Fast, Approximate Vector Queries on Very Large Unstructured Datasets
summary
Auncel provides bounded query error and bounded latency for billion-item vector search by building a per-query error-latency profile that predicts how many clusters to sample using local geometric volumes. It calibrates per-shard errors to prevent amplification in distributed merges, implements a C++ prototype integrated with Faiss, and reports up to 10× lower latency for the same ≤10% error bound. On DEEP1B Auncel processes a query in ~25 ms using four c5.metal instances and scales linearly with workers.
AI-written summary · about
chapters (7)
- 0:00Start
- 4:58Bounded Performance: Error Bound
- 9:31Bounded Performance: Distributed Settings
- 10:56Evaluation: Set up
- 11:35Evaluation: End-to-end Latency
- 12:05Evaluation: Effectiveness
- 12:44Evaluation: Scalability
related talks
01h 09mVector Search: Ask Me Anything!
16mTowards High-throughput and Low-latency Billion-scale Vector Search via CPU/GPU...
35mImpact of Vector Search
15mQuake: Adaptive Indexing for Vector Search
26mThe Rise of Vector Databases - Lessons from the Milvus Community
24mVector Search at Uber
40mUnder the hood of vector search with JVector
12mAchieving Low-Latency Graph-Based Vector Search via Aligning Best-First Search...