Embeddings drift after fine-tuning — how to maintain search quality?

Issue

After fine-tuning text-embedding-3-small on our domain corpus, the embedding space shifted enough that our existing HNSW index returns irrelevant results for queries that used to work perfectly.

Numbers

• 50k documents re-embedded: ~$12 one-time cost
• But now 40% of top-1 retrievals are wrong
• Rolling back the fine-tune fixes it, but loses the quality gains

Question

Is there a principled way to incrementally update an HNSW index after a model change, or do we have to nuke and rebuild every time we fine-tune?