GCP Generative AI Leader Practice Question 98: Feature Store Purpose
Google Cloud Generative AI Leader · Test 10 · question 8 of 10
A retailer's data science team computes customer features, such as average order value and days since last purchase, once for model training. In production, a different team recomputes the same features from scratch for real-time predictions, and the two calculations sometimes disagree, degrading model quality. Which Agent Platform capability is designed to prevent this?
Show answer and explanation
Answer:
Feature Store centralizes feature computation so the same values used to train a model are also served for real-time predictions, at low latency. That shared repository is exactly what avoids the disagreement the retailer is seeing, known as training-serving skew, because there is no longer a second, separately maintained calculation to drift out of sync.
Option A is wrong because Model Garden is a catalog for discovering and deploying models, not a place to manage feature values. Option B is wrong because Vector Search solves nearest-neighbor retrieval over embeddings, not feature reuse between training and serving. Option D is wrong because grounding improves the factual accuracy of generated text; it has nothing to do with how numeric features are computed or served.
Reference: About Feature Store on Gemini Enterprise Agent Platform