All Products
Search
Document Center

OpenSearch:QGraph (Quantized Graph) configuration

Last Updated:Oct 19, 2026

This topic describes how to configure the Build parameters and Search parameters of the QGraph algorithm.

Build parameters for QGraph indexes

Parameter

Type

Default value

Description

proxima.hnsw.builder.max_neighbor_count

uint32

100

The maximum number of neighbors for each node in the graph. A larger value indicates better graph connectivity but also increases the graph construction cost and index size.

proxima.hnsw.builder.efconstruction

uint32

500

The size of the neighbor scanning area during graph construction. A larger value results in better offline graph quality but slower index building. We recommend that you start with an initial value of 400.

proxima.hnsw.builder.thread_count

uint32

0

The number of threads used during building. If you set this parameter to 0, the number of CPU cores is used.

proxima.qgraph.builder.quantizer_class

string

None

The quantizer. Valid values: Int4QuantizerConverter, Int8QuantizerConverter, and HalfFloatConverter. In general, fewer integer bits result in a smaller index, higher performance, and lower recall rate. Because of limitations in the underlying CPU instruction set, int16 quantization delivers almost the same performance and recall rate as no quantization.

Search parameters for QGraph indexes

Parameter name

Type

Default value

Description

proxima.hnsw.searcher.ef

uint32

500

Controls the size of the subgraph to scan during online retrieval. A larger value increases the retrieval rate but degrades performance. The recommended value range is [100, 1000].

proxima.hnsw.searcher.max_scan_ratio

float

None

Controls the ratio of points to scan during online retrieval. A larger value increases the retrieval rate but degrades performance.

proxima.hnsw.searcher.brute_force_threshold

uint32

None

Controls the minimum number of points to scan during online retrieval. If this value is greater than the number of documents in a single shard, the query degrades to a brute-force search.