AlloyDB ScaNN reaches 10B vectors

Google says a preview four-level tree for AlloyDB’s ScaNN index can serve more than 10 billion vectors at ≤51 ms p95 latency and 95% recall in internal tests, a scale jump aimed at enterprise agentic workloads. The architecture narrows traversal hierarchically and uses memory-conscious sampling, but the figures are Google’s internal results rather than an independent benchmark or a price/performance guarantee. Teams scaling retrieval should compare the workload against the archive’s provider-swap search benchmark: vector capacity only matters if recall and tail latency survive the actual agent loop.