Spot the production bug in this embedding-migration code

Attempt it

Click any words you think contain an error. Click again to unmark.

Mark at least one word to submit.

Quick reference

TL;DR

The migration is broken because it leaves old vectors in place. A query has to compare against vectors from one model, not two unrelated coordinate systems; recall silently collapses.

Memory aid

Sign in for the mnemonic that makes this stick. See Pro pricing.

Easy to grasp

Picture a library where half the books are catalogued by the Dewey Decimal System and half by a private numbering scheme invented last week. When you ask the librarian for 'the books nearest to call number 510.3', she gives you Dewey books that match and private-system books that happen to have the same digits, even though those digits mean something completely different on the new shelves. The library does not warn you. The search just brings back a mix of the right answers and accidental matches. Upgrading a search catalog midway behaves the same way: leaving old cards in place means the librarian is now searching across two incompatible filing schemes at once.

Key concepts

Embeddings Retrieval-augmented generation

Detailed answer & concept explanation

~5 min read

Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.

Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example.

Sign in for 5 free unlocks/day or go Pro for unlimited. See Pro pricing.

Quick scan · 2–4 min

Everything important, quickly.

4 min: locate the continue-branch bug + geometric reason mixed-model fails + parallel-index migration stages + cost and storage math + dual-write, shadow-test, atomic cutover + model_id discipline.

Real-world examples

Real products, models, and research that use this idea.

OpenAI's text-embedding-3-large migration guidance explicitly warns against mixed-model indexes and recommends a parallel-index pattern with full corpus re-embedding.
Pinecone's 2026 embedding-migration playbook documents the dual-write plus shadow-test plus atomic-cutover pattern as the supported approach.
Production RAG stacks at Notion and Cursor maintain model_id and model_version metadata on every vector so a botched migration is at least auditable.

Interviewer follow-ups

What an interviewer would ask next. Try answering before peeking at the approach.

QHow would you size the cost of re-embedding a billion-doc corpus on text-embedding-3-large?

Sign in or go Pro to see the approach for this follow-up.

QHow do you detect that someone has already shipped a mixed-model index in production?

Sign in or go Pro to see the approach for this follow-up.

QCan you ever justify a mixed-model index in production?

Sign in or go Pro to see the approach for this follow-up.

Don't say this

Red flags and common mistakes that signal junior thinking. Click to expand.

Most common mistake

Treating the migration as 'only re-embed new docs' to save the cost of re-embedding the back catalogue, which silently breaks recall across the entire index.

Sign in or go Pro to see all red flags and common mistakes. See Pro pricing.

60-second cheat sheet

The night-before-the-interview bullets. Scan these on the way to the call.

Sign in for 5 unlocks/day or go Pro for unlimited revision sheets. See Pro pricing.

References

Primary sources. Skim if you want the original framing.