Amazon Aurora PostgreSQL Improves Vector Search Performance and Relevance with pgvector 0.8.0

With Amazon Aurora PostgreSQL-Compatible Edition now supporting pgvector 0.8.0, vector search performance and relevance have improved significantly. The key is iterative_scan, which mitigates the overfiltering problem where filtered ANN searches returned too few results or missed results entirely.

pgvector 0.8.0 offers three modes.

• off: existing behavior
• strict_order: maintains exact distance ordering
• relaxed_order: returns results faster with approximate ordering

Also, thanks to improved cost estimation, traditional indexes like B-tree are now better chosen for complex filtering queries. For HNSW, at the default search setting of hnsw.ef_search = 40, if only 10% of the data matched the filter, only about 4 results could be obtained; but iterative_scan continues the index scan until it fills the required number of results, improving recall.

The tuning points are clear. For HNSW indexes, hnsw.max_scan_tuples defaults to 20,000, and hnsw.scan_mem_multiplier can be used to adjust the memory cap as a multiple of work_mem. However, using relaxed_order may require re-sorting the results at the end, so, as in the example, a method that forces a separate sort may be needed.

The benchmark was conducted on a synthetic e-commerce dataset built from 10 million products. Two Aurora PostgreSQL clusters on db.r8g.4xlarge were compared, and p99 latency was measured for queries A through E. Notably, Query A improved from 123.3ms → 13.1ms, about 9.4x faster, and Query E improved from 913.4ms → 160.3ms, a 5.7x improvement. Even for searches with category filters, the optimal configuration of 0.8.0 was faster than 0.7.4 while still returning a complete result set.