DiskANN configuration-OpenSearch(Open Search)-阿里云帮助中心

DiskANN is a disk-based approximate nearest neighbor (ANN) search algorithm for ultra-large-scale datasets that exceed available memory. It stores the Vamana graph index on disk and uses Product Quantization (PQ) to compress vectors in memory, enabling efficient vector search under memory-constrained conditions.

DiskANN operates in two phases:

Index build phase — constructs the Vamana graph on disk. Build parameters control index quality, PQ compression ratio, and construction cost.
Search phase — traverses the graph using compressed vectors to find approximate nearest neighbors. Search parameters control recall and query latency at runtime.

Index build parameters

Parameter	Type	Description
`proxima.diskann.builder.pq_dimension_scale`	Int	The PQ compression scale. The PQ dimension is calculated as: vector dimension / `pq_dimension_scale`. Default: `8`, which compresses vectors to 1/8 of the original dimension. A higher value compresses more aggressively, reducing memory usage at the cost of lower index precision. A lower value preserves more precision at the cost of higher memory usage.
`proxima.diskann.builder.max_degree`	Int	The maximum number of neighbor edges per node in the Vamana graph. A higher value produces a more accurate graph but increases construction time and disk usage.
`proxima.diskann.builder.list_size`	Int	The candidate set size for edge selection during graph construction. A higher value improves graph accuracy but increases construction time.
`proxima.diskann.builder.thread_count`	Int	The number of threads used for index building.

Index search parameters

Parameter	Type	Description
`proxima.diskann.searcher.list_size`	Int	The result candidate set size. A higher value increases recall but also increases query time.
`proxima.diskann.searcher.io_limit`	Int	The disk I/O limit per query. The search performs at most `io_limit` disk read operations, which mainly affects how deeply the algorithm traverses the Vamana graph. A higher value increases recall at the cost of more I/O operations and longer query time.
`proxima.diskann.searcher.beam_search_width`	Int	The number of parallel disk I/O operations per query. A higher value increases disk read throughput, reducing the latency impact of a given `io_limit`.