RAG & Retrieval · Ground answers in evidence
What is multi-vector retrieval and when does it help?
Short answer
Multi-vector retrieval stores several vectors per document, such as passage or token embeddings, and scores them jointly at query time. It improves fine-grained matching at higher storage and compute cost.
A single vector compresses a whole chunk into one point, which loses detail. Multi-vector approaches keep finer representations, such as one vector per passage or token-level late interaction, so a query can match a specific part of a long document. The trade-off is index size and query cost.
- ColBERT-style late interaction gives strong precision but increases storage significantly.
- Multiple passage vectors per document can improve long-document retrieval.
- Query cost rises because scoring is more expensive than a single nearest-neighbour lookup.
- Use it when reranking alone does not recover enough precision.
Follow-ups you should expect
Interviewers use these to check whether the first answer was understanding or memorisation.
- 1How would you estimate the storage cost of late interaction?
- 2When would a reranker be a better investment?
- 3What kind of queries benefit most from multi-vector retrieval?
Say it in the interview
Multi-vector retrieval keeps more than one representation per document, so a query can match a specific passage or token rather than a single blended vector. It improves precision on long or nuanced documents, but storage and query cost go up. I would try better chunking and a reranker first, because they are simpler. If precision is still the bottleneck, multi-vector is a reasonable next step.
Use this as a model, not a script. Saying it in your own words is what makes it survive follow-up pressure.
Related questions
Reading is step one
Now say it out loud.
Understanding an answer and delivering it under pressure are different skills. Rehearse this topic in a live RAG & Retrieval mock interview and get scored feedback on depth, clarity and confidence.