PARAMETERS.md

Parameters

Parameter	Description	PQ	SQ	PCA	Flash
K	Specifies the top K nearest neighbors returned by the algorithm.	√	√	√	√
M	Number of neighbors in higher layers. The base layer contains twice as many neighbors as higher layers. Must be a multiple of VECTOR_PER_BLOCK in Flash. Corresponds to $R$ in the original paper.	√	√	√	√
EF_CONSTRUCTION	Maximum number of candidate neighbors considered during index construction. The algorithm prunes these candidates to form the final graph structure. Corresponds to $C$ in the original paper. A higher value improves graph quality but increases indexing time.	√	√	√	√
EF_SEARCH	Maximum number of candidates retained during the search phase. The algorithm selects and evaluates unvisited candidates iteratively. A higher value improves search quality but increases search time.	√	√	√	√
SAMPLE_NUM	In PQ: Number of points sampled for k-means clustering. In PCA: Number of points used to extract principal components.	√		√	√
NUM_THREADS	Number of threads used for parallel processing on the CPU.	√	√	√	√
SUBVECTOR_NUM	Number of subvectors created by splitting the original dimensions. Alternatively, this represents the dimensionality after applying PQ. Must be a multiple of VECTOR_PER_BLOCK in Flash.	√			√
SUBVECTOR_LENGTH	Number of dimensions in each subvector. Total dimensions = SUBVECTOR_NUM * SUBVECTOR_LENGTH. Deprecated.	√			√
CLUSTER_NUM	Number of clusters used in PQ for k-means clustering.	√			√
MAX_ITERATIONS	Maximum number of iterations for k-means clustering in PQ.	√			√
VECTORS_PER_BLOCK	Number of neighbors Flash can process simultaneously. Relevant when PQLINK_CALC is enabled and the CPU supports SIMD. Typically 16 for SSE/AVX; may change to 64 with AVX512.				√
PRINCIPAL_DIM	Number of dimensions retained after PCA transformation.			√	√

Macro Definition

Parameter	Description
INT8/INT16/INT32 (General)	Data type for data_t. PQ and PCA only support INT32. Unsigned data types are used when ALL_POSITIVE_NUMBER is enabled.
ALL_POSITIVE_NUMBER	Modifies nmslib/hnswlib/hnswalg.h to ensure all numbers are positive, allowing the use of unsigned data types.
PQLINK_STORE	Stores the neighbor vectors for each node.
PQLINK_CALC	Calculates neighbor data in batches. Must be used together with PQLINK_STORE.
USE_PCA	Applies PCA to reduce dimensions in Flash.
USE_PCA_OPTIMAL	Groups subvectors based on the sum of cumulative variance in PCA. May slow down related algorithms.
RERANK	Searches for $2K$ points and re-ranks them using the original distance metric.
SAVE_MEMORY	Skips saving the distance table when using SDC to calculate distances.
TAU	Parameter used in tau-MG
VBASE	use vbase to calculate distance
VBASE_WINDOW_SIZE	the window size of vbase
ADSAMPLING	use ADSampling to calculate distance
ADSAMPLING_EPSILON	the epsilon of ADSampling
ADSAMPLING_DELTA_D	the delta_d of ADSampling