Convert Figma logo to code with AI

spotify logoannoy

Approximate Nearest Neighbors in C++/Python optimized for memory usage and loading/saving to disk

13,073
1,149
13,073
59

Top Related Projects

2,552

A Scala API for Apache Beam and Google Cloud Dataflow.

Benchmarks of approximate nearest neighbor libraries in Python

30,390

A library for efficient similarity search and clustering of dense vectors.

3,373

Non-Metric Space Library (NMSLIB): An efficient similarity search library and a toolkit for evaluation of k-NN methods for generic non-metric spaces.

7,343

Uniform Manifold Approximation and Projection

29,291

A cloud-native vector database, storage for next generation AI applications

Quick Overview

Spotify's Annoy (Approximate Nearest Neighbors Oh Yeah) is a C++ library with Python bindings for searching for nearest neighbors in high-dimensional spaces. It's optimized for memory usage and fast index creation, making it suitable for large-scale recommendation systems and similarity search applications.

Pros

  • Fast index creation and querying, especially for high-dimensional data
  • Memory-efficient, with the ability to load indices from disk
  • Supports both Euclidean distance and angular (cosine) distance metrics
  • Easy-to-use Python interface

Cons

  • Limited to static datasets; doesn't support online updates
  • May sacrifice some accuracy for speed compared to exact nearest neighbor methods
  • Only supports L2 and angular distance metrics
  • Primarily designed for read-heavy workloads

Code Examples

  1. Creating and building an index:
from annoy import AnnoyIndex

dim = 100
t = AnnoyIndex(dim, 'angular')
for i in range(1000):
    v = [random.gauss(0, 1) for z in range(dim)]
    t.add_item(i, v)

t.build(10)  # 10 trees
t.save('test.ann')
  1. Querying the index:
u = AnnoyIndex(dim, 'angular')
u.load('test.ann')
results = u.get_nns_by_vector([1.0] * dim, 10)  # Return 10 nearest neighbors
  1. Using memory mapping for faster loading:
u = AnnoyIndex(dim, 'angular')
u.load('test.ann', prefault=True)  # Prefault pages into memory

Getting Started

To use Annoy, first install it via pip:

pip install annoy

Then, in your Python script:

from annoy import AnnoyIndex

# Create an index
dim = 100
index = AnnoyIndex(dim, 'angular')

# Add items to the index
for i in range(1000):
    vector = [random.random() for _ in range(dim)]
    index.add_item(i, vector)

# Build the index
index.build(10)

# Search for nearest neighbors
results = index.get_nns_by_vector([0.5] * dim, 10)
print(results)

This example creates an index with 1000 random vectors, builds it with 10 trees, and then searches for the 10 nearest neighbors of a query vector.

Competitor Comparisons

2,552

A Scala API for Apache Beam and Google Cloud Dataflow.

Pros of Scio

  • Designed for large-scale data processing with Apache Beam
  • Supports both batch and streaming data processing
  • Offers a Scala API for more expressive and type-safe code

Cons of Scio

  • Steeper learning curve due to Scala and Apache Beam concepts
  • More complex setup and configuration compared to Annoy
  • Limited to data processing tasks, not optimized for nearest neighbor search

Code Comparison

Scio (data processing example):

sc.textFile("input.txt")
  .flatMap(_.split(" "))
  .countByValue
  .map(kv => s"${kv._1}: ${kv._2}")
  .saveAsTextFile("output")

Annoy (nearest neighbor search example):

t = AnnoyIndex(f, 'angular')
for i in range(1000):
    v = [random.gauss(0, 1) for z in range(f)]
    t.add_item(i, v)
t.build(10)

Key Differences

  • Scio is a data processing framework, while Annoy is a library for approximate nearest neighbor search
  • Scio is written in Scala and uses Apache Beam, Annoy is written in C++ with Python bindings
  • Scio is more suitable for complex data pipelines, while Annoy excels in high-dimensional vector search tasks

Benchmarks of approximate nearest neighbor libraries in Python

Pros of ann-benchmarks

  • Comprehensive benchmarking suite for various ANN algorithms
  • Regularly updated with new algorithms and datasets
  • Provides visualizations and comparisons across multiple metrics

Cons of ann-benchmarks

  • More complex to set up and use compared to Annoy
  • Requires more computational resources for running benchmarks
  • Not optimized for production use, primarily for evaluation purposes

Code comparison

ann-benchmarks:

import ann_benchmarks
from ann_benchmarks.algorithms.annoy import Annoy

dataset = ann_benchmarks.datasets.random(1000, 100)
algo = Annoy(metric='angular', n_trees=10)
algo.fit(dataset)

Annoy:

from annoy import AnnoyIndex

f = 100
t = AnnoyIndex(f, 'angular')
for i in range(1000):
    v = [random.gauss(0, 1) for z in range(f)]
    t.add_item(i, v)
t.build(10)

Both repositories serve different purposes. ann-benchmarks is a comprehensive benchmarking tool for comparing various ANN algorithms, including Annoy. It's ideal for researchers and developers looking to evaluate and compare different ANN implementations. Annoy, on the other hand, is a specific ANN library optimized for production use, offering a simpler API and faster query times for specific use cases.

30,390

A library for efficient similarity search and clustering of dense vectors.

Pros of Faiss

  • Supports GPU acceleration for faster processing
  • Offers a wider range of indexing algorithms and distance metrics
  • Scales better for very large datasets (billions of vectors)

Cons of Faiss

  • Steeper learning curve and more complex API
  • Requires more memory and computational resources
  • Less suitable for small to medium-sized datasets

Code Comparison

Annoy:

from annoy import AnnoyIndex

f = 40
t = AnnoyIndex(f, 'angular')
for i in range(1000):
    v = [random.gauss(0, 1) for z in range(f)]
    t.add_item(i, v)
t.build(10)

Faiss:

import faiss
import numpy as np

d = 64
nb = 100000
xb = np.random.random((nb, d)).astype('float32')
index = faiss.IndexFlatL2(d)
index.add(xb)

Both Annoy and Faiss are popular libraries for approximate nearest neighbor search, but they cater to different use cases. Annoy is simpler to use and works well for smaller datasets, while Faiss offers more advanced features and better performance for large-scale applications. The choice between them depends on the specific requirements of your project, such as dataset size, performance needs, and available computational resources.

3,373

Non-Metric Space Library (NMSLIB): An efficient similarity search library and a toolkit for evaluation of k-NN methods for generic non-metric spaces.

Pros of nmslib

  • Supports a wider range of distance metrics and index types
  • Generally faster for high-dimensional data and large datasets
  • More flexible and customizable for advanced use cases

Cons of nmslib

  • Steeper learning curve and more complex API
  • Less memory-efficient for some use cases
  • Requires more configuration and tuning for optimal performance

Code Comparison

nmslib:

import nmslib
index = nmslib.init(method='hnsw', space='cosine')
index.addDataPointBatch(data)
index.createIndex({'post': 2})
ids, distances = index.knnQuery(query, k=10)

Annoy:

from annoy import AnnoyIndex
index = AnnoyIndex(f, 'angular')
for i, v in enumerate(data):
    index.add_item(i, v)
index.build(10)
ids = index.get_nns_by_vector(query, 10)

Both nmslib and Annoy are popular libraries for approximate nearest neighbor search, but they cater to different use cases. nmslib offers more advanced features and better performance for high-dimensional data, while Annoy provides a simpler API and is more memory-efficient for certain scenarios. The choice between the two depends on the specific requirements of your project, such as dataset size, dimensionality, and desired trade-offs between speed, accuracy, and memory usage.

7,343

Uniform Manifold Approximation and Projection

Pros of UMAP

  • More effective for dimensionality reduction and visualization
  • Better preservation of global structure in data
  • Faster runtime for large datasets

Cons of UMAP

  • More complex algorithm, potentially harder to understand and implement
  • May require more parameter tuning for optimal results
  • Less suitable for exact nearest neighbor search

Code Comparison

UMAP example:

import umap
reducer = umap.UMAP()
embedding = reducer.fit_transform(data)

Annoy example:

from annoy import AnnoyIndex
t = AnnoyIndex(f, 'angular')
for i, v in enumerate(data):
    t.add_item(i, v)
t.build(10)

UMAP is primarily used for dimensionality reduction and visualization, while Annoy focuses on approximate nearest neighbor search. UMAP offers better preservation of global structure and faster runtime for large datasets, but may require more parameter tuning. Annoy is simpler to implement and use, especially for exact nearest neighbor search, but may not perform as well for dimensionality reduction tasks.

Both libraries have their strengths and are suited for different use cases. UMAP is preferred for complex data visualization and dimensionality reduction, while Annoy excels in efficient approximate nearest neighbor search for large datasets.

29,291

A cloud-native vector database, storage for next generation AI applications

Pros of Milvus

  • Scalable and distributed architecture for handling large-scale vector datasets
  • Supports multiple index types and similarity metrics
  • Offers both standalone and cluster deployment options

Cons of Milvus

  • More complex setup and configuration compared to Annoy
  • Requires more system resources and infrastructure
  • Steeper learning curve for beginners

Code Comparison

Annoy (Python):

from annoy import AnnoyIndex

f = 40
t = AnnoyIndex(f, 'angular')
for i in range(1000):
    v = [random.gauss(0, 1) for z in range(f)]
    t.add_item(i, v)
t.build(10)

Milvus (Python):

from milvus import Milvus, IndexType, MetricType

client = Milvus(host='localhost', port='19530')
collection_name = 'example_collection'
dimension = 128
client.create_collection({'collection_name': collection_name, 'dimension': dimension})
client.create_index(collection_name, IndexType.IVF_FLAT, {'nlist': 16384})

Annoy is simpler to use and integrate, making it ideal for smaller-scale applications or quick prototyping. Milvus, on the other hand, offers more advanced features and scalability, suitable for large-scale production environments with complex vector search requirements. Annoy is better for in-memory searches, while Milvus excels in distributed scenarios with massive datasets.

Convert Figma logo designs to code with AI

Visual Copilot

Introducing Visual Copilot: A new AI model to turn Figma designs to high quality code using your components.

Try Visual Copilot

README

Annoy

.. figure:: https://raw.github.com/spotify/annoy/master/ann.png :alt: Annoy example :align: center

.. image:: https://github.com/spotify/annoy/actions/workflows/ci.yml/badge.svg :target: https://github.com/spotify/annoy/actions

Annoy (Approximate Nearest Neighbors <http://en.wikipedia.org/wiki/Nearest_neighbor_search#Approximate_nearest_neighbor>__ Oh Yeah) is a C++ library with Python bindings to search for points in space that are close to a given query point. It also creates large read-only file-based data structures that are mmapped <https://en.wikipedia.org/wiki/Mmap>__ into memory so that many processes may share the same data.

Install

To install, simply do pip install --user annoy to pull down the latest version from PyPI <https://pypi.python.org/pypi/annoy>_.

For the C++ version, just clone the repo and #include "annoylib.h".

Background

There are some other libraries to do nearest neighbor search. Annoy is almost as fast as the fastest libraries, (see below), but there is actually another feature that really sets Annoy apart: it has the ability to use static files as indexes. In particular, this means you can share index across processes. Annoy also decouples creating indexes from loading them, so you can pass around indexes as files and map them into memory quickly. Another nice thing of Annoy is that it tries to minimize memory footprint so the indexes are quite small.

Why is this useful? If you want to find nearest neighbors and you have many CPU's, you only need to build the index once. You can also pass around and distribute static files to use in production environment, in Hadoop jobs, etc. Any process will be able to load (mmap) the index into memory and will be able to do lookups immediately.

We use it at Spotify <http://www.spotify.com/>__ for music recommendations. After running matrix factorization algorithms, every user/item can be represented as a vector in f-dimensional space. This library helps us search for similar users/items. We have many millions of tracks in a high-dimensional space, so memory usage is a prime concern.

Annoy was built by Erik Bernhardsson <http://www.erikbern.com>__ in a couple of afternoons during Hack Week <http://labs.spotify.com/2013/02/15/organizing-a-hack-week/>__.

Summary of features

  • Euclidean distance <https://en.wikipedia.org/wiki/Euclidean_distance>, Manhattan distance <https://en.wikipedia.org/wiki/Taxicab_geometry>, cosine distance <https://en.wikipedia.org/wiki/Cosine_similarity>, Hamming distance <https://en.wikipedia.org/wiki/Hamming_distance>, or Dot (Inner) Product distance <https://en.wikipedia.org/wiki/Dot_product>__
  • Cosine distance is equivalent to Euclidean distance of normalized vectors = sqrt(2-2*cos(u, v))
  • Works better if you don't have too many dimensions (like <100) but seems to perform surprisingly well even up to 1,000 dimensions
  • Small memory usage
  • Lets you share memory between multiple processes
  • Index creation is separate from lookup (in particular you can not add more items once the tree has been created)
  • Native Python support, tested with 2.7, 3.6, and 3.7.
  • Build index on disk to enable indexing big datasets that won't fit into memory (contributed by Rene Hollander <https://github.com/ReneHollander>__)

Python code example

.. code-block:: python

from annoy import AnnoyIndex import random

f = 40 # Length of item vector that will be indexed

t = AnnoyIndex(f, 'angular') for i in range(1000): v = [random.gauss(0, 1) for z in range(f)] t.add_item(i, v)

t.build(10) # 10 trees t.save('test.ann')

...

u = AnnoyIndex(f, 'angular') u.load('test.ann') # super fast, will just mmap the file print(u.get_nns_by_item(0, 1000)) # will find the 1000 nearest neighbors

Right now it only accepts integers as identifiers for items. Note that it will allocate memory for max(id)+1 items because it assumes your items are numbered 0 … n-1. If you need other id's, you will have to keep track of a map yourself.

Full Python API

  • AnnoyIndex(f, metric) returns a new index that's read-write and stores vector of f dimensions. Metric can be "angular", "euclidean", "manhattan", "hamming", or "dot".
  • a.add_item(i, v) adds item i (any nonnegative integer) with vector v. Note that it will allocate memory for max(i)+1 items.
  • a.build(n_trees, n_jobs=-1) builds a forest of n_trees trees. More trees gives higher precision when querying. After calling build, no more items can be added. n_jobs specifies the number of threads used to build the trees. n_jobs=-1 uses all available CPU cores.
  • a.save(fn, prefault=False) saves the index to disk and loads it (see next function). After saving, no more items can be added.
  • a.load(fn, prefault=False) loads (mmaps) an index from disk. If prefault is set to True, it will pre-read the entire file into memory (using mmap with MAP_POPULATE). Default is False.
  • a.unload() unloads.
  • a.get_nns_by_item(i, n, search_k=-1, include_distances=False) returns the n closest items. During the query it will inspect up to search_k nodes which defaults to n_trees * n if not provided. search_k gives you a run-time tradeoff between better accuracy and speed. If you set include_distances to True, it will return a 2 element tuple with two lists in it: the second one containing all corresponding distances.
  • a.get_nns_by_vector(v, n, search_k=-1, include_distances=False) same but query by vector v.
  • a.get_item_vector(i) returns the vector for item i that was previously added.
  • a.get_distance(i, j) returns the distance between items i and j. NOTE: this used to return the squared distance, but has been changed as of Aug 2016.
  • a.get_n_items() returns the number of items in the index.
  • a.get_n_trees() returns the number of trees in the index.
  • a.on_disk_build(fn) prepares annoy to build the index in the specified file instead of RAM (execute before adding items, no need to save after build)
  • a.set_seed(seed) will initialize the random number generator with the given seed. Only used for building up the tree, i. e. only necessary to pass this before adding the items. Will have no effect after calling a.build(n_trees) or a.load(fn).

Notes:

  • There's no bounds checking performed on the values so be careful.
  • Annoy uses Euclidean distance of normalized vectors for its angular distance, which for two vectors u,v is equal to sqrt(2(1-cos(u,v)))

The C++ API is very similar: just #include "annoylib.h" to get access to it.

Tradeoffs

There are just two main parameters needed to tune Annoy: the number of trees n_trees and the number of nodes to inspect during searching search_k.

  • n_trees is provided during build time and affects the build time and the index size. A larger value will give more accurate results, but larger indexes.
  • search_k is provided in runtime and affects the search performance. A larger value will give more accurate results, but will take longer time to return.

If search_k is not provided, it will default to n * n_trees where n is the number of approximate nearest neighbors. Otherwise, search_k and n_trees are roughly independent, i.e. the value of n_trees will not affect search time if search_k is held constant and vice versa. Basically it's recommended to set n_trees as large as possible given the amount of memory you can afford, and it's recommended to set search_k as large as possible given the time constraints you have for the queries.

You can also accept slower search times in favour of reduced loading times, memory usage, and disk IO. On supported platforms the index is prefaulted during load and save, causing the file to be pre-emptively read from disk into memory. If you set prefault to False, pages of the mmapped index are instead read from disk and cached in memory on-demand, as necessary for a search to complete. This can significantly increase early search times but may be better suited for systems with low memory compared to index size, when few queries are executed against a loaded index, and/or when large areas of the index are unlikely to be relevant to search queries.

How does it work

Using random projections <http://en.wikipedia.org/wiki/Locality-sensitive_hashing#Random_projection>__ and by building up a tree. At every intermediate node in the tree, a random hyperplane is chosen, which divides the space into two subspaces. This hyperplane is chosen by sampling two points from the subset and taking the hyperplane equidistant from them.

We do this k times so that we get a forest of trees. k has to be tuned to your need, by looking at what tradeoff you have between precision and performance.

Hamming distance (contributed by Martin Aumüller <https://github.com/maumueller>__) packs the data into 64-bit integers under the hood and uses built-in bit count primitives so it could be quite fast. All splits are axis-aligned.

Dot Product distance (contributed by Peter Sobot <https://github.com/psobot>__ and Pavel Korobov <https://github.com/pkorobov>) reduces the provided vectors from dot (or "inner-product") space to a more query-friendly cosine space using a method by Bachrach et al., at Microsoft Research, published in 2014 <https://www.microsoft.com/en-us/research/wp-content/uploads/2016/02/XboxInnerProduct.pdf>.

More info

  • Dirk Eddelbuettel <https://github.com/eddelbuettel>__ provides an R version of Annoy <http://dirk.eddelbuettel.com/code/rcpp.annoy.html>__.
  • Andy Sloane <https://github.com/a1k0n>__ provides a Java version of Annoy <https://github.com/spotify/annoy-java>__ although currently limited to cosine and read-only.
  • Pishen Tsai <https://github.com/pishen>__ provides a Scala wrapper of Annoy <https://github.com/pishen/annoy4s>__ which uses JNA to call the C++ library of Annoy.
  • Atsushi Tatsuma <https://github.com/yoshoku>__ provides Ruby bindings for Annoy <https://github.com/yoshoku/annoy.rb>__.
  • There is experimental support for Go <https://github.com/spotify/annoy/blob/master/README_GO.rst>__ provided by Taneli Leppä <https://github.com/rosmo>__.
  • Boris Nagaev <https://github.com/starius>__ wrote Lua bindings <https://github.com/spotify/annoy/blob/master/README_Lua.md>__.
  • During part of Spotify Hack Week 2016 (and a bit afterward), Jim Kang <https://github.com/jimkang>__ wrote Node bindings <https://github.com/jimkang/annoy-node>__ for Annoy.
  • Min-Seok Kim <https://github.com/mskimm>__ built a Scala version <https://github.com/mskimm/ann4s>__ of Annoy.
  • hanabi1224 <https://github.com/hanabi1224>__ built a read-only Rust version <https://github.com/hanabi1224/RuAnnoy>__ of Annoy, together with dotnet, jvm and dart read-only bindings.
  • Presentation from New York Machine Learning meetup <http://www.slideshare.net/erikbern/approximate-nearest-neighbor-methods-and-vector-models-nyc-ml-meetup>__ about Annoy
  • Annoy is available as a conda package <https://anaconda.org/conda-forge/python-annoy>__ on Linux, OS X, and Windows.
  • ann-benchmarks <https://github.com/erikbern/ann-benchmarks>__ is a benchmark for several approximate nearest neighbor libraries. Annoy seems to be fairly competitive, especially at higher precisions:

.. figure:: https://github.com/erikbern/ann-benchmarks/raw/master/results/glove-100-angular.png :alt: ANN benchmarks :align: center :target: https://github.com/erikbern/ann-benchmarks

Source code

It's all written in C++ with a handful of ugly optimizations for performance and memory usage. You have been warned :)

The code should support Windows, thanks to Qiang Kou <https://github.com/thirdwing>__ and Timothy Riley <https://github.com/tjrileywisc>__.

To run the tests, execute python setup.py nosetests. The test suite includes a big real world dataset that is downloaded from the internet, so it will take a few minutes to execute.

Discuss

Feel free to post any questions or comments to the annoy-user <https://groups.google.com/group/annoy-user>__ group. I'm @fulhack <https://twitter.com/fulhack>__ on Twitter.