Your Compiler

Vector Databases & Embeddings

How embeddings represent meaning, how vector databases index and search them at scale, and the tradeoffs behind both.

5 of 23 published

Beginner

What exactly is an embedding, and why does AI engineering need them?Beginner6 min

Opener question. They're checking whether you can explain a core concept without hiding behind jargon.

What happens inside an embedding model when it turns text into numbers?BeginnerComing soon

Separates people who've used an embeddings API from people who know what it's doing.

Sparse vs dense vectors — when does each one make sense?BeginnerComing soon

A setup question. Your answer here decides whether they ask you about hybrid search next.

Cosine, dot product, Euclidean: which distance metric should you actually use?BeginnerComing soon

Almost everyone can define them. Far fewer can say when the choice changes your results.

Why can't you just use Postgres for vector search?Beginner1 min

Tests whether you reach for new infrastructure by default or because something specific required it.

What metadata should you store alongside your vectors, and why?BeginnerComing soon

A schema-design question in disguise. Gets asked because most people under-plan this and pay for it later.

Intermediate

How does ANN search trade accuracy for speed, and where's the knob?Intermediate1 min

They want to hear you say "approximate" and mean it. Knowing there's a recall cost at all is the bar.

Picking an embedding model: what actually matters beyond the leaderboard?IntermediateComing soon

Tests whether you evaluate for your domain or copy whatever's top of MTEB.

Does a bigger embedding dimension make search better? What does it cost?IntermediateComing soon

A cost-awareness question. Common follow-up when you mention a model choice.

How do multi-modal embeddings put images and text in the same space?IntermediateComing soon

Comes up if the role touches search, recommendations, or anything with images.

Quantizing embeddings: how much can you compress before quality drops?Intermediate1 min

Asked at companies actually paying an infrastructure bill.

Why pure vector search loses to hybrid search more often than people expectIntermediateComing soon

One of the strongest signals that you've shipped RAG rather than read about it.

How do you actually measure whether your embeddings are any good?IntermediateComing soon

The question most candidates fumble. If you can name real metrics, you stand out immediately.

Two-word queries return garbage. What's going wrong?IntermediateComing soon

A debugging question. They're watching how you narrow down a cause, not whether you know the answer.

Advanced

Rolling out a new embedding model without breaking production searchAdvancedComing soon

A migration question. Tests whether you think about the system while it's serving traffic.

Multi-tenancy in vector search: namespaces, filters, or separate indexes?AdvancedComing soon

Asked at any B2B company. There's no single right answer, which is the point.

What breaks first when you go from a million vectors to a billion?Advanced1 min

Scale question. They want to know if your mental model has real limits in it.

When is fine-tuning an embedding model worth it, and how do you do it?AdvancedComing soon

Half the signal here is knowing when NOT to.

Your RAG index is eating all your RAM. Where does the memory actually go?AdvancedComing soon

Tests whether you know what's actually stored in an index versus what you assumed.

Vector search got slow as you scaled. How do you find the real bottleneck?AdvancedComing soon

A diagnosis question. Guessing loses; having an order to check wins.

New model, different dimensions, existing vectors in prod. Now what?AdvancedComing soon

A specific, nasty situation. Usually asked by someone who has lived it.

Similarity score is 0.91 and the result is wrong. Why?AdvancedComing soon

The best question on this list. Understanding why a high score can be meaningless is what separates senior from mid.

Search quality collapsed overnight after a model swap. How do you respond?AdvancedComing soon

An incident-response question. They're assessing how you behave when something is on fire, not what you know.