Indexes

The main index interfaces are listed here and in the next sections. Refer to this table when making decisions about which index to use.

• svs::Vamana - A graph-based similarity search engine suitable for high recall, high throughput, and low latency.

• svs::DynamicVamana - A dynamic graph-based similarity search engine suitable for high recall, high throughput, and low latency.

• svs::Flat - An exhaustive search engine. Provides precise results at the expense of low throughput and high latency for large datasets.

Contents

• Vamana Index
  • svs::Vamana
    • search_parameters_type
    • get_alpha()
    • get_graph_max_degree()
    • get_construction_window_size()
    • get_max_candidates()
    • get_prune_to()
    • get_full_search_history()
    • save()
    • batch_iterator()
    • get_distance()
    • size()
    • dimensions()
    • query_types()
    • get_num_threads()
    • set_threadpool()
    • get_threadpool_handle()
    • assemble()
    • build()
  • Helper Classes
    • svs::index::vamana::VamanaBuildParameters
      • alpha
      • graph_max_degree
      • window_size
      • max_candidate_pool_size
      • prune_to
      • use_full_search_history
    • svs::index::vamana::VamanaSearchParameters
      • buffer_config_
      • search_buffer_visited_set_
      • prefetch_lookahead_
      • prefetch_step_
  • Example
    • Preamble
    • Building and Index
    • Searching the Index
    • Saving the Index
    • Reloading a Saved Index
    • Using Vector Compression
    • Building using Vector Compression
    • Entire Example
• Dynamic Vamana Index
  • svs::DynamicVamana
    • search_parameters_type
    • DynamicVamana()
    • set_search_window_size()
    • get_search_window_size()
    • get_alpha()
    • get_graph_max_degree()
    • get_construction_window_size()
    • get_max_candidates()
    • get_prune_to()
    • get_full_search_history()
    • has_id()
    • all_ids()
    • reconstruct_at()
    • batch_iterator()
    • get_distance()
    • size()
    • dimensions()
    • query_types()
    • get_num_threads()
    • set_threadpool()
    • get_threadpool_handle()
    • svs::DynamicVamana::AssembleTag
• Vamana Iterator
  • svs::VamanaIterator
    • VamanaIterator()
    • parameters_for_current_iteration()
    • query_type()
    • batch_number()
    • size()
    • results()
    • next()
    • restart_next_search()
    • done()
    • update()
  • Example
    • Constructing an Iterator
    • Using the Iterator
    • Stepping the Iterator
    • Exhausting the Candidate Pool
  • Main BatchIterator Class
    • svs::index::vamana::BatchIterator
      • iterator
      • const_iterator
      • BatchIterator()
      • update()
      • adapt()
      • begin()
      • end()
      • begin()
      • end()
      • cbegin()
      • cend()
      • contents()
      • size()
      • batch_number()
      • done()
      • restart_next_search()
      • parameters_for_current_iteration()
      • next()
  • Full Example
• Flat Orchestrator
  • svs::Flat
    • search_parameters_type
    • save()
    • get_distance()
    • size()
    • dimensions()
    • query_types()
    • get_num_threads()
    • set_threadpool()
    • get_threadpool_handle()
    • assemble()
    • svs::Flat::AssembleTag

Compatible Loaders

The table below lists the compatibility between different index and the different kinds of dataset (e.g., uncompressed, lvq etc).

Index and Loader Compatibility Matrix

Loader / Right: Index	svs::Vamana	svs::Flat
svs::VectorDataLoader	Build and Search	Search Only
svs::quantization::lvq::LVQDataset	Build and Search	Search Only

Note

As an implementation detail, the indexes here are type-erased instances of more specialized, heavily templated implementations. These low level implementations generally have a similar API to the top level indexes and can be more expressive as functions calls do not have to cross a virtual function boundary. However, this specialization results in a different concrete type for each implementation which can be undesireable. More documentation regarding the low level implementation will be available in the future.