Upgrading your embedding model changes the geometry of the vector space. Old embeddings are useless for the new query vectors.

Data Flow

Day 0:  docs embedded with model M1   V1-M1   saved -> vector store
Day 30: QA team upgrades to M2 (better accuracy)

query embedded with M2  Q-M2

cos(Q-M2, V1-M1)  <-- nonsense, scores scrambled

What it is

The mapping text -> vector changes. Old text -> vector mappings were calibrated to M1’s geometry and are misaligned with M2’s queries.

How to detect

• Compare retrieval Recall@5 against a gold eval set BEFORE and AFTER the upgrade.
• Watch cache_hit_rate and avg_retrieval_score; a step change = drift.

How to fix

• Re-index in the background; dual-write during cutover; serve reads from old while writing to new; switch when reindex finishes.
• Pin the embedding model version; reject queries with different versions.

Analogy

Replacing the measuring stick used to map a forest: if the old stick was in inches and the new one in millimeters, you cannot compare coordinates.

Interview tip: Run the embedding migration with dual-write + shadow-eval, not a hot swap.