Embedding Drift
Also called Embedding Drift
When you change embedding models and old vectors no longer line up with new ones.
Think of it like
Two people mapping the same city with different grid systems — the coordinates don’t match, so directions between the maps fall apart.
Example
A team upgrades to a newer embedding model but only re-embeds new documents; searches now compare apples to oranges and results quietly degrade.
How it actually works
Embeddings are only comparable within the same model — each model places meaning in its own coordinate space. Swap the model and every stored vector must be regenerated, or you’re measuring distances between incompatible spaces. Drift also creeps in subtler ways, like a provider silently updating a hosted model. The rule is blunt: change the embedding model, re-embed everything.
For product teams
A quiet quality regression that follows a "harmless" model upgrade if you don’t re-index.
For engineers
Vectors are only comparable within one embedding model’s space; on any model change, re-embed the entire corpus before serving.
Related
- Embeddings — The vectors that fall out of alignment.
- Index Refresh — The fix is a full re-index.
- Cosine similarity — Distances become meaningless across spaces.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome