Vectors from two versions of an embedding model are not comparable, and nothing in the stack will tell you so. The cosine between the same document's old and new vector is roughly zero, which is why a half-finished migration returns confident nonsense at full speed.
Updated Aug 2026 · Grounded in real Forward Deployed Engineer interview loops and written to a senior-engineer editorial bar.
Vectors from two versions of an embedding model are not comparable, and nothing in the stack will tell you so. The cosine between the same document's old and new vector is roughly zero, which is why a half-finished migration returns confident nonsense at full speed.
The point candidates miss is that this failure cannot raise an error, because nothing invalid happens. Two vectors of the same dimension always have a cosine, so a half-migrated index returns confident nonsense at full speed. Strong answers reach for the detectable signature rather than a theory: in a half-migrated index the newly embedded documents took 90% of the top ten slots against an unbiased 50%, and that skew is visible in production logs weeks before anyone files a quality ticket.
No comments yet — be the first to share your approach.
