FDEInterviews logo
← RAG & Agent System Design / 57
hardGleanDatabricksOpenAI

You upgraded the embedding model and re-embedded the corpus. Nothing errored, no alert fired, and answer quality fell. Find out why.

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 Sep 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.

20 answers per topic instead of 10, plus saved progress and bookmarks · no cardor unlock all 543 remaining answers · ₹2,000 / $25

Every answer, concept and course, all published video lessons, hands-on FDE Lab missions, Premium PDF guides and companion files, the full practice-test bank and work-sample downloads. Referral Premium excludes guide PDFs and their companion files.

6 months · One payment · No auto-renewal

Study alongside free video lessons.

UP NEXT ON YOUR JOURNEY
Next in this trackYour agent writes what it learns into memory. A wrong conclusion from three weeks ago now shapes every answer. Find it and remove it.Next in this trackThe agent works in the pilot and falls apart somewhere between ten and fifty concurrent runs. Diagnose it before you add capacity.Next in this trackThe search index has to be rebuilt weekly, the rebuild takes ninety minutes, and users must not see stale results. Design it.
FEDITOR'S NOTE

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.

DISCUSSION · 0

No comments yet — be the first to share your approach.