Same size. Total noise.

transcript

A model upgrade ships. Every search still returns 10 results. Every one of them is noise.

The deploy embeds queries with a new model. Version 2. But 3.2 million stored vectors are still version 1. Re-embedding is next sprint.

Here's the trap. Both models make 1,024 numbers. So nothing errors. No alert fires.

Each model places meaning in its own vector space. And the two spaces aren't aligned. So nearest, between a v2 query and v1 documents, means nothing.

Catch it first. Store the model version with every vector. Refuse a query whose model differs from the index. Run recall evals on every deploy. A drop to near zero is unmissable.

The fix? Keep querying with v1. Build a complete v2 index beside it. When recall and relevance pass, switch both together.

Same dimensions don't mean the same meaning. Treat a model change like a schema migration. Version it, build beside it, then flip.

Full breakdown in the lecture.