Business Strategies for Generative AI Solutions →mediumMultiple Choice

Generative AI Leader Practice Question: Business Strategies for Generative AI Solutions

A regional insurance company wants to let claims adjusters ask natural-language questions about 12 years of policy documents and claim histories stored in Cloud Storage. Leadership wants a working prototype in two weeks, minimal model-tuning effort, and answers that cite the exact source document. Which Google Cloud approach should the team implement?

⚠ Common exam trap

The trap here is assuming that fine-tuning or a larger context window can substitute for retrieval when the real requirement is traceable, up-to-date grounding in a large document corpus.

Answer choices

Why each option matters

Answer the question above first, then reveal the full breakdown to understand why each option is right or wrong.

Correct answer & explanation

✓

Build a retrieval-augmented generation flow using Vertex AI Search to index the corpus and ground Gemini responses in retrieved passages.

Retrieval-augmented generation pairs a managed search index with a foundation model, so answers are grounded in retrieved passages and can point back to the originating document. Vertex AI Search handles ingestion and indexing of the Cloud Storage corpus, while Gemini synthesizes the retrieved context into an answer. This avoids costly fine-tuning and oversized prompts while meeting the citation and speed requirements.

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

✓
Build a retrieval-augmented generation flow using Vertex AI Search to index the corpus and ground Gemini responses in retrieved passages.
Why this is correct
Vertex AI Search indexes the Cloud Storage corpus and returns relevant passages at query time, which Gemini then uses as grounding context. This delivers citable, source-linked answers without retraining, and the managed indexing pipeline keeps the prototype within a two-week window while honoring the citation requirement.
✗
Increase the Gemini model's context window to the maximum and paste the entire document corpus into every prompt.
Why it's wrong here
Even the largest context windows cannot hold 12 years of policy and claim documents, and cost and latency scale with prompt size, so this is economically and technically unworkable. It also provides no retrieval ranking, so the model may overlook the relevant passage and cannot reliably cite the source document.
✗
Fine-tune a Gemini model on the historical claims corpus, then deploy the tuned endpoint behind an internal web app.
Why it's wrong here
Fine-tuning changes model weights to absorb style and task patterns, but it does not give the model a live, citable index of 12 years of documents, and citations cannot be traced to a source passage. It also requires curated training data and repeated tuning cycles, which conflicts with the two-week prototype goal and the citation requirement.
✗
Train a custom text embedding model from scratch on the claims corpus, then serve similarity search from a self-managed cluster.
Why it's wrong here
Training embeddings from scratch requires large labeled datasets, GPU time, and ML expertise that a two-week prototype cannot absorb. Self-managing the vector store also adds operational burden, and embeddings alone do not generate grounded answers or citations, so the adjusters would still lack a usable question-answering experience.

Go deeper

Generative AI Leader Practice Question: Business Strategies for Generative AI Solutions

Why each option matters

Option-by-option breakdown

Related to this question