GCP Generative AI Leader Practice Question 8: Grounding Internal Policies
Google Cloud Generative AI Leader · Test 1 · question 8 of 10
An HR assistant built on Gemini answers employee questions about benefits and leave. Users report that it sometimes gives confident but wrong answers. It has cited a parental-leave policy the company replaced last quarter and invented a reimbursement limit. HR policies change several times a year, and employees want to see which document an answer came from.
Which approach best addresses these problems?
Show answer and explanation
Answer:
The failures come from missing or outdated knowledge. Grounding connects model output to verifiable sources. When the assistant retrieves from the current policy documents (retrieval-augmented generation, for example through Agent Search or RAG Engine on Gemini Enterprise Agent Platform), its answers are tied to that content. Google cites three benefits of grounding: fewer hallucinations, responses anchored to your own data, and links to sources for auditability. When a policy changes, HR updates the document store, and the next answer reflects the change. No retraining is needed, and the citations meet the employees' request.
Option A is slow and expensive because it means a new tuning run for every policy change. It also does not reliably stop the model from inventing details, and it produces no source citations.
Option B uses the wrong source. Google Search grounds answers in public web content, which does not include the company's internal leave or reimbursement policies. It suits questions about world knowledge, not first-party enterprise data.
Option D makes answers more repeatable, but a model that lacks the current policy will simply repeat the same wrong answer. Temperature changes randomness, not knowledge.
Reference: Grounding overview