TL;DR

SOAR (Spilling with Orthogonality-Amplified Residuals) recovers recall lost at partition boundaries in ANN indices by spilling boundary queries to neighboring partitions using orthogonality-weighted residuals.

Summary

Partition-based Approximate Nearest Neighbor (ANN) indices systematically miss queries that fall near partition boundaries, since only the assigned partition is searched. SOAR augments these indices with a residual spilling mechanism: boundary queries are forwarded to neighboring partitions using residuals amplified by their angular deviation from the partition boundary — prioritizing the most informative spills. The technique is compatible with existing IVF-style indices as an augmentation layer and achieves superior recall-latency tradeoffs over ScaNN and HNSW on standard ANN benchmarks at matched latency budgets.

Key contributions

  1. Identifies partition-boundary recall loss as a systematic failure mode in standard ANN indices and proposes a targeted solution.
  2. Introduces residual spilling as a mechanism to forward boundary queries to neighboring partitions without full reindexing.
  3. Proposes orthogonality amplification as a principled weighting of residuals by angular deviation from partition boundaries.
  4. Demonstrates superior recall-latency tradeoffs over ScaNN and HNSW on standard ANN benchmarks at matched latency.

When to cite

  1. When optimizing ANN recall at fixed latency and seeking alternatives to brute-force multi-probe search.
  2. When augmenting an existing IVF-style vector index to recover recall at partition boundaries.
  3. When comparing vector search index designs and needing a reference that beats ScaNN and HNSW on recall-latency curves.
  4. When arguing that boundary handling in partitioned indices is a distinct optimization axis from quantization.

AI VDB RAG