Member Junction
    Preparing search index...
    interface VectorAccelerationOptions {
        ANN: VectorANNOptions;
        ClusterOffloadMinRows: number;
        NativeMinWork: number;
        OffloadMinWork: number;
        PartitionWork: number;
        PoolSize: number;
        TaskTimeoutMs: number;
        UseNative: boolean;
        WorkerScriptPath?: string;
    }
    Index

    Properties

    Approximate nearest-neighbour search. Unlike every other path here it can miss true neighbours (scores are still exact), so it is opt-in and only applies to cosine searches over stores of at least MinRows rows.

    ClusterOffloadMinRows: number

    Smallest store, in rows, whose clustering is sent to a worker. Default 200.

    NativeMinWork: number

    Smallest unfiltered search, in rows × dimensions, routed to native code. Default 100,000.

    OffloadMinWork: number

    Smallest search, in rows × dimensions, worth sending to a worker. Below this the message round trip costs more than the scan. Default 2,000,000 (about 1,300 rows of 1,536-dimension embeddings).

    PartitionWork: number

    Rows × dimensions per partition when one search is split across several workers. Default 8,000,000.

    PoolSize: number

    Worker threads for off-thread search and clustering. 0 keeps everything in-process.

    TaskTimeoutMs: number

    A worker task running longer than this is abandoned and its worker replaced. Default 120,000 ms.

    UseNative: boolean

    Use the native usearch backend when it is installed. Default true.

    WorkerScriptPath?: string

    Path to the compiled worker script. Defaults to the one shipped beside this package's dist.