The correct option is Implement Retrieval Augmented Generation that calls a live inventory database and a real time trend analytics feed before responding.

This approach connects the model to current external sources before it generates a response. By retrieving stock status from a live inventory database and pulling trend signals from a real time analytics feed, the model grounds its recommendations in fresh data. This addresses the out of stock risk and adapts to fast moving fashion trends that can shift within two hours. With Retrieval Augmented Generation you can add connectors or extensions that call APIs and databases at inference time so responses reflect what is true right now.

Human in the loop review and prompt engineering does not supply the model with live data and it is not practical for real time mobile recommendations at scale. Prompt tweaks cannot ensure the model knows current inventory or rapidly changing trends.

Fine tune the model on last quarter sales and raise the temperature relies on stale training data and does not fetch new information at inference time. Increasing temperature only changes randomness and diversity of outputs and it does not improve freshness or accuracy about stock or trends.

Nightly retraining with BigQuery ML and Cloud Memorystore caching of inventory snapshots cannot keep up with changes that occur within hours and cached snapshots can quickly become outdated. This batch workflow does not guarantee that recommendations reflect real time availability or trend shifts.

Full AWS Practitioner Certification Question