Embeddings Drift Like Clocks Do, and Nobody Is Running NTP for Them
Last Updated on October 29, 2026 by Editorial Team
Author(s): “The AI Engineer”
Originally published on Towards AI.
Your vector index can lose all of its retrieval quality without throwing a single error. I built a simulation to see how, and here is the fix borrowed from how computers keep time.
Your laptop is wrong about the time right now. It was wrong yesterday by a slightly different amount, and it will be wrong tomorrow by another one.
After introducing the time-drift analogy, the article explains what embedding drift is and why it’s hard to detect: providers can retire models (sometimes switching to incompatible vector spaces even with the same dimension), pipelines can change preprocessing (chunking, cleaning, tokenization, etc.), and the world/corpus evolves. It then shows a worst-case simulation where a “new model” implemented as a rotation plus noise causes retrieval to fall to chance with no warnings, demonstrates how partially reembedded (“mixed”) indexes behave and why they can break queries, and notes that learned adapters can bridge old and new spaces when trained on paired embeddings. The author then proposes an “embedding NTP” built from three pieces—(1) a golden set of real queries to track Recall@10 baseline on deploys, (2) daily anchor strings to measure encoder consistency via embedding similarity, and (3) version stamps in vector metadata to block unsafe writes when model/preprocessing changes. Finally, it clarifies where the clock analogy doesn’t hold (steps vs continuous drift) and summarizes the takeaway: with reference checks, measurements, and gradual/controlled correction strategies (dual collections, adapter testing, and gated cutover), retrieval quality can stay trustworthy even when models change silently.
Read the full blog for free on Medium.
Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.
Published via Towards AI
Towards AI Academy
We Build Enterprise-Grade AI. We'll Teach You to Master It Too.
15 engineers. 100,000+ students. Towards AI Academy teaches what actually survives production.
Start free — no commitment:
→ AI Engineering For Production — our new book, coming before the end of 2026. Join the waitlist.
→ Agents Architecture Cheatsheet — 3 years of architecture decisions in 6 pages
Our courses:
→ AI Engineering Certification — 90+ lessons from project selection to deployed product. The most comprehensive practical LLM course out there.
→ Agent Engineering Course — Hands on with production agent architectures, memory, routing, and eval frameworks — built from real enterprise engagements.
→ AI for Work — Understand, evaluate, and apply AI for complex work tasks.
Note: Article content contains the views of the contributing authors and not Towards AI.
Related posts
Recent Posts
![[Day 37/100] Build Your Own Agent Memory Layer with Postgres and pgvector](https://miro.medium.com/v2/resize:fit:700/1*-Doos2nmUT3FPZ-Qxbk3Hw.png "[Day 37/100] Build Your Own Agent Memory Layer with Postgres and pgvector")
