GenAI interview guides / Embeddings, Vector Search & Reranking Interview Guide
Embeddings, Vector Search & Reranking Interview Guide
Design retrieval stacks from embedding choice and ANN indexes through hybrid retrieval and reranking.
Embedding layer
Choose embeddings using domain quality, language coverage, dimension, latency, cost, and update strategy.
ANN trade-offs
Approximate indexes trade recall for memory and latency; discuss build/update cost, filters, sharding, and recall@k.
Hybrid retrieval
Dense semantic retrieval and lexical exact-match retrieval cover different failure modes.
Reranking
Retrieve cheaply, then apply a stronger reranker to a smaller candidate set and justify the latency/cost increment.