RAG Inc. is designing a RAG pipeline and must choose an embedding model and similarity search configuration to use Amazon OpenSearch Service. (select TWO)

The two correct options describe best practices for building a RAG pipeline with Amazon OpenSearch Service. They both involve using a dedicated embedding model (like those from Bedrock or hosted on SageMaker) to generate dense, high-dimensional vectors, which are then stored in OpenSearch using a densevector field. The OpenSearch k-NN plugin, specifically with HNSW (Hierarchical Navigable Small Worlds) and cosine similarity, is crucial for efficient Approximate Nearest Neighbor (ANN) search, which is fundamental for RAG. Tuning efconstruction and efsearch allows for balancing recall and latency. The incorrect options are flawed because: Using Amazon Comprehend for sentiment analysis doesn't generate general-purpose embeddings suitable for semantic search. Term-based inverted indices are not designed for semantic similarity. Storing embeddings in DynamoDB and using GSI for cosine similarity is not an efficient or scalable approach for vector similarity search compared to specialized vector databases or k-NN plugins. Relying solely on term matching (BM25) in OpenSearch without vector embeddings will not achieve semantic similarity, which is a core requirement for RAG.