---
id: pynndescent
version: "0.6.0"
license: BSD-2-Clause
license_treatment: permissive
maintenance: active
---

# pynndescent — Nearest Neighbor Descent

License: permissive · Maintenance: active · Downloads: 7.0M/mo

## What it is and what it does

PyNNDescent is a Python library for building approximate nearest neighbor (ANN) search indexes using the nearest neighbor descent algorithm, supplemented with random projection trees for initialization. It supports a large variety of distance metrics including euclidean, manhattan, cosine, hamming, wasserstein, and many others, as well as custom user-defined metrics.

The library provides a simple two-operation interface: construct an index from training data, then query it for the k nearest neighbors of new points. It integrates with scikit-learn and can serve as a drop-in replacement for KNeighborTransformer in algorithms that use nearest neighbor computations. The implementation relies on numba for JIT compilation and joblib for parallelization.

Use it for:

- Build a searchable index for similarity-based recommendation systems or content retrieval
- Perform fast approximate k-nearest neighbor queries on high-dimensional embeddings or feature vectors
- Integrate nearest neighbor search into scikit-learn pipelines as a preprocessing or feature transformation step
- Find similar items across different distance metrics (e.g., cosine for text embeddings, euclidean for image features)
- Construct k-neighbor graphs for clustering, graph-based learning, or dimensionality reduction algorithms

## Worth the install?

AI-flagged interpretation of the facts above — verify before relying.

PyNNDescent builds approximate nearest neighbor search indexes using nearest neighbor descent algorithms, supporting a wide variety of distance metrics for fast k-neighbor queries on high-dimensional data.

Yes. PyNNDescent is actively maintained, has low install friction, carries no known vulnerabilities, and uses a permissive BSD-2-Clause license. It solves a well-defined problem with broad metric support and scikit-learn integration. Suitable for production use in recommendation systems, similarity search, and machine learning pipelines.

## Install

pip install pynndescent
uv add pynndescent
poetry add pynndescent

## Installing pynndescent

Before you install: Low friction: pure Python wheel with five runtime dependencies (scikit-learn, scipy, numba, llvmlite, joblib) that are standard in scientific Python. Last release 218 days ago with active repository maintenance and 969 stars.

License in practice: BSD-2-Clause permissive license allows commercial and private use with minimal restrictions; suitable for most projects.

Quickstart:

pip install pynndescent

from pynndescent import NNDescent
index = NNDescent(data)
index.query(query_data, k=15)

Verify before relying:

- Whether the 80%-100% accuracy rate claim applies to all distance metrics or only specific ones
- Performance comparison details with other ANN libraries beyond the ann-benchmarks reference

## Package facts

- License: BSD-2-Clause (permissive)
- Python support: unspecified
- Install friction: low
- Maintenance: active
- Downloads: 7.0M/month (top 5,000 on PyPI)
- Known vulnerabilities: none known

## Tags

approximate nearest neighbor search, knn indexing library, ann algorithm python, nearest neighbor graph construction, fast similarity search, k-nearest neighbors index, distance metric search, approximate-search, knn-indexing, distance-metrics

[View on SkillFed](https://skillfed.io/packages/pynndescent) · [View on PyPI](https://pypi.org/project/pynndescent/)