PUFFINN

Overview

PUFFINN (Parameterless and Universal Fast FInding of Nearest Neighbors) is a parameterless LSH-based index for solving the k-nearest neighbor problem with probabilistic guarantees. It provides an easily configurable library for finding approximate nearest neighbors of arbitrary points.

Key Features

• Parameterless Design: Users only need to specify memory usage and desired result quality (recall), not complex algorithm parameters
• Probabilistic Guarantees: Each near neighbor is guaranteed to be found with the specified recall probability, regardless of query difficulty
• LSH-Based: Uses Locality Sensitive Hashing with an adaptive query mechanism
• Multiple Similarity Measures: Supports Cosine similarity (SimHash/cross-polytope LSH) and Jaccard similarity (MinHash)
• Multi-Language Support: Available in both C++ and Python with feature parity

Technical Approach

Under the hood, PUFFINN uses Locality Sensitive Hashing (LSH) with an adaptive query mechanism. This approach provides:

• Configurable memory budgets
• Adjustable recall levels
• Consistent performance across different query types

Distance Functions

Currently Supported:

• Cosine Similarity: Using SimHash or cross-polytope LSH
• Jaccard Similarity: Using MinHash

Implementation

• Languages: C++ (primary implementation) with Python bindings
• License: Open source
• Platform: Cross-platform support

Use Cases

• Applications requiring guaranteed recall rates
• Similarity search with predictable performance
• Systems where parameter tuning overhead should be minimized
• Cosine similarity search (text embeddings, document similarity)
• Jaccard similarity search (set similarity, recommendation systems)

Academic Background

Published at the 27th Annual European Symposium on Algorithms (ESA 2019)

Authors: Martin Aumüller, Tobias Christiani, Rasmus Pagh, and Michael Vesterli

Paper: Available on arXiv (arXiv:1906.12211)

Benchmarks

PUFFINN has been included in various ANN benchmarking efforts and demonstrates competitive performance with probabilistic quality guarantees.

Resources

• GitHub: https://github.com/puffinn/puffinn
• Publication: ESA 2019 (27th Annual European Symposium on Algorithms)
• : 1906.12211

Advantages

The parameterless design makes PUFFINN particularly attractive for developers who want:

• Simple configuration (just memory and recall)
• Guaranteed search quality
• No need for extensive parameter tuning
• Predictable behavior across different datasets