v0.1.0

What's Changed

• Creation of SVS C++ Runtime Bindings package (#208)
• Addition of libsvs conda and python packages along with existing tarballs
• Added shared library support for Clang-20
• LVQ8x8 support added to SVS python package
• Fixes related to VamanaBuildParameters (#237) and unmasked AVX2 load (#239)

Full Changelog: v0.0.11...v0.1.0

v0.0.11

What's Changed

IVF support for SQDataset (#184), LVQ, and LeanVec

Various bug resolutions, fixes, and tooling support

Full Changelog: v0.0.10...v0.0.11

v0.0.10

What's Changed

• Support for OpenMP threading (#170, #171, #172)
• Enable Inverted File index (#156)
• Resolution of inf/nan crashes in LVQ (#174, #178)
• SVS_LAZY update to address alpine issue (#175)
• Updates to build parameter defaults (#176)
• MKL symbols not exported from SVS libs

Full Changelog: v0.0.9...v0.0.10

v1.0.0-dev

Rolling release page for next major release, 1.0.0. This page provides nightly builds with the latest commits to main. See v0.0.9...main

v0.0.9

What's Changed

• fix: Multi-vector dynamic vamana index Save/Load functionality by @rfsaliev in #162
• feature: save and load in flat index by @yuejiaointel in #163
• Handle corner in dynamic index with insufficient valid search results by @ibhati in #164
• change to v0.0.9 by @ahuber21 in #169

Full Changelog: v0.0.8...v0.0.9

v0.0.8

• Addition of 8-bit scalar quantization support to C++ interface

• Introduced multi-vector index and batch iterator support that allows multiple vectors to be mapped to the same external ID

• Automatic ISA dispatching with optimizations based on AVX support

• Enabled compatibility with ARM and MacOS

• Enhanced logging capabilities

• Updated vamana iterator API

• Broader shared library support:

  • gcc-11+, clang-18+, glibc 2.26+ compatibility

  • Static library provided in addition to .so

  • Intel(R) MKL linked within the shared library - no need for Intel(R) MKL in user environment

Note

SVS shared and static libraries are included in the tarballs. Binary variants included in this release:

• svs-shared-library-0.0.8.tar.gz: Standard build for GCC 11+ and glibc 2.28+ — compatible with most modern Linux distributions
• glibc2_26 suffix: Build for GCC 11+ and glibc 2.26 (e.g. Amazon Linux 2)
• clang suffix: Build for Clang-18+
• reduced suffix: Lighter-weight builds for specific integrations — not recommended for general use

v0.0.8-dev

Note that the shared library binaries are built with gcc-11 (unless suffixed with "clang" - which is built with clang-18), GLIBC 2.28, oneAPI 2024.1, cmake 3.26.5. These versions or later are required for use of the shared library (except oneAPI/MKL which is only required if using LeanVec). All 0.0.8 binaries are statically linked to MKL and do not require it in user env.

Note that most of these are compatible with AVX2 and above, unless specifically suffixed with AVX512 (these are compatible with AVX512 only)

v0.0.7

SVS 0.0.7 Release Notes

• Implemented batch iterator support for hybrid search

• Added support for custom threading and memory allocation

• Introduced a timeout feature for search calls

• Introduced reuse_empty flag in dynamic Vamana, enabling users to choose whether to reuse empty entries that may exist after deletion and consolidation

• Enhanced heuristics in the Vamana construct to improve efficiency when adding a small number of points.

Note that the shared library binaries are built with gcc 11.2.0, GLIBC 2.28, oneAPI 2024.1, cmake 3.26.5. These versions or later are required for use of the shared library (except oneAPI/MKL which is only required if using LeanVec). Also note that some of these are under active development and may change.

Use the avx512 binaries on machines with AVX512 instruction support for best performance.

v0.0.6

SVS 0.0.6 Release Notes

Please note that this repository only contains the open-source portion of the SVS library, which supports all functionalities and features described in the documentation, except for our proprietary vector compression techniques, specifically LVQ [ABHT23] and Leanvec [TBAH24]. These techniques are closed-source and supported exclusively on Intel hardware. We provide shared library and PyPI package to enable these vector compression techniques in C++ and Python, respectively.

v0.0.3

SVS 0.0.3 Release Notes

Highlighted Features

• Turbo LVQ: A SIMD optimized layout for LVQ that can improve end-to-end search
  performance for LVQ-4 and LVQ-4x8 encoded datasets.

• Split-buffer: An optimization that separates the search window size used during greedy
  search from the actual search buffer capacity. For datasets that use reranking (two-level
  LVQ and LeanVec), this allows more neighbors to be passed to the reranking phase without
  increasing the time spent in greedy search.

• LeanVec dimensionality reduction is now included as
  an experimental feature!
  This two-level technique uses a linear transformation to generate a primary dataset with
  lower dimensionality than full precision vectors.
  The initial portion of a graph search is performed using this primary dataset, then uses
  the full precision secondary dataset to rerank candidates.
  Because of the reduced dimensionality, LeanVec can greatly accelerate index constructed
  for high-dimensional datasets.

  As an experimental feature, future changes to this API are expected.
  However, the implementation in this release is sufficient to enable experimenting with
  this technique on your own datasets!

New Dependencies

• MKL:
  Required by LeanVec.
• toml: Required by testing infrastructure.

pysvs (Python)

Additions and Changes

• Added the LeanVecLoader class as a dataset loader enabling use of
  LeanVec dimensionality reduction.

  The main constructor is shown below:

  pysvs.LeanVecLoader(
      loader: pysvs.VectorDataLoader,
      leanvec_dims: int,
      primary: pysvs.LeanVecKind = pysvs.LeanVecKind.lvq8,
      secondary: pysvs.LeanVecKind = pysvs.LeanVecKind.lvq8
  )
  

  where:

  • loader is the loader for the uncompressed dataset.
  • leanvec_dims is the target reduced dimensionality of the primary dataset.
    This should be less than loader.dims to provide a performance boost.
  • primary is the encoding to use for the reduced-dimensionality dataset.
  • secondary is the encoding to use for the full-dimensionality dataset.

  Valid options for pysvs.LeanVecKind are: float16, float32, lvq4, lvq8.

  See the documentation for docstrings and an example.

• Search parameters controlling recall and performance for the Vamana index are now set and
  queried through a pysvs.VamanaSearchParameters configuration class. The layout of this
  class is as follows:

  class VamanaSearchParameters
  
  Parameters controlling recall and performance of the VamanaIndex.
  See also: `Vamana.search_parameters`.
  
  Attributes:
      buffer_config (`pysvs.SearchBufferConfig`, read/write): Configuration state for the
          underlying search buffer.
      search_buffer_visited_set (bool, read/write): Enable/disable status of the search
          buffer visited set.
  

  with pysvs.SearchBufferConfig defined by

  class pysvs.SearchBufferConfig
  
  Size configuration for the Vamana index search buffer.
  See also: `pysvs.VamanSearchParameters`, `pysvs.Vamana.search_parameters`.
  
  Attributes:
      search_window_size (int, read-only): The number of valid entries in the buffer
          that will be used to determine stopping conditions for graph search.
      search_buffer_capacity (int, read-only): The (expected) number of valid entries that
          will be available. Must be at least as large as `search_window_size`.
  

  Example usage is shown below.

  index = pysvs.Vamana(...);
  # Get the current parameters of the index.
  parameters = index.search_parameters
  print(parameters)
  # Possible Output: VamanaSearchParameters(
  #    buffer_config = SearchBufferConfig(search_window_size = 0, total_capacity = 0),
  #    search_buffer_visited_set = false
  # )
  
  # Update our local copy of the search parameters
  parameters.buffer_config = pysvs.SearchBufferConfig(10, 20)
  # Assign the modified parameters to the index. Future searches will be affected.
  index.search_parameters = parameters

• Split search buffer for the Vamana search index. This is achieved by using different
  values for the search_window_size and search_buffer_capacity fields of the
  pysvs.SearchBufferConfig class described above.

  An index configured this way will maintain more entries in its search buffer while still
  terminating search relatively early. For implementation like two-level LVQ that use
  reranking, this can boost recall without significantly increasing the effective
  search window size.

  For uncompressed indexes that do not use reranking, split-buffer can be used to decrease
  the search window size lower than the requested number of neighbors (provided the
  capacity is at least the number of requested neighbors). This enables continued trading
  of recall for search performance.

• Added pysvs.LVQStrategy for picking between different flavors of LVQ. The values
  and meanings are given below.

  • Auto: Let pysvs decide from among the available options.
  • Sequential: Use the original implementation of LVQ which bit-packs subsequent vector
    elements sequentially in memory.
  • Turbo: Use an experimental implementation of LVQ that permutes the packing of
    subsequent vector elements to permit faster distance computations.

  The selection of strategy can be given using the strategy keyword argument of
  pysvs.LVQLoader and defaults to pysvs.LVQStrategy.Auto.

• Index construction and loading methods will now list the registered index specializations.

• Assigning the padding keyword to LVQLoader will now be respected when reloading a
  previously saved LVQ dataset.

• Changed the implementation of the greedy-search visited set to be effective when operating
  in the high-recall/high-neighbors regime. It can be enabled with:

  index = pysvs.Vamana(...)
  p = index.search_parameters
  p.search_buffer_visited_set = True
  index.search_parameters = p

Experimental Features

Features marked as experimental are subject to rapid API changes, improvement, and
removal.

• Added the experimental_backend_string read-only parameter to pysvs.Vamana to aid in
  recording and debugging the backend implementation.

• Introduced pysvs.Vamana.experimental_calibrate to aid in selecting the best runtime
  performance parameters for an index to achieve a desired recall.

  This feature can be used as follows:

  # Create an index
  index = pysvs.Vamana(...)
  # Load queries and groundtruth
  queries = pysvs.read_vecs(...)
  groundtruth = pysvs.read_vecs(...)
  # Optimize the runtime state of the index for 0.90 10-recall-at-10
  index.experimental_calibrate(queries, groundtruth, 10, 0.90)

  See the documentation for a more detailed explanation.

Deprecations

• Versions 0.0.1 and 0.0.2 of VamanaConfigParameters (the top-level configuration file
  for the Vamana index) are deprecated. The current version is now v0.0.3. Older versions
  will continue to work until the next minor release of SVS.

  To upgrade, use the convert_legacy_vamana_index binary utility described below.

• The attribute pysvs.Vamana.visisted_set_enabled is deprecated and will be removed in the
  next minor release of SVS. It is being replaced with pysvs.Vamana.search_parameters.

• The LVQ loader classes pysvs.LVQ4, pysvs.LVQ8, pysvs.LVQ4x4, pysvs.LVQ4x8 and
  pysvs.LVQ8x8 are deprecated in favor of a single class pysvs.LVQLoader. This class
  has similar arguments to the previous family, but encodes the number of bits for the
  primary and residual datasets as run-time values.

  For example,

  # Old
  loader = pysvs.LVQ4x4("dataset", dims = 768, padding = 32)
  # New
  loader = pysvs.LVQLoader("dataset", primary = 4, residual = 4, dims = 768, padding = 32)

• Version v0.0.2 of serialized LVQ datasets is broken, the current version is now
  v0.0.3. This change was made to facilitate a canonical on-disk representation of LVQ.

  Goind forward, previously saved LVQ formats can be reloaded using different runtime
  alignments and different packing strategies without requiring whole dataset recompression.

  Any previously saved datasets will need to be regenerated from uncompressed data.

Build System Changes

Building pysvs using cibuildwheel now requires a custom docker container with MKL.
To build the container, run the following commands:

cd ./docker/x86_64/manylinux2014/
./build.sh

libsvs (C++)

Changes

• Added svs::index::vamana::VamanaSearchParameters and
  svs::index::vamana::SearchBufferConfig. The latter contains parameters for the search
  buffer sizing while the former groups all algorithmic and performance parameters of search
  together in a single class.
• API addition of get_search_parameters() and set_search_parameters() to svs::Vamana
  and svs::DynamicVamana as the new API for getting and setting all search parameters.
• Introducing split-buffer for the search buffer (see description in the Python section)
  to potentially increase recall when using reranking.
• Overhauled LVQ implementation, adding an additional template parameter to
  lvq::CompressedVectorBase and friends. This parameter assumes the following types:
  • lvq::Sequential: Store dimension encodings sequentially in memory. This corresponds
    to the original LVQ implementation.
  • lvq::Turbo<size_t Lanes, size_t ElementsPerLane>: Use a SIMD optimized format,
    optimized to use Lanes SIMD lanes, storing ElementsPerLane. Selection of these
    parame...