How do I evaluate Vector Search quality?

Every MongoDB Atlas Vector Search implementation we ship includes an evaluation suite. Standard approach: build a golden set of 50 to 200 real queries with known-relevant results, then score MRR (Mean Reciprocal Rank), Recall@K, and NDCG on every deployment. Tools like Ragas, Promptfoo, and Langfuse make this repeatable inside your CI. Without an eval suite, you have no signal on whether an embedding model swap, a chunking change, or a retrieval tweak actually improved anything. This is the single most valuable engineering artifact on any Vector Search project.