GCP Generative AI Leader Practice Question 95: RAG Control Tradeoff
Google Cloud Generative AI Leader · Test 10 · question 5 of 10
A legal technology team is grounding an assistant in scanned contracts that require a non-standard chunking strategy and a specific embedding model already used by the firm's existing search index, so the retrieved passages line up with that index. The team has the engineering capacity to build and maintain this pipeline. Which approach best fits their requirement?
Show answer and explanation
Answer:
RAG Engine is the lower-level toolkit for building a custom retrieval pipeline: a team can pick the vector database to use, including their own, choose the embedding model, and configure the retrieval and ranking steps, so retrieved passages can be made to line up with an index the firm already runs elsewhere. That component-level control over the vector store and the retrieval/ranking pipeline is exactly what this team needs and has the engineering capacity to build and maintain.
Option A fails on data source, not on customization: a website data store points Agent Search's crawler at public web pages, and the team's contracts are scanned files in the firm's own storage, not a website to crawl, so this connector is the wrong tool for the source regardless of how much chunking or embedding control it exposes. Option C addresses individual document summarization inside Docs, not a searchable, grounded retrieval pipeline across the contract archive. Option D grounds answers in public web content, not the firm's own scanned contracts, so it does not touch the requirement at all.
Reference: RAG Engine overview