Package io.objectbox.annotation

Annotation Interface HnswIndex

@Retention(CLASS) @Target(FIELD) public @interface HnswIndex
Parameters to configure HNSW-based approximate nearest neighbor (ANN) search. Some of the parameters can influence index construction and searching. Changing these values causes re-indexing, which can take a while due to the complex nature of HNSW.

  • Required Element Summary

    Required Elements
    Modifier and Type
    Required Element
    Description
    long
    dimensions
    Dimensions of vectors; vector data with fewer dimensions are ignored.

  • Optional Element Summary

    Optional Elements
    Modifier and Type
    Optional Element
    Description
    VectorDistanceType
    distanceType
    The distance type used for the HNSW index.
    HnswFlags
    flags
    See HnswFlags.
    long
    indexingSearchCount
    Aka "efConstruction": the number of neighbor searched for while indexing (default: 100).
    long
    neighborsPerNode
    Aka "M": the max number of connections per node (default: 30).
    float
    reparationBacklinkProbability
    When repairing the graph after a node was removed, this gives the probability of adding backlinks to the repaired neighbors.
    long
    vectorCacheHintSizeKB
    A non-binding hint at the maximum size of the vector cache in KB (default: 2097152 or 2 GB/GiB).

  • Element Details

    • dimensions

      long dimensions
      Dimensions of vectors; vector data with fewer dimensions are ignored. Vectors with more dimensions than specified here are only evaluated up to the given dimension value. Changing this value causes re-indexing.

    • neighborsPerNode

      long neighborsPerNode
      Aka "M": the max number of connections per node (default: 30). Higher numbers increase the graph connectivity, which can lead to more accurate search results. However, higher numbers also increase the indexing time and resource usage. Try e.g. 16 for faster but less accurate results, or 64 for more accurate results. Changing this value causes re-indexing.
      Default:
      0L

    • indexingSearchCount

      long indexingSearchCount
      Aka "efConstruction": the number of neighbor searched for while indexing (default: 100). The higher the value, the more accurate the search, but the longer the indexing. If indexing time is not a major concern, a value of at least 200 is recommended to improve search quality. Changing this value causes re-indexing.
      Default:
      0L

    • flags

      HnswFlags flags
      See HnswFlags.
      Default:
      @io.objectbox.annotation.HnswFlags

    • distanceType

      VectorDistanceType distanceType
      The distance type used for the HNSW index. Changing this value causes re-indexing.
      Default:
      DEFAULT

    • reparationBacklinkProbability

      float reparationBacklinkProbability
      When repairing the graph after a node was removed, this gives the probability of adding backlinks to the repaired neighbors. The default is 1.0 (aka "always") as this should be worth a bit of extra costs as it improves the graph's quality.
      Default:
      1.0f

    • vectorCacheHintSizeKB

      long vectorCacheHintSizeKB
      A non-binding hint at the maximum size of the vector cache in KB (default: 2097152 or 2 GB/GiB). The actual size max cache size may be altered according to device and/or runtime settings. The vector cache is used to store vectors in memory to speed up search and indexing.

      Note 1: cache chunks are allocated only on demand, when they are actually used. Thus, smaller datasets will use less memory.

      Note 2: the cache is for one specific HNSW index; e.g. each index has its own cache.

      Note 3: the memory consumption can temporarily exceed the cache size, e.g. for large changes, it can double due to multi-version transactions.

      Default:
      0L