Index for ANN Search

Overview of Deep Lake's Index implementation for ANN search.

How Deep Lake Implements an Index for ANN Search

Deep Lake implements the Hierarchical Navigable Small World (HSNW) index for Approximate Nearest Neighbor (ANN) search. The index is based on the OSS Hsnwlib package with added optimizations. The implementation enables users to run queries on >35M embeddings in less than 1 second.

Unique aspects of Deep Lake's HSNW implementation

• Rapid index creation with multi-threading optimized for Deep Lake

• Efficient memory management that reduces RAM usage

Memory Management in Deep Lake

RAM Cost >> On-disk Cost >> Object Storage Cost

Minimizing RAM usage and maximizing object store significantly reduces costs of running a Vector Database. Deep Lake has a unique implementation of memory allocation that minimizes RAM requirement without any performance penalty:

Memory Architecture for the Deep Lake Vector Store

Using the Index

By default, the index is turned off in Deep Lake. To enable the index, during Vector Store initialization or loading, specify the Vector Store length threshold above which the index will be applied:

vectorstore = VectorStore(path, index_params = {threshold: <threshold_int>})

Limitations

The following limitations of the index are being implemented in upcoming releases:

• Index does not support incremental updates. If any update is made to the dataset, the index is re-created.

• If the search is performed using a combination of attribute and vector search, the index is not used and linear search is applied instead.