Task 2.4: Describe how Google Cloud empowers developers to build with AI.

• ASet a Model Garden organization policy that allows only the vetted models

  Correct

  A Model Garden organization policy at the organization, folder or project level can allow access to specific vetted models and deny all others.

• BMove the vetted models into Model Registry and delete the rest

  Incorrect

  Model Registry manages versions of your own models; it does not control which Model Garden models developers can reach.

• CTurn on Model Monitoring for every endpoint that serves one of the models

  Incorrect

  Model Monitoring does not control which models developers can use.

• DAsk developers to fine-tune each model before they use it

  Incorrect

  Tuning customizes a model; it is not an access control.

Model choice and governance go together: Model Garden is open by design, and organization policies narrow it to what the company has approved.

• AGemini Notebook Enterprise

  Incorrect

  Gemini Notebook Enterprise is an application for people, not a RAG service for developers.

• BSupervised fine-tuning of Gemini

  Incorrect

  Tuning teaches a model behavior from labeled examples; it does not retrieve current passages from the manuals at question time.

• CRAG Engine on Agent Platform

  Correct

  RAG Engine is a configurable managed RAG service: it covers ingestion, transformation such as chunking, embedding, indexing and retrieval, with a choice of embedding models and vector databases.

• DModel Monitoring

  Incorrect

  Model Monitoring plays no part in retrieval.

Agent Search gives RAG out of the box; when the team wants to configure the pipeline's parts, RAG Engine is the managed, configurable option.

• AAgent Studio

  Incorrect

  Agent Studio is the low-code, visual way to build agents; teams that need exact control in code use ADK.

• BAgent Gallery in Gemini Enterprise

  Incorrect

  Agent Gallery is the catalog in Gemini Enterprise where people find agents to use; it is not a development framework.

• CAgent Development Kit (ADK)

  Correct

  ADK is an open-source, code-first framework for building, evaluating and deploying agents, with native support for multi-agent architectures and flexible orchestration.

• DAgent Platform AutoML

  Incorrect

  AutoML trains custom image and tabular models with little ML effort; it does not build agents.

Low-code tools for speed, ADK for code-level control over multi-agent logic.

• AAgent Garden

  Incorrect

  Agent Garden does not host or scale an agent in production; it is not a runtime.

• BAgent Runtime

  Correct

  Agent Runtime is a fully managed environment for deploying, hosting and scaling agents built with ADK or other frameworks, LangGraph among them, with integrated sessions.

• CModel Registry

  Incorrect

  Model Registry tracks and deploys versions of ML models; it is not the runtime for an agent application.

• DCompute Engine VMs that the team manages

  Incorrect

  VMs can run anything, but the team would manage the servers and scaling itself, which it wants to avoid.

Build with any framework, then let the managed runtime handle hosting, scaling and sessions.

• AWith a Provisioned Throughput commitment

  Incorrect

  Provisioned Throughput is a capacity commitment, the wrong tool for a one-off job that can wait.

• BWith Priority pay-as-you-go requests

  Incorrect

  Priority PayGo charges a premium rate for more reliable, faster processing, which this job does not need.

• CThrough context caching on every request

  Incorrect

  Context caching cuts the cost of reusing the same long input across requests; each description here is different.

• DAs a batch inference job on Agent Platform

  Correct

  Batch inference is the cost-optimized option for high-volume, asynchronous processing whose results are needed within a longer timeframe, billed at a discounted per-token rate.

Match the consumption option to the workload: large, patient jobs go to batch inference.

• ACustom training with the team's own code

  Incorrect

  Custom training gives full control but needs ML engineers to write the training code, which the retailer lacks.

• BAgent Platform AutoML for tabular forecasting

  Correct

  AutoML trains high-quality models with minimal effort and ML expertise, and supports tabular objectives including forecasting.

• CGrounding a Gemini model with Google Search

  Incorrect

  Grounding adds web information to generated answers; it does not build a forecasting model on the retailer's data.

• DAgent Garden agent samples adapted for forecasting

  Incorrect

  Agent Garden provides agent samples, not demand-forecasting models.

AutoML is the low-expertise path to a custom model on your own data; custom training is the full-control path.

• AMemory Bank

  Correct

  Memory Bank generates long-term memories from conversations and keeps personalized information that persists across sessions, so the agent can personalize later responses.

• BVector Search

  Incorrect

  Vector Search finds similar embeddings at scale; it does not keep each customer's preferences for the agent.

• CModel Registry

  Incorrect

  Model Registry versions ML models; it does not store facts about users.

• DContext caching

  Incorrect

  Context caching reduces the cost of reusing the same long input; it is not a per-user memory.

Long-term, per-user memory across conversations is what Memory Bank adds to an agent.

• AAgent Garden

  Correct

  Agent Garden is a curated library of prebuilt agent samples for common AI patterns, such as RAG-based support, with source code for customization.

• BModel Garden

  Incorrect

  Model Garden is a library of models to discover and deploy, not of agent templates.

• CAgent Platform Pipelines

  Incorrect

  Pipelines are not a library of agent templates for developers to start from.

• DAgent Gallery in Gemini Enterprise

  Incorrect

  Agent Gallery is where employees find and run agents in Gemini Enterprise, not a set of code samples for developers.

Agent Garden is for developers starting a build; Agent Gallery is for employees using finished agents.

• ASimilarity scores from Vector Search

  Incorrect

  Vector Search scores how close each stored passage is to the query; it says nothing about whether the generated answer's claims are supported.

• BModel Armor screening of each response

  Incorrect

  Model Armor screens prompts and responses against security and safety policies; it does not check an answer's claims against the retrieved sources.

• CContent filter scores on each response

  Incorrect

  Content filters score how likely and how severe harmful content is for each harm category; a safe answer can still be unsupported by the sources.

• DThe check grounding API of Agent Search

  Correct

  The check grounding API takes an answer candidate and a set of reference facts and returns a support score, with citations showing which facts support which claims.

Agent Search's RAG APIs cover the steps of a RAG pipeline, including checking that answers are grounded before they go out. Retrieval scores and safety checks answer different questions.

• AVector Search

  Correct

  Vector Search finds semantically similar items among very large numbers of embeddings with low latency, the core of similarity search and many RAG systems.

• BModel Registry

  Incorrect

  Model Registry organizes and versions models; it does not search embeddings.

• CBatch inference

  Incorrect

  Batch inference runs large asynchronous model jobs; it is not built for millisecond similarity lookups.

• DAgent Platform Pipelines

  Incorrect

  Pipelines do not serve low-latency similarity queries.

Embeddings capture meaning; Vector Search finds the nearest ones at scale.