The correct answer is B because HNSW with the Faiss engine is the best fit when the system needs approximate k-NN search over millions of high-dimensional vectors while maintaining low latency and strong recall. OpenSearch approximate k-NN is designed for large, high-dimensional datasets where brute-force exact search becomes inefficient, and OpenSearch documentation shows HNSW configured with the Faiss engine for knn_vector fields. Exact search with a scoring script is accurate but does not scale well for 5 million vectors under a strict latency requirement. IVF can scale but usually requires training and tuning and may trade off recall. Binary quantization reduces memory but can reduce accuracy. Therefore, HNSW with Faiss is the strongest exam answer.