Counting 3,663 Big Data & Machine Learning Frameworks, Toolsets, and Examples...
Suggestion? Feedback? Tweet @stkim1

Last Commit
Mar. 21, 2019
Jul. 10, 2013

Pypi version Build Status Windows Build Status Join the chat at

Non-Metric Space Library (NMSLIB)

The latest pre-release is Note that the manual is not updated to reflect some of the changes. In particular, we changed the build procedure for Windows. Also note that the manual targets primiarily developers who will extend the library. For most other folks, Python binding docs should be sufficient. The basic parameter tuning/selection guidelines are also available online.

Non-Metric Space Library (NMSLIB) is an efficient cross-platform similarity search library and a toolkit for evaluation of similarity search methods. The core-library does not have any third-party dependencies.

The goal of the project is to create an effective and comprehensive toolkit for searching in generic non-metric spaces. Being comprehensive is important, because no single method is likely to be sufficient in all cases. Also note that exact solutions are hardly efficient in high dimensions and/or non-metric spaces. Hence, the main focus is on approximate methods.

NMSLIB is an extendible library, which means that is possible to add new search methods and distance functions. NMSLIB can be used directly in C++ and Python (via Python bindings). In addition, it is also possible to build a query server, which can be used from Java (or other languages supported by Apache Thrift). Java has a native client, i.e., it works on many platforms without requiring a C++ library to be installed.

Main developers : Bilegsaikhan Naidan, Leonid Boytsov, Yury Malkov, David Novak, Ben Frederickson.

Other contributors: Lawrence Cayton, Wei Dong, Avrelin Nikita, Dmitry Yashunin, Bob Poekert, @orgoro, Maxim Andreev, Daniel Lemire, Nathan Kurz, Alexander Ponomarenko.

Citing: If you find this library useful, feel free to cite our SISAP paper [BibTex] as well as other papers listed in the end. One crucial contribution to cite is the fast Hierarchical Navigable World graph (HNSW) method [BibTex]. Please, also check out the stand-alone HNSW implementation by Yury Malkov, which is released as a header-only HNSWLib library.

Leo(nid) Boytsov is a maintainer. Leo was supported by the Open Advancement of Question Answering Systems (OAQA) group and the following NSF grant #1618159: "Matching and Ranking via Proximity Graphs: Applications to Question Answering and Beyond". Bileg was supported by the iAd Center.

Should you decide to modify the library (and, perhaps, create a pull request), please, use the develoment branch. For generic questions/inquiries, please, use Gitter (see the badge above). Bug reports should be submitted as GitHub issues.

NMSLIB is generic yet fast!

Even though our methods are generic (see e.g., evaluation results in Naidan and Boytsov 2015), they often outperform specialized methods for the Euclidean and/or angular distance (i.e., for the cosine similarity). Below are the results (as of May 2016) of NMSLIB compared to the best implementations participated in a public evaluation code-named ann-benchmarks. Our main competitors are:

  1. A popular library Annoy, which uses a forest of trees (older version used random-projection trees, the new one seems to use a hierarchical 2-means).
  2. A new library FALCONN, which is a highly-optimized implementation of the multiprobe LSH. It uses a novel type of random projections based on the fast Hadamard transform.

The benchmarks were run on a c4.2xlarge instance on EC2 using a previously unseen subset of 5K queries. The benchmarks employ the following data sets:

  1. GloVe : 1.2M 100-dimensional word embeddings trained on Tweets
  2. 1M of 128-dimensional SIFT features

As of May 2016 results are:

1.19M 100d GloVe, cosine similarity. 1M 128d SIFT features, Euclidean distance:

What's new in version 1.6 (see this page for more details )

  1. Improved portability (Can now be built on MACOS)
  2. Easier build: core NMSLIB has no dependencies
  3. Improved Python bindings: dense, sparse, and generic bindings are now in the single module! We also have batch addition and querying functions.
  4. New baselines, including FALCONN library
  5. New spaces (Renyi-divergence, alpha-beta divergence, sparse inner product)
  6. We changed the semantics of boolean command line options: they now have to accept a numerical value (0 or 1).

General information

A detailed description is given in the manual. The manual also contains instructions for building under Linux and Windows, extending the library, as well as for debugging the code using Eclipse. Note that the manual is not fully updated to reflect 1.6 changes. Also note that the manual targets primiarily developers who will extend the library. For most other folks, Python binding docs should be sufficient.

Most of this code is released under the Apache License Version 2.0

  • The LSHKIT, which is embedded in our library, is distributed under the GNU General Public License, see
  • The k-NN graph construction algorithm NN-Descent due to Dong et al. 2011 (see the links below), which is also embedded in our library, seems to be covered by a free-to-use license, similar to Apache 2.
  • FALCONN library's licence is MIT.


  1. A modern compiler that supports C++11: G++ 4.7, Intel compiler 14, Clang 3.4, or Visual Studio 14 (version 12 can probably be used as well, but the project files need to be downgraded).
  2. 64-bit Linux is recommended, but most of our code builds on 64-bit Windows and MACOS as well.
  3. Only for Linux/MACOS: CMake (GNU make is also required)
  4. An Intel or AMD processor that supports SSE 4.2 is recommended
  5. Extended version of the library requires a development version of the following libraries: Boost, GNU scientific library, and Eigen3.

To install additional prerequisite packages on Ubuntu, type the following

sudo apt-get install libboost-all-dev libgsl0-dev libeigen3-dev


  1. Currently only static data sets are supported
  2. HNSW currently duplicates memory to create optimized indices
  3. Range/threshold search is not supported by many methods including SW-graph/HNSW

We plan to resolve these issues in the future.

Quick start on Linux

To compile, go to the directory similarity_search and type:

cmake .

To build an extended version (need extra library):

cmake . -DWITH_EXTRAS=1

You can also download almost every data set used in our previous evaluations (see the section Data sets below). The downloaded data needs to be decompressed (you may need 7z, gzip, and bzip2). Old experimental scripts can be found in the directory previous_releases_scripts. However, they will work only with previous releases.

Note that the benchmarking utility supports caching of ground truth data, so that ground truth data is not recomputed every time this utility is re-run on the same data set.

Query server (Linux-only)

The query server requires Apache Thrift. We used Apache Thrift 0.9.2, but, perhaps, newer versions will work as well.
To install Apache Thrift, you need to build it from source. This may require additional libraries. On Ubuntu they can be installed as follows:

sudo apt-get install libboost-dev libboost-test-dev libboost-program-options-dev libboost-system-dev libboost-filesystem-dev libevent-dev automake libtool flex bison pkg-config g++ libssl-dev libboost-thread-dev make

After Apache Thrift is installed, you need to build the library itself. Then, change the directory to query_server/cpp_client_server and type make (the makefile may need to be modified, if Apache Thrift is installed to a non-standard location). The query server has a similar set of parameters to the benchmarking utility experiment. For example, you can start the server as follows:

 ./query_server -i ../../sample_data/final8_10K.txt -s l2 -m sw-graph -c NN=10,efConstruction=200,initIndexAttempts=1 -p 10000

There are also three sample clients implemented in C++, Python, and Java. A client reads a string representation of a query object from the standard stream. The format is the same as the format of objects in a data file. Here is an example of searching for ten vectors closest to the first data set vector (stored in row one) of a provided sample data file:

export DATA_FILE=../../sample_data/final8_10K.txt
head -1 $DATA_FILE | ./query_client -p 10000 -a localhost  -k 10

It is also possible to generate client classes for other languages supported by Thrift from the interface definition file, e.g., for C#. To this end, one should invoke the thrift compiler as follows:

thrift --gen csharp  protocol.thrift

For instructions on using generated code, please consult the Apache Thrift tutorial.

Python bindings

We provide Python bindings for Python 2.7+ and Python 3.5+, which have been tested under Linux, OSX and Windows. To install:

pip install nmslib

For examples of using the Python API, please, see the README in the python_bindings folder. More detailed documentation is also available (thanks to Ben Frederickson).

Quick start on Windows

Building on Windows requires Visual Studio 2015 Express for Desktop and CMake for Windows. First, generate Visual Studio solution file for 64 bit architecture using CMake GUI. You have to specify both the platform and the version of Visual Studio. Then, the generated solution can be built using Visual Studio. Attention: this way of building on Windows is not well tested yet. We suspect that there might be some issues related to building truly 64-bit binaries.

Data sets

We use several data sets, which were created either by other folks, or using 3d party software. If you use these data sets, please, consider giving proper credit. The download scripts prints respective BibTex entries. More information can be found in the manual.

Here is the list of scripts to download major data sets:

The downloaded data needs to be decompressed (you may need 7z, gzip, and bzip2)

Related publications

Most important related papers are listed below in the chronological order: