Yi: In-place graph-based vector index boosts update throughput 1.75x
New vector database technique handles real-time updates with 1.8x faster search at 73% memory.
Deep Dive
Yi proposes in-place graph-based vector index updates, achieving 1.75x higher update throughput and 1.8x higher concurrent search throughput than state-of-the-art systems on an 800M dataset, using only 73% of peak memory and fewer CPU cores. Its key innovations include a vector-level update mechanism with a tasklet-based execution engine, asynchronous buffer manager, and vector file system.
Key Points
- Yi achieves 1.75x higher update throughput and 1.8x higher concurrent search throughput on an 800M dataset vs SOTA.
- Uses only 73% of peak memory and fewer CPU cores, improving resource efficiency.
- Core innovation: vector-level in-place updates via tasklet execution engine, async buffer manager, and vector file system.
Why It Matters
Real-time vector updates at lower cost enable LLM applications to capture rapidly evolving data without sacrificing search quality.