Slipstream boosts streaming nearest neighbor search 30.8x
New algorithm exploits vector stream continuity for up to 30.8x higher throughput.
Slipstream, developed by Shubing Yang and Dongfang Zhao, is a novel method for streaming approximate nearest neighbor search (ANNS). Traditional graph indexes like HNSW require repeated index construction when new vectors arrive continuously, creating a bottleneck. Slipstream exploits locality in vector streams: instead of searching from the entry point for each new insertion, it starts from promising candidates found during the previous insertion. An adaptive controller narrows or widens the search range based on the stream's stability, reducing computational cost. The authors also provide a theoretical model to characterize performance bounds, moving beyond heuristic approaches.
Implemented as a drop-in enhancement for popular open-source libraries Faiss and HNSWLib, Slipstream was tested against four baselines on five streaming vector datasets. Results show up to 30.8x improvement in end-to-end throughput while maintaining at least 0.95 recall@10. This makes Slipstream highly suitable for real-time applications such as recommendation systems, anomaly detection, and search in continuously updating databases, where high recall and low latency are critical. The approach could become a standard optimization for streaming vector search in production systems.
- Slipstream reduces insertion cost by starting searches from previous candidates, not entry points.
- Achieves up to 30.8x higher throughput than baselines at 0.95 recall@10.
- Implemented in Faiss and HNSWLib with an adaptive controller for stream stability.
Why It Matters
Enables real-time vector search on continuous data streams with minimal latency, crucial for modern AI applications.