Concept guide · Embeddings and vector search
What are embeddings?
A slower explanation for when the short interview answer is not enough. Start with the intuition, then connect it to the technical idea and the way an interviewer expects you to reason about it.
ML area
Generative AI → Embeddings and semantic representation
Interview format
Foundations · Concept
Question
GENAI-033
Start with the idea
In simple terms
An embedding turns text, an image, or another object into a list of numbers. The list is useful because items with similar meaning can end up near one another, so an application can search by meaning instead of only matching exact words.
Make it concrete
A user asks, “How can I change my billing address?” A semantic search can find a document titled “Updating account payment details” even though the wording is different.
Build the answer
How to reason about it
- 1Explain that an embedding is a learned representation, not a human-readable summary.
- 2Connect distance in the vector space to a retrieval or matching task.
- 3Name the limits: model choice, exact terms, metadata, permissions, and evaluation.
Go one level deeper
The technical explanation
An embedding model maps an input into a vector space learned from its training objective. Similarity in that space is useful only for the task and data distribution it was trained to represent. Embeddings do not prove that two passages are interchangeable, do not preserve every fact equally, and do not enforce access permissions. A retrieval system still needs metadata, filters, ranking, and evaluation.
Say this in the interview
Embeddings are numeric representations learned so that items with useful semantic relationships tend to be near one another in a vector space. They support retrieval and similarity tasks, but the geometry is task- and model-dependent; an embedding is not a universal database of meaning.
Keep preparing