Concept guide · Embeddings and vector search
What are multi-vector or late-interaction retrieval methods?
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 → Multi-vector and late-interaction retrieval
Interview format
Advanced · Scenario
Question
GENAI-044
Start with the idea
In simple terms
Single-vector retrieval compresses a whole passage into one vector. Multi-vector or late-interaction methods keep several token- or passage-level representations so detailed matches can be compared more precisely.
Make it concrete
A technical document may contain several independent symbols, names, and conditions. A single vector can blur them together, while token-level interactions can preserve more of the matching detail.
Build the answer
How to approach it
- 1Identify whether single-vector compression is causing a fine-grained retrieval miss.
- 2Explain the quality benefit of keeping multiple representations.
- 3Compare the gain with storage, latency, and operational complexity.
Go one level deeper
The technical explanation
Late-interaction systems delay some of the query-document comparison until retrieval, rather than forcing all meaning into one vector. This can improve fine-grained relevance for technical or long documents, but it increases storage, indexing, query computation, and serving complexity. Treat it as a measured retrieval upgrade, not a universal replacement for dense or hybrid search.
Say this in the interview
They represent a document or query with multiple token- or passage-level vectors and compare finer-grained interactions instead of compressing everything into one vector. This can improve matching for detailed text, but storage, indexing, and serving are more complex than single-vector retrieval.
Keep preparing