RAG & Retrieval · Ground answers in evidence
How do you choose an embedding model?
Short answer
Evaluate on your own retrieval data across languages, domains and query types, then weigh dimension, latency, cost and hosting constraints. Re-embed the corpus when you switch.
Embedding quality is task-specific. A model that leads a public benchmark may underperform on your product names, legal language or code. The decision should come from retrieval recall on real queries, not from a leaderboard alone.
- Test multilingual coverage if users write in more than one language.
- Check the maximum input length and how the model handles long chunks.
- Compare vector dimension: larger costs more storage and slows search.
- Consider hosted versus self-hosted for privacy, latency and cost.
- Keep the model version pinned; upgrades require re-embedding.
Follow-ups you should expect
Interviewers use these to check whether the first answer was understanding or memorisation.
- 1How would you benchmark two embedding models fairly?
- 2What would make you self-host an embedding model?
- 3How do you migrate a large index to a new model?
Say it in the interview
I evaluate embeddings on my own corpus, because public benchmarks rarely match product language. I look at recall on real queries, then weigh dimension, latency, cost and whether self-hosting is required. I pin the model version, because changing it means re-embedding everything. I also check the instruction format, since some models need different prefixes for queries and documents, and getting that wrong quietly hurts retrieval.
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.