A 30-question deep dive into embeddings, vector search and retrieval

techNmak · x · 2026-08-26

A systematic 30-question breakdown of embeddings, vector search and retrieval:

Fundamentals: what "closeness" in embedding space means, why pooling strategies can't change at serving time, when cosine/dot product/Euclidean rank identically, why similarity scores aren't calibrated probabilities
Training: contrastive learning for relevance; in-batch negatives, hard negatives and the false-negative trap
Retrieval: how BM25 ranks, when lexical beats dense, SPLADE, hybrid retrieval and RRF (and why hybrid isn't automatically better)
Reranking: bi- vs cross-encoders, rerank depth, ColBERT late interaction
ANN indexes: how HNSW really works (and why it's not O(log N) in production), IVF probing trade-offs, PQ, choosing between exact/HNSW/IVF/IVF+PQ
Production: how metadata filtering breaks ANN search, inserts/deletes/freshness, embedding model upgrades, memory estimation
Evaluation: Precision@k, Recall@k, MRR, nDCG; debugging embeddings vs index; why the top leaderboard model may be wrong for production

More from coding & agent