Interview on Baidu's Open‑Source Large‑Scale Vector Search Engine Puck

Recently, Baidu announced the open‑source release of its self‑developed vector search engine Puck under the Apache 2.0 license, marking the first domestic open‑source engine designed for ultra‑large‑scale datasets. Vector search algorithms are crucial in personalized recommendation, multimodal retrieval, and natural language processing, especially for handling massive and high‑dimensional data.

The name “Puck” is taken from the intelligent hero in the MOBA game DOTA, symbolizing agility. After years of internal refinement, Puck achieved four first‑place wins in the 2021 BIGANN vector search competition organized by Nerulps. InfoQ interviewed Baidu Search Content Technology Department Chief Architect Ben to discuss the project’s evolution and core strengths.

Ben explained his career at Baidu, from mobile search to co‑founding the multimodal search team, and now leading content‑related technologies such as content acquisition, understanding, computation, and generation.

Motivated by the success of Facebook AI Research’s FAISS, Baidu decided to open‑source Puck to foster community growth, improve code quality, accelerate innovation, and explore new business models.

Puck originated from the visual search business, requiring an ANN engine capable of handling billions of similar‑image queries with high throughput, low latency, high accuracy, low memory usage, and flexibility. Since its first launch in 2017, Puck has expanded beyond image search to semantic retrieval powered by transformer‑based embeddings, and is now used across Baidu’s search, recommendation, cloud storage, and knowledge graph products, supporting trillion‑scale indexes.

The development roadmap can be divided into four phases:

2016‑2019: Core algorithm and implementation optimization, focusing on performance in Baidu’s own scenarios.

2019‑2021: Enhancing usability, extensibility, and feature diversity (e.g., real‑time insertion, multi‑condition search, distributed indexing) as internal open‑source adoption grew.

2021‑2022: Scaling to single‑instance ultra‑large datasets through large‑scale quantization and index structure improvements, demonstrated by winning four categories in the BIGANN competition.

2022‑present: Introducing new algorithms for varied data scenarios, adding features, and preparing for external open‑source release.

Ben highlighted three core advantages of Puck: superior performance across benchmarks ranging from millions to billions of vectors, ease of use via simple APIs with sensible defaults, and proven reliability from years of production deployment supporting over a trillion‑level index and massive query traffic.

The open‑source release includes two Baidu‑developed retrieval algorithms—Puck (for ultra‑large datasets) and Tinker (for medium‑size datasets)—covering most retrieval scenarios.

Regarding the future of open‑source in the AI era, Ben believes that while competition will intensify, it will drive innovation, lower R&D costs, and accelerate AI progress. He expects open‑source projects to become more specialized, cross‑domain, practical, and focused on data and algorithm sharing.

In practice, Puck powers Baidu Search and the information‑flow recommendation in the mobile Baidu app. Community feedback after open‑sourcing has been valuable, revealing bugs, suggesting improvements, and fostering collaborative development.

Ben expressed strong personal commitment to Puck’s continued evolution, hoping for broader community adoption, ongoing algorithmic innovation, and deeper impact on AI applications across industries.