Creating Vector Indexes in an Elasticsearch Cluster

Prerequisites

You have created a cluster by referring to Cluster Node Specifications Selection for Vector Search. The cluster must be an Elasticsearch cluster of version 7.6.2 or 7.10.2, or an OpenSearch cluster of version 1.3.6.

(Optional) Preparations

Before creating a vector index, configure advanced settings for the cluster based on service needs.

• When importing data offline, you are advised to set refresh_interval of indexes to -1 to disable automatic index refreshing and thus improve batch write performance.
• You are advised to set number_of_replicas to 0. After the offline data import is complete, you can modify the parameter value again as needed.
• Table 1 describes the other advanced settings.

  Table 1 Parameters for advanced cluster settings

  Parameter	Description
  native.cache.circuit_breaker.enabled	Whether to enable the circuit breaker for off-heap memory. Default value: true
  native.cache.circuit_breaker.cpu.limit	Upper limit of off-heap memory usage of the vector index. For example, if the overall memory of a host is 128 GB and the heap memory occupies 31 GB, the default upper limit of the off-heap memory usage is 43.65 GB, that is, (128 - 31) x 45%. If the off-heap memory usage exceeds its upper limit, the circuit breaker will be triggered. Default value: 45%
  native.cache.expire.enabled	Whether to enable the cache expiration policy. If this parameter is set to true, some cache items that have not been accessed for a long time will be cleared. Value: true or false Default value: false
  native.cache.expire.time	Expiration time. Default value: 24h
  native.vector.index_threads	Number of threads used for creating underlying indexes. Each shard uses multiple threads. Set a relatively small value to avoid resource preemption caused by the build queries of too many threads. Default value: 4

(Optional) Pre-Building and Registering a Center Point Vector

If index algorithms IVF_GRAPH and IVF_GRAPH_PQ are selected when creating a vector index, you need to pre-build and register a center point vector.

The vector index acceleration algorithms IVF_GRAPH and IVF_GRAPH_PQ are suitable for ultra-large-scale computing. These two algorithms allow you to narrow down the query scope by dividing a vector space into subspaces through clustering or random sampling. Before pre-build, you need to obtain all center point vectors by clustering or random sampling. Center point vectors are pre-built into the GRAPH or GRAPH_PQ index and register them with the Elasticsearch cluster. All nodes in the cluster can share this index file. Reuse of the center index among shards can effectively reduce the training overhead and the number of center index queries, improving the write and query performance.

1. On the Clusters page, locate the target cluster, and click Access Kibana in the Operation column.
2. Click Dev Tools in the navigation tree on the left.
3. Create a center point index table.
   • For example, if the created index is named my_dict, number_of_shards of the index must be set to 1. Otherwise, the index cannot be registered.
   • If you want to use the IVF_GRAPH index, set algorithm of the center point index to GRAPH.
   • If you want to use the IVF_GRAPH_PQ index, set algorithm of the center point index to GRAPH_PQ.

   PUT my_dict 
    { 
      "settings": { 
        "index": { 
          "vector": true 
        }, 
        "number_of_shards": 1, 
        "number_of_replicas": 0 
      }, 
      "mappings": { 
        "properties": { 
          "my_vector": { 
            "type": "vector", 
            "dimension": 2, 
            "indexing": true, 
            "algorithm": "GRAPH", 
            "metric": "euclidean" 
          } 
        } 
      } 
    }

4. Write the center point vector to the created index.

   Write the center point vector obtained through sampling or clustering into the created my_dict index. For details, see Importing Vector Data.

5. Call the registration API.

   Register the created my_dict index with a Dict object with a globally unique identifier name (dict_name).

   PUT _vector/register/my_dict 
    { 
      "dict_name": "my_dict" 
    }

6. Create an IVF_GRAPH or IVF_GRAPH_PQ index.

   You do not need to specify the dimension or metric information. Simply specify the registered dictionary name.

   PUT my_index 
    { 
      "settings": { 
        "index": { 
          "vector": true,
          "sort.field": "my_vector.centroid" # Set the centroid subfield of each vector field as a sorting field.
        } 
      }, 
      "mappings": { 
        "properties": { 
          "my_vector": { 
            "type": "vector", 
            "indexing": true, 
            "algorithm": "IVF_GRAPH", 
            "dict_name": "my_dict", 
            "offload_ivf": true 
          } 
        } 
      } 
    }

   Table 2 Field mappings parameters

   Parameter	Description
   dict_name	Specifies the name of the depended central point index. The vector dimensions and metrics of the index are the same as those of the Dict index.
   offload_ivf	Unloads the IVF inverted index implemented by the underlying index to Elasticsearch. The reduces the use of non-heap memory and the overhead of write and merge operations. You are advised to set this parameter to true. Value: true or false Default value: false

Creating a Vector Index

1. Log in to the CSS management console.
2. Choose Clusters in the navigation pane. On the Clusters page, locate the target cluster and click Access Kibana in the Operation column.
3. Click Dev Tools in the navigation tree on the left and run the following command to create a vector index.

   Create an index named my_index that contains a vector field my_vector and a text field my_label. The vector field creates the graph index and uses Euclidean distance to measure similarity.

   PUT my_index 
   {
     "settings": {
       "index": {
         "vector": true
       }
     },
     "mappings": {
       "properties": {
         "my_vector": {
           "type": "vector",
           "dimension": 2,
           "indexing": true,
           "algorithm": "GRAPH",
           "metric": "euclidean"
         },
         "my_label": {
           "type": "keyword"
         }
       }
     }
   }

   Table 3 Parameters for creating an index

   Type	Parameter	Description
   Index settings parameters	vector	To use a vector index, set this parameter to true.
   	vector.exact_search_threshold	The minimum size of the intermediate result set that triggers a switchover from pre-filtering search to brute-force search. When the size of the intermediate result set in a segment is lower than this threshold, brute-force search starts. The default value is null, which disables a switchover to brute-force search.
   Field mappings parameters	type	Field type, for example, vector.
   	dimension	Vector dimensionality. Value range: [1, 4096]
   	indexing	Whether to enable vector index acceleration. The value can be: false: disables vector index acceleration. If this parameter is set to false, vector data is written only to docvalues, and only ScriptScore and Rescore can be used for vector query. true: enables vector index acceleration. If this parameter is set to true, an extra vector index is created. The index algorithm is specified by the algorithm field and VectorQuery can be used for data query. Default value: false
   	algorithm	Index algorithm. This parameter is valid only when indexing is set to true. The value can be: FLAT: brute-force algorithm that calculates the distance between the target vector and all vectors in sequence. The algorithm relies on sheer computing power and its recall rate reaches 100%. You can use this algorithm if you require high recall accuracy. GRAPH: Hierarchical Navigable Small Worlds (HNSW) algorithm for graph indexes. This algorithm is mainly used in scenarios where high performance and precision are required and the data records of a single shard is fewer than 10 million. GRAPH_PQ: combination of the HNSW algorithm and the PQ algorithm. The PQ algorithm reduces the storage overhead of original vectors, so that HNSW can easily search for data among hundreds of millions of records. IVF_GRAPH: combination of IVF and HNSW. The entire space is divided into multiple cluster centroids, which makes search much faster but slightly inaccurate. You can use this algorithm if you require high performance when searching for data among hundreds of millions of records. IVF_GRAPH_PQ: combination of the PQ algorithm with the IVF or HNSW algorithm to further improve the system capacity and reduce the system overhead. This algorithm is applicable to scenarios where there are more than 1 billion files in shards and high retrieval performance is required. Default value: GRAPH NOTE: If IVF_GRAPH or IVF_GRAPH_PQ is specified, you need to pre-build and register a central point index. For details, see (Optional) Pre-Building and Registering a Center Point Vector.
   	Other optional parameters	If Indexing is set to true, CSS provides optional parameters for vector search that you can configure to achieve higher query performance or precision. For more information, see Table 4.
   	metric	Method of calculating the distance between vectors. The value can be: euclidean: Euclidean distance inner_product: inner product distance cosine: cosine distance hamming: Hamming distance, which can be used only when dim_type is set to binary. Default value: euclidean
   	dim_type	Type of the vector dimension value. The value can be binary and float (default).

   Table 4 Optional parameters

   Type	Parameter	Description
   Graph index configuration parameters	neighbors	Number of neighbors of each vector in a graph index. The default value is 64. A larger value indicates higher query precision. A larger index results in a slower build and query speed. Value range: [10, 255]
   	shrink	Cropping coefficient during HNSW build. The default value is 1.0f. Value range: (0.1, 10)
   	scaling	Scaling ratio of the upper-layer graph nodes during HNSW build. The default value is 50. Value range: (0, 128]
   	efc	Queue size of the neighboring node during HNSW build. The default value is 200. A larger value indicates a higher precision and slower build speed. Value range: (0, 100000]
   	max_scan_num	Maximum number of nodes that can be scanned. The default value is 10000. A larger value indicates a higher precision and slower indexing speed. Value range: (0, 1000000]
   PQ index configuration parameters	centroid_num	Number of cluster centroids of each fragment. The default value is 255. Value range: (0, 65535]
   	fragment_num	Number of fragments. The default value is 0. The plug-in automatically sets the number of fragments based on the vector length. Value range: [0, 4096]

Importing Vector Data

Run the following command to import vector data. When writing vector data to the my_index index, you need to specify the vector field name and vector data.

• If the input vector data is an array of floating-point numbers separated by commas (,):

  POST my_index/_doc
  {
    "my_vector": [1.0, 2.0]
  }

• If the input vector data is a Base64 string encoded using little endian:
  When writing binary vectors or high dimensional vectors that have a large number of valid bits, the Base64 encoding format is efficient for data transmission and parsing.

  POST my_index/_doc
  {
    "my_vector": "AACAPwAAAEA="
  }

• To write a large amount of data, bulk operations are recommended.

  POST my_index/_bulk
  {"index": {}}
  {"my_vector": [1.0, 2.0], "my_label": "red"}
  {"index": {}}
  {"my_vector": [2.0, 2.0], "my_label": "green"}
  {"index": {}}
  {"my_vector": [2.0, 3.0], "my_label": "red"}