A team wants to use Retrieval Augmented Generation (RAG) with an open-source LLM running on Amazon Bedrock. The RAG source documents (CSV and DOCX files) are stored in an S3 bucket. Which approach requires the least operational overhead to enable RAG over those files?

Creating a knowledge base for Amazon Bedrock is the most operationally efficient solution. Bedrock knowledge bases are purpose-built for RAG, automatically handling data ingestion, chunking, embedding, and vector storage (using Amazon OpenSearch Service or Pinecone/Redis Enterprise Cloud). You simply point it to your S3 bucket, and Bedrock manages the entire RAG pipeline, minimizing manual setup and maintenance. The other options involve significantly more operational overhead: SageMaker Pipelines for a new model or fine-tuning an existing LLM are for model development and training, not directly for RAG over existing documents, and require managing model deployment and inference. Converting data to vectors and storing it in Amazon Neptune would require manual vectorization, managing the Neptune database, and integrating it with Bedrock, adding complexity.