EmbedFlow, Before the Backfill

Field notes

The proposed workflow keeps the old model and existing index live while a target index is populated progressively, shifting re-encoding work after deployment rather than blocking the upgrade. The page grounds its scale example in Qwen3-8B H100 throughput of 106.98 documents per second, projecting 25,960 H100 hours for 10 billion documents. It also reports 63 studied migrations and a 19.8 ms measured warm p50 overhead, though the underlying method is not published here.