Annoy : Scalable similarity search for embeddings

Annoy (Approximate Nearest Neighbors Oh Yeah) is an open-source C++ library developed by Spotify for approximate nearest neighbor (ANN) search in high-dimensional spaces. Optimized for read-heavy workloads, Annoy is designed to quickly search large sets of static vectors using efficient tree-based indexing, making it a popular choice for recommendation engines, music similarity, content-based filtering, and semantic search.

Annoy is particularly useful when you have a large number of embeddings that rarely change and require low-latency querying. It builds indexes that can be saved to disk and memory-mapped for efficient loading and querying in production environments.

Key benefits include:

• Extremely fast read performance with low memory overhead
• On-disk indexes for efficient loading and sharing across processes
• Minimal dependencies and easy to use in Python or C++

What are the main features of Annoy?

Approximate nearest neighbor (ANN) search

Annoy implements fast ANN search using multiple random projection trees.

• Efficient for high-dimensional vector spaces
• Supports k-nearest neighbor (k-NN) queries
• Works well with metrics like angular (cosine), Euclidean, Manhattan, and Hamming distance

Disk-based index and memory mapping

Annoy builds read-only indexes that are saved to disk, making them ideal for production.

• Indexes can be memory-mapped for low-latency access
• Enables multiple processes to share the same index without duplication
• Especially suited for read-heavy workloads and static datasets

Lightweight and dependency-free

Annoy is written in C++ with Python bindings, and has no external dependencies.

• Simple to compile and integrate
• Python interface is intuitive and widely used in ML pipelines
• Easily embeddable in applications with limited resource environments

Support for multiple distance metrics

Annoy supports several distance functions to match different use cases.

• Angular (cosine similarity)
• Euclidean (L2)
• Manhattan (L1)
• Hamming (for binary vectors)

Scales well for large static datasets

Annoy is optimized for use cases with many vectors that don’t change frequently.

• Can handle millions of high-dimensional vectors
• Performance improves with more trees (configurable trade-off between speed and accuracy)
• Good fit for personalized recommendations, image or music similarity, and precomputed vector search

Why choose Annoy?

• Optimized for read-only use: perfect for static embeddings and production serving
• Disk-efficient: builds indexes that are fast to load and share
• Simple and portable: lightweight C++ core with easy Python access
• Multi-metric support: handles various distance functions out of the box
• Proven at scale: used by Spotify and others for real-world recommendation systems