Safe Postgres Embedding Migration Keeps Course-Page Search Live During Reindexing

Migrating embedding models should be treated as a data migration: a second index is built under an explicit version while the current index keeps serving traffic. The new index is only promoted to reads after it passes a retrieval-quality check and a latency target. Vectors from different models must never be combined in one searchable set, which keeps alerts about course-page changes reliable during background re-embedding. The write-up argues that uptime alone is misleading, since recall can quietly degrade while a service still responds quickly. To prevent stale data, each page revision is stored as the durable unit of work and writes are checked against the expected revision, with deletions applied to both indexes.