Option B is correct because Amazon Bedrock Knowledge Bases supports structured document metadata that can be stored separately from the text and used as filters during retrieval. Case type, jurisdiction, and case outcome are ideal filterable metadata attributes, while author and version fields preserve provenance for returned documents. AWS documentation confirms that S3 knowledge-base data sources can use sidecar metadata files and that metadata remains available for filtering even when it is excluded from the embedding. This gives the application precise semantic retrieval plus structured constraints without custom extraction at query time. Option A lacks structured filtering. Option C adds unnecessary Comprehend processing. Option D embeds metadata inside document text, which does not provide reliable field-level filtering and adds unnecessary agent-side extraction complexity. AWS Documentation

================