What is an embedding?
An embedding is a numeric vector representation of data such as text. It lets systems compare meaning using similarity math.
Interview Questions: What Embeddings Are
Use these to explain embeddings clearly without overselling them.
Why does RAG use embeddings?
RAG uses embeddings to retrieve chunks that are semantically related to the user's question, even when exact words differ.
What is cosine similarity?
Cosine similarity measures the angle between vectors. Similar directions usually mean related meaning.
Are embeddings proof of truth?
No. Similarity means likely related, not necessarily correct, current, authorized, or complete.
What is the difference between dense and sparse retrieval?
Dense retrieval uses learned semantic vectors. Sparse retrieval uses exact or weighted term matching such as TF-IDF or BM25.
When is hybrid retrieval useful?
When both meaning and exact terms matter, such as support docs with error codes, legal clauses, SKUs, or configuration keys.
What happens when you change embedding models?
You usually need to re-embed the corpus and rebuild the vector index because vectors from different models may not be comparable.
What metadata should you store with embeddings?
Chunk id, source id, source text, title, section, access labels, version, effective date, chunker version, embedding model, and index version.
What are embedding production risks?
Privacy, cost, latency, drift, stale vectors, dimension mismatch, weak evals, and overtrusting semantic similarity.
How do you evaluate embeddings in RAG?
Use golden questions with expected chunks and measure Recall@k, MRR, gold chunk rank, citation support, and downstream answer quality.