受查询。

查询向量索引

可以通过 MongoOperations 发起一个使用 VectorSearchOperation 的聚合来查询向量索引，如下例所示

Java

VectorSearchOperation search = VectorSearchOperation.search("vector_index") 1
  .path("plotEmbedding") 2
  .vector( ... )
  .numCandidates(150)
  .limit(10)
  .withSearchScore("score"); 3

AggregationResults<MovieWithSearchScore> results = mongoTemplate
  .aggregate(newAggregation(Movie.class, search), MovieWithSearchScore.class);

• Provide the name of the vector index to query since a collection may hold multiple ones.

• The name of the path used for comparison.

• Optionally add the search score with given name to the result document.

Mongo Shell

db.embedded_movies.aggregate([
  {
    "$vectorSearch": {
      "index": "vector_index",
      "path": "plot_embedding", // <1>
      "queryVector": [ ... ],
      "numCandidates": 150,
      "limit": 10
    }
  },
  {
    "$addFields": {
      "score": { $meta: "vectorSearchScore" }
    }
  }
])

• Field name plotEmbedding got mapped to plot_embedding considering a @Field(name = "…​") annotation.