A legal tech firm must let its assistant answer from 200,000 private contracts that update daily, cite source clauses, avoid retraining a model, and keep data within its VPC boundary. Answers must stay current as PDFs arrive in S3. Which RAG ingestion and retrieval strategy best meets these grounding and freshness requirements?
Select an answer to reveal the explanation.
Short Explanation
You’ll get fresh cited answers without pricey retraining, like a librarian fetching exact pages on demand. Bedrock knowledge bases plus embeddings keep retrieval grounded, current, and private.
Full Explanation
The Bedrock knowledge base with S3 chunking, vector embeddings, and VPC endpoints is correct because it delivers retrieval-augmented generation without training: documents are chunked, embedded and stored for semantic retrieval, data source sync keeps 200,000 contracts current as PDFs arrive, grounding with source attribution provides clause citations, and VPC endpoints keep embeddings, queries and private text inside the boundary. The weekly in-VPC full rebuild with re-embedding and clause metadata fails on freshness because periodic batch reindexing creates a stale window where daily arrivals are unsearchable and superseded clauses remain retrievable, violating continuous currency despite correct isolation and citation mechanics. The auto-syncing managed vector service with citations but without private connectivity fails on data residency because default managed endpoints let prompts, embeddings and contract text transit outside the VPC, violating confinement despite correct freshness and grounding. The nightly VPC-hosted fine-tuning approach fails on the no-retraining constraint because updating parametric weights needs training jobs, risks catastrophic forgetting and hallucinated citations, and replaces verifiable retrieval with learned memory that cannot guarantee current clause-linked answers. For the exam, beware valid production patterns that omit one constraint word such as daily, cited, or VPC endpoints, as RAG sync versus fine-tuning versus hosted search are often contrasted this way; as a practical check, verify continuous S3 sync status, confirm interface VPC endpoints and policies for Bedrock and the vector store with no internet path, and test a newly ingested contract returns an exact retrievable clause identifier.