ScaNN Index reference

This page provides reference material for the ScaNN Index.

Tuning parameters

The following index parameters and database flags are used together to find the right balance of recall and QPS.

Tuning parameter Description Option type
max_num_levels The maximum number of centroid levels of the K-means clustering tree.
  • Two-level tree index: Set to 1 by default for a two-level tree (1 centroid level + bottom leaf level).
  • Three-level tree index: Set to 2 by default for a three-level tree (2 centroid levels + bottom leaf level)
  • Set the value to 2 if the number of vector rows exceeds 100 million rows.
  • Set the value to 1 if the number of vector rows are less than 10 million rows.
  • Set to either 1 or 2 if the number of vector rows lie between 10 million and 100 million rows to optimize for index build time (set to 2) or optimize for search recall (set to 1).
 Index creation
(optional)
num_leaves The number of partitions to apply to this index. The number of partitions you apply to when creating an index affects the index performance. By increasing partitions for a set number of vectors, you create a more fine-grained index, which improves recall and query performance. However, this comes at the cost of longer index creation times.

Since three-level trees build faster than two-level trees, you can increase the num_leaves_value when creating a three-level tree index to achieve better performance.
  • Two-level index: Set this value to any value between 1 and 1048576.

    For an index that balances fast index build and good search performance, use sqrt(ROWS) as a starting point, where ROWS is the number of vector rows. The number of vectors that each partition holds is calculated by
    ROWS/sqrt(ROWS) = sqrt(ROWS).

    Since a two-level tree index can be created on a dataset with less than 10 million vector rows, each partition will hold less than (sqrt(10M)) vectors, which is 3200 vectors. For optimal vector search quality, it's recommended to minimize the number of vectors in each partition. The recommended partition size is about 100 vectors per partition, so set num_leaves to ROWS/100. If you have 10 million vectors you would set num_leaves to 100,000.
  • Three-level index: Set this value to any value between 1 and 1048576.

    If you are unsure about selecting the exact value, use power(ROWS, 2/3) as a starting point, where ROWS is the number of vector rows. The number of vectors that each partition holds is calculated by
    ROWS/power(ROWS, 2/3) = power(ROWS, 1/3).

    Since a three-level tree index can be created on a dataset with vector rows more than 100 million, each partition will hold more than
    (power(100M, 1/3)) vectors, which is 465 vectors. For optimal performance, it's recommended to minimize the number of vectors in each partition. The recommended partition size is about 100 vectors per partition, so set num_leaves to ROWS/100. If you have 100 million vectors you would set num_leaves to 1 million.
 Index creation
(required)
quantizer The type of quantizer you want to use for the K-means tree. The default value is set to SQ8 which provides better query performance with minimal recall loss (typically less than 1-2%).

Set it to FLAT if a recall of 99% or higher is required.		 Index creation
(optional)
scann.enable_pca Enables Principal Component Analysis (PCA), which is a dimension reduction technique used to automatically reduce the size of the embedding when possible. This option is enabled by default.

Set to false if you observe deterioration in recall.		 Index creation
(optional)
scann.num_leaves_to_search This database flag controls the absolute number of leaves or partitions to search which lets you trade off between recall and QPS. The default value is 1% of the value set in num_leaves.

A higher value will result in better recall but lower QPS. Similarly, a lower value will result in lower recall but higher QPS.		 Query runtime
(optional)
scann.pre_reordering_num_neighbors The database flag, when set, specifies the number of candidate neighbors to consider during the reordering stages after the initial search identifies a set of candidates. Set this parameter to a value higher than the number of neighbors you want the query to return.

A higher value results in better recall, but a lower QPS. Set this value to 0 to disable reordering. The default is 0 if PCA is not enabled during index creation. Otherwise, the default is 50 x K, where K is the LIMIT specified in the query. 		Query runtime
(optional)
scann.num_search_threads The number of searcher threads for multi-thread search. This can help reduce single query latency by using more than one thread for ScaNN ANN search in latency-sensitive applications. This setting doesn't improve single query latency if the database is already cpu-bound. The default value is 2. Query runtime
(optional)

What's next