puck

Source: https://github.com/baidu/puck

Category: open-sources

Tags: open-source, vector-search, similarity-search, embedding

---

Description

Puck is an open-source, high-performance vector search engine designed for fast similarity search and retrieval of embedding vectors. It is intended for large-scale industrial applications where memory constraints, computational resources, and database size are critical factors.

---

Features

• Approximate Nearest Neighbor (ANN) Search: Supports fast and efficient similarity search.
• Two Algorithms:
  • Puck: Optimized for large-scale datasets, with memory efficiency and high recall-vs-latency performance. Uses a two-layered architectural design for inverted indices and multi-level quantization.
  • Tinker: Designed for smaller datasets (e.g., 10M, 100M). Offers better performance than Nmslib in benchmarks but uses more memory than Puck.
• Written in C++: Provides Python 3 wrappers for integration with Python projects.
• Similarity Metrics Supported:
  • Cosine similarity
  • L2 (Euclidean) distance
  • Inner Product (IP, with transformation to cosine distance)
• Memory Efficiency:
  • Puck uses compressed vectors (after product quantization), reducing memory usage to about 1/4 of the original size by default.
  • Tinker requires more memory to store similarity point relationships.
• Benchmark Results:
  • Puck demonstrated top performance on multiple 1B-datasets in NeurIPS'21 competition track.
  • Performance improvements since initial release (up to 70% increase).
• Flexible Configuration:
  • Configurable via files for training, building, and searching.
  • Supports different vector formats (.fvecs) and raw little endian storage.
• Build & Deployment:
  • Requires MKL, Python 3.6+, and CMake 3.21+ for building.
  • Includes demos and tools for training, building, and searching.
• Benchmark Tools:
  • Includes scripts and configs for benchmarking against other ANN libraries (e.g., Faiss, Nmslib).

---

Pricing

Puck is open-source and available for free under its respective license.

---

Installation & Usage

• Requires Intel MKL, Python 3.6+, CMake 3.21+.
• Build instructions and demos provided in the repository.
• Tools and configs are included for formatting datasets, training, building indices, and performing searches.

---

Additional Resources

• Documentation and more details
• Benchmark details