Starry Vector Retrieval Platform: Architecture, Features, and Performance

Background and Challenges – With the rapid development of deep learning, unstructured data such as images, videos, audio, and user behavior are represented as vectors, creating massive K‑nearest‑neighbor search demands across Kuaishou’s internal services (search, recommendation, safety, content understanding). Open‑source libraries like Faiss, HNSW, and ScaNN provide a foundation, but gaps remain in performance, usability, ultra‑large‑scale support, complex retrieval scenarios, algorithm hyper‑parameter tuning, and rapid onboarding.

Platform Goal – The "Starry" vector retrieval platform (繁星) was built jointly by the Search Technology and AI Platform teams to address these challenges, offering a high‑performance, highly reliable, and user‑friendly service that supports both generic and specialized business needs.

Architecture Overview – The platform adopts a layered, low‑coupling design that abstracts algorithm optimization, feature enrichment, high concurrency, and ease of use. It provides independent services at each layer, enabling rapid iteration and component reuse. The architecture includes online and offline indexing pipelines, a proxy layer for request routing, and a Lambda‑style big‑data processing backbone for batch and real‑time updates.

Key Feature 1: Algorithm Diversity and Advancement – Starry integrates multiple state‑of‑the‑art ANN algorithms, including HNSW, NSSG, ScaNN (Google), Faiss, GPU‑accelerated solutions (SONG, Faiss‑GPU), and several proprietary algorithms such as IMI‑PQ, GPU‑KNN, and multi‑objective GPU‑KNN. Continuous R&D improves throughput and reduces hardware costs.

Algorithm Optimizations – Specific enhancements to HNSW and ScaNN (vector compression, SIMD acceleration, graph re‑ordering, cache and read‑only optimizations) yield 25‑36% performance gains. MIPS (Maximum Inner Product Search) techniques are refined using ip‑NSW and anisotropic quantization, achieving significant speed‑up over baseline methods.

Key Feature 2: Rich Engine Capabilities – The Starry engine abstracts common business needs, offering structured query support (combining vector similarity with Boolean filters), multiple index types (vector, hash, B‑tree), data unit management (buckets, versions, segments), a trainer‑worker near‑real‑time indexing framework, and the Hive2Hive offline large‑scale index query tool for batch K‑NN tasks.

Key Feature 3: Distributed Scalable Indexing & Online Retrieval – Leveraging a Lambda architecture, the platform provides second‑level index availability, supports billions of vectors, separates static and dynamic indexes for read‑write isolation, and implements a proxy layer with asynchronous protocols and Brpc optimizations to handle millions of QPS with minimal resources.

Key Feature 4: Automated Deployment – Built on the Starry Platform Web Server, the system integrates with internal infrastructure (container cloud, KwaiFlow, Kconf, Btqueue) to enable one‑click index construction and service deployment, reducing onboarding time to minutes.

Future Roadmap – Ongoing work includes deeper integration with the company’s AI platform, continual algorithm research, solving ultra‑large‑scale indexing challenges, automated hyper‑parameter tuning, and extending vector retrieval to multi‑objective, semantic, and multimodal scenarios such as product photo search.