You discover that switching from text-embedding-ada-002 to text-embedding-3-large improves retrieval quality by 18% on your eval set. Your production system has 50M vectors indexed in Pinecone, serves 5,000 QPS, and cannot go down. Design the complete migration plan. Include: how to validate the new model does not regress on your specific domain before committing, the dual-index strategy during migration, traffic routing during cutover, how to handle the 2-week migration window where new documents must be indexed in both models, rollback triggers and plan, and cost of the migration in tokens and dollars.