Prepare and test your skills
Prepare and test your skills
Google Cloud's strength in generative AI comes from turning long-term research into enterprise-ready systems. Research breakthroughs such as the Transformer architecture became foundation models and optimized infrastructure that businesses can test, tune, and deploy. Data flows from customer datasets into Vertex AI (opens in a new tab) services, models flow into deployment, and deployed models serve predictions to applications.
Traditional AI and generative AI differ in what they learn from and what they produce. Traditional AI trains on historical structured data such as tables and images to predict outcomes or assign categories, including spam filtering, sales forecasting, price estimation, and customer segmentation. Generative AI trains on broad patterns across text, images, video, and audio, then creates new content such as summaries, answers, code, translations, and media.
| Approach | What it learns from | What it produces |
|---|---|---|
| Traditional AI | Historical tabular data and images | Classifications, forecasts, predictions, and segments |
| Generative AI | Broad patterns across text, images, video, and audio | Summaries, conversations, code, media, and insights |
Generative AI excels at content creation, conversational search and chatbots, workflow automation, finding insights in documents, and generating code. The two approaches can combine when traditional AI output becomes part of a generative prompt, such as a churn forecast explained in natural language by a chatbot.
The path from research to business use runs through Vertex AI Model Garden, Vertex AI Studio, and managed deployment. Model Garden provides Google models such as Gemini and Imagen alongside select open-source and third-party models, including Hugging Face text-generation models. Vertex AI Studio supports prompt design for models including Gemini and PaLM 2, while Colab Enterprise notebooks and software development kits (SDKs) help teams compare outputs and move from experiment to production.
The development lifecycle moves in order, with the first four stages forming a repeating evaluation loop:
Models selected in Model Garden can deploy in Vertex AI or self-host on Google Kubernetes Engine (GKE) or Compute Engine, where graphics processing units (GPUs) attached to virtual machine (VM) instances accelerate generative workloads. Embedded help brings the same model family into existing workflows through Gemini Cloud Assist, Gemini Code Assist, Gemini in BigQuery, Gemini in Looker, and Gemini in Security Command Center.
Enterprise readiness on Google Cloud connects data boundaries, responsible-use checks, and managed scaling. Customer data used in managed search and agent services is not used to train foundation models. The "your data is your data" commitment means customer data stays isolated, encrypted, and reachable only by identities the customer approves.
Data protection applies in a fixed order. First, sensitive data is placed in its own perimeter using VPC Service Controls drawn around Google-managed services so data cannot be copied out without authorization. Then Identity and Access Management (IAM) controls which approved identities can reach data inside the perimeter. Then multifactor authentication is required for users who need access to sensitive data. Sensitive Data Protection finds and classifies sensitive data such as credit card numbers and personally identifiable information (PII), while Assured Workloads adds boundaries for regulated workloads such as FedRAMP and CJIS.
Responsible AI adds request-time safety checks and design-time evaluation. In Vertex AI, a user prompt first passes through safety classification, then the model generates a response, then the response passes through safety classification again before flowing back to the application. Administrators choose a threshold from permissive to most restrictive for hate speech, harassment, sexually explicit content, and dangerous content. Teams also define quality, groundedness, and fairness metrics, run curated test prompts, score responses, and retest until outputs meet standards.
Exam tip: VPC Service Controls governs whether data can move across a perimeter, while IAM governs which identities can access data inside it.
AI-optimized infrastructure co-designs accelerators, networking, storage, and software to run generative AI at scale. Data flows from cloud storage to the accelerator attached to the VM, where training consumes it, and the trained model then serves inference requests. AI Hypercomputer (opens in a new tab) combines accelerators, networking, storage, consumption models, and cluster management into one system.
Task-matched accelerators handle different jobs. GPUs are NVIDIA accelerators for general workloads from prototyping to large-scale training and serving. Cloud TPUs (Tensor Processing Units) are Google custom chips for large matrix operations with on-chip high-bandwidth memory for larger models and batch sizes. Machine choice depends on scale:
BigQuery supplies feature data to Vertex AI Feature Store, where online serving retrieves the latest values for real-time inference and offline serving exports volumes for training. Grounding connects models to real-time facts through retrieval-augmented generation (RAG) using Vertex AI Search, Vertex AI RAG Engine, Vertex AI Vector Search, or SQL results from Cloud SQL or Firestore placed directly into the prompt. Open choice prevents lock-in because the same open models can run on Vertex AI for managed deployment, on GKE for Kubernetes portability, or on Cloud Run (opens in a new tab) for serverless inference that scales to zero. Low-code access comes from managed notebooks, prompt tools, pre-trained speech, language, vision, and document APIs, and Model Garden discovery with model cards, vulnerability scanning, and organization policies that allow only vetted models.
Continue with this study guide: 2.1 Describe Google Cloud's strengths in the field of gen AI
Prebuilt generative AI offerings bring help into daily work without building models from scratch. Custom Gems save reusable expertise, role-based assistants help inside developer and data tools, and enterprise search and Workspace assistants answer from company and work context. Together they trade custom development for speed and consistency.
Custom Gems are reusable AI experts inside Gemini Advanced that save a role, tone, and set of rules for a recurring job. A Gem holds the background instructions once, such as how to review contracts, coach support replies, or summarize research, and applies that same pattern every time it is used. This reduces setup time and keeps answers aligned when different people need the same specialized behavior.
Role-based assistants divide help by where the user already works. A data analyst gets help in BigQuery or Looker, a developer gets help in Code Assist, and a security analyst gets help in Security Command Center:
Gemini Enterprise is an intranet search, AI assistant, and agentic platform that connects Google and third-party systems into one permissions-aware interface. Data flows from Google and third-party sources through prebuilt connectors into dedicated data stores, and a Gemini Enterprise app serves results from those data stores to the user. Supported connections include Confluence, Jira Cloud, Microsoft SharePoint, ServiceNow, plus Google content and other SaaS systems.
The trust boundary sits between the user and the stored data. Traffic flows from the user query to the app, the app checks single sign-on and user-level access, and only content the user is allowed to see flows back as results. Because answers are grounded in those data stores, they include citations to enterprise sources. Cross-media work lets the assistant analyze uploaded PDFs and images, summarize complex files, answer follow-up questions, generate images from text prompts, and create reports with streamed research.
Gemini Notebook Enterprise and Gemini Enterprise answer different needs. Notebook Enterprise is a focused research assistant that answers only from a curated notebook of documents, websites, and notes. Gemini Enterprise is a broad search and agentic platform that searches many systems and uses agents to act across them. Knowledge flows in order: first Gemini Enterprise searches broadly and surfaces valuable sources, then the user adds those sources into a notebook, then future chat and content stay grounded in that curated set.
Custom agents automate multi-step processes. Within Gemini Enterprise, an app must connect to one or more data stores to use their data, while actions let the platform create or update records in real time, such as creating Calendar events or editing Jira issues. Teams can build agents with Workflow Builder, start from prebuilt agents in the Agent Gallery, or import outside agents through Agent Registry. Observability provides real-time traces and audit logs to diagnose latency, while analytics provides usage and search quality through a Looker-powered dashboard.
Exam tip: Choose Gemini Notebook Enterprise when answers must stay confined to a known set of sources, and choose Gemini Enterprise when the starting point is unknown and search must span all enterprise data.
Gemini for Google Workspace is a generative AI collaboration product that assists users inside everyday work apps. The request flows from the open Workspace app to the Gemini models, and the response flows back into the same document, message, spreadsheet, presentation, or meeting. The user describes what is wanted in natural language, reviews the draft Gemini creates, and edits the final version.
Core assistance follows the same describe, review, and edit pattern across five apps. Gmail drafts replies, summarizes threads, and adjusts tone and length. Google Docs drafts from a prompt, summarizes, rewrites, and proofreads. Google Slides generates images, drafts outlines, and summarizes slides. Google Sheets organizes tables, generates formulas, and summarizes patterns. Google Meet captures summaries, highlights key points, and supports follow-up.
Enterprise-grade privacy rests on encryption, identity controls, and limits on training. Gemini does not use prompts or responses to train models, and data sent to models is used strictly to serve the response. Prompts are encrypted in transit, and data is encrypted at rest and in transit. Cloud Run (opens in a new tab) and Vertex AI show the same managed pattern for hosting agents, while IAM controls who can use Gemini features and VPC Service Controls can create a perimeter around resources such as BigQuery. Outputs can seem plausible but be factually incorrect, so users should validate all output, and Gemini in BigQuery does not support the same compliance offerings or individual-location data residency as BigQuery itself.
Google Cloud improves customer experience by connecting generative models to current information and by automating contact center work. External search grounding keeps answers fact-based, commerce search personalizes product discovery, and the Customer Engagement Suite combines virtual agents, human assistance, and analytics. Interaction data then flows back into operations to improve future service.
Grounding is the process of linking a Gemini model response to real search results before the final answer is returned. An ungrounded model answers only from patterns learned during training, which can lead to outdated or invented facts. A grounded model sends the user prompt to a search tool, receives relevant evidence, and then builds its answer from that evidence.
Google Search grounding and Vertex AI Search differ in source and flow. Google Search grounding is a tool enabled inside a Vertex AI generation request. The application builds a request that names a model such as gemini-2.5-flash, includes user content, and adds a Tools field containing GoogleSearch. The request is sent through a prediction client to a regional endpoint, and the model returns candidates whose text reflects live Google Search results. Setting a low temperature such as 0.0 makes output more factual and consistent. Vertex AI Search with RAG follows a different path from business data stores to the model to the user, where semantic retrieval understands meaning instead of matching only exact keywords. Supported patterns include RAG with Vertex AI Vector Search for similarity search, RAG with AlloyDB for PostgreSQL for combining vector search with operational data, and RAG with GKE and Cloud SQL plus open-source tools for customizable hosting.
Exam tip: Google Search grounding connects a response to live Google Search results through the GoogleSearch tool, not only to the model's training data.
AI Commerce Search is Google Cloud's managed search and recommendations offering for websites and mobile applications. Data flows from the product catalog and user events into AI models, and the models send ranked search results and recommendations back to the storefront. The engine applies query understanding and personalization to broad queries, long-tail searches, and conversational requests.
Personalization is driven by search history, shopping context, and product SKUs. Guided search refines a broad query through back-and-forth conversation, dynamic filtering, and product image tiles, while autocomplete and semantic search help shoppers narrow choices in real time. The service supports text and image discovery for similar or complementary items across channels. Because results reflect actual behavior and catalog information, shoppers encounter fewer dead ends, less search abandonment, and fewer No Results Found pages.
Commerce data creates business value when interaction data flows from the storefront into ranking and then into analytics. AI Commerce Search uses interaction and ranking models to order results for engagement, revenue, and conversions, leading to higher click-through, higher order value and conversion rates, and stronger engagement. The same catalog and user event data also informs merchandising when exported to BigQuery, where teams review dashboards in Looker and generate forecasts with Vertex AI. The platform is fully managed and handles preprocessing, training, tuning, load balancing, and provisioning, and the same ingested data serves both search and recommendations without ingesting twice. Strong access controls keep customer data isolated, and personalized results from one shopper must never be cached and returned to a different shopper.
Google's Customer Engagement Suite modernizes the contact center by combining automation, human assistance, and analytics in one connected flow. Conversational Agents resolve common requests first, Agent Assist helps human representatives during live conversations, and Conversational Insights analyzes those conversations to improve service. CCAI Platform, the Google Cloud Contact Center as a Service (CCaaS), provides the cloud contact center where voice and digital interactions run.
Conversational Agents are virtual agents that automate conversations across digital and voice channels by understanding what a customer says, keeping conversation context, and responding or taking action. Dialogflow is the core service for building them. Designers use a visual console to create agents, while developers use APIs to connect those agents to applications.
| Option | Agent size | Complexity |
|---|---|---|
| Dialogflow ES | Small to medium agents | Simple to moderately complex conversations |
| Dialogflow CX | Large agents, including contact center agents | Large and complex conversations |
Traffic flows from the customer to the virtual agent first, and the virtual agent resolves common requests or collects information before routing to a human only when needed. Voice and language APIs let the same logic serve chat and phone: spoken audio flows to Speech-to-Text (opens in a new tab) for transcription, the agent processes text meaning with Cloud Natural Language, written replies flow to Text-to-Speech for synthesis back to voice, and Cloud Translation translates across languages.
Conversational Insights and CCaaS play connected roles. Conversational Insights, delivered as Customer Experience Insights, turns raw conversations into call drivers, sentiment, and intent. CCaaS, delivered as CCAI Platform, runs agents, queues, IVR, CRM integration, and digital channels. A typical CCAI Platform setup runs in order: add agents and teams, create IVR queues and assign agents, integrate with a CRM such as Salesforce or Zendesk, test the integration, port the support phone number, extend to mobile, web, and SMS, then use real-time dashboards for reporting. Agent Assist then helps representatives during active conversations with generative knowledge assist, summarization for after-call work, smart reply, and live translation, which reduces average handle time and prevents repeat contacts.
Developers move from foundation to production by discovering a model, grounding it in enterprise facts, and assembling it into a custom agent. Vertex AI Platform provides discovery, training, evaluation, and deployment with consistent governance. Data flows from customer datasets to training or retrieval services, models flow into Model Registry for versioning, and deployed versions serve predictions to applications.
Vertex AI Model Garden is the part of Vertex AI used to discover, manage, and deploy foundation models. A foundation model is trained on broad data and can be adapted through fine-tuning to many tasks. Model Garden brings together Google proprietary models and third-party and open-source alternatives so a team can match a model to task, licensing, and infrastructure needs, including Gemini for text and multimodal prompts, Codey for code, Imagen for images, and text-embedding models for search and classification.
Deployment choice depends on how much infrastructure management the team wants to keep:
| Option | Management | Choose when |
|---|---|---|
| Vertex AI Model Garden | Managed by Vertex AI | The team wants fully managed infrastructure with generative tools and APIs |
| Google Kubernetes Engine (GKE) | Self-hosted | The team wants Kubernetes-based deployment of open-source models |
| Cloud Run | Self-hosted | The team wants serverless containerized inference for smaller models |
| Compute Engine | Self-hosted | The model needs a specialized kernel, legacy OS, or special licensing |
After selection, Vertex AI Model Registry organizes and tracks versions and assigns versions to endpoints directly or with aliases. Vertex AI Model Evaluation then assesses performance and quality with metrics and visualizations to identify bias and support deployment decisions. To manage a procured Marketplace model on Vertex AI, the team opens the product page in Cloud Marketplace and clicks Manage on Vertex AI.
Exam tip: Model Garden manages discovery and deployment of Marketplace and open models, while Model Registry manages the lifecycle of machine learning models and deploys versions to endpoints.
Vertex AI AutoML is a managed training service that builds custom models from labeled data with minimal coding. Data flows from the customer dataset to the AutoML training job, the resulting model flows to Model Registry, and the registry version flows to an endpoint for inference. AutoML supports image classification and object detection, text classification and extraction, tabular prediction including time-series forecasting, and video action analysis. Training uses about 80% of data to teach patterns, with separate validation and test sets to check performance.
The tradeoff is simplicity versus control. AutoML requires a predefined objective and performs automated hyperparameter tuning that the user cannot view or modify, while custom training allows manual tuning and control over machine type, disk size, and container image. The user controls only the training-hour budget and early stopping. Training new AutoML models for text classification, entity extraction, and sentiment analysis is being replaced by Gemini prompts and tuning, while existing AutoML text models remain usable for a transition period.
BigQuery ML brings AI functions directly into the warehouse workflow. Managed AI functions perform semantic analysis over structured rows without calling a large language model for every row, using an efficient mode that trains a lightweight distilled model on a sample and then uses it for inference on most rows. Task-specific functions send each type from BigQuery to its matching Cloud AI model for natural language processing, translation, document processing, audio transcription, and computer vision, with results flowing back into the same query.
Retrieval-Augmented Generation (RAG) grounds a model in company or real-time data before it answers. The retrieval layer finds relevant content, the retrieved content flows into the model prompt, and the model generates a response based on that context. A hallucination happens when a model generates fluent but incorrect content because training knowledge is incomplete or out of date.
Grounding follows a fixed order:
A process flow showing the three-step RAG grounding workflow: the user query is embedded and compared against an indexed search engine, the most relevant passages are retrieved, and the query plus retrieved passages are sent to the Gemini model to generate a grounded answer instead of a hallucination.
An embedding is a numerical representation of text that lets search compare meaning, not just keywords. Google Cloud offers Vertex AI Search (opens in a new tab) as a fully managed solution without direct embedding access, Vertex AI RAG Engine (opens in a new tab) as a managed orchestrator for customizable retrieval, Vertex AI Vector Search as a specialized low-latency vector database, and AlloyDB, Cloud SQL, and BigQuery for operational and analytical data. High-impact uses include semantic search across repositories, context-aware support assistants, and internal knowledge management.
Vertex AI Agent Builder assembles custom agents from models and enterprise data using a low-code app graph where nodes for models, input sources, and destinations define data flow. Pre-trained models are ready to use for general tasks, while user-trained models require supplied sample data and training with AutoML or custom training. The builder flow runs in order: select a model and review configuration, connect it to an input source, then add the model as a node so its output can flow to streams, other models, or a media warehouse. Governance includes customer-managed encryption keys, VPC Service Controls, data residency, evaluation with precision and recall, and monitoring. For isolated environments, a Distributed Cloud air-gapped appliance runs select services on one NVIDIA A100 80 GB GPU, so only one GPU-dependent API such as OCR or Translation can run at a time, while CPU-only Speech-to-Text can run alongside either.
Continue with this study guide: 2.4 Describe how Google Cloud empowers developers to build with AI
Agent tooling lets a generative AI agent move beyond frozen training data to use live systems and current information. A language model alone cannot know today's news, current database rows, or live account state. Tools close that gap by letting the agent retrieve real-time data or perform an API-driven action, then use the returned result to finish the task.
The three core tool categories are extensions, function calling, and data stores, and each answers a different need. Extensions connect an agent to external systems and live APIs. Function calling lets an agent run a defined operation or custom code, such as a calculation or workflow step. Data stores ground answers in private organizational data.
Data flows from the external API or database into the model prompt, then from the model to the user as a grounded answer or action. Grounding connects an agent to verifiable information before it answers, so outputs are based on facts instead of only training data. The RAG workflow runs in order: first the user query is converted into an embedding and compared against indexed data, then the system retrieves the most relevant documents or API results, and only after retrieval does the model receive the query plus retrieved facts to generate the final answer.
An embedding converts text, images, or video into an array of floating point numbers called a vector, where closer proximity in vector space means more similar meaning. Many engines store embeddings in a vector database optimized for similarity search. Google Cloud offers Vertex AI Search as a managed engine without managing embeddings, Vertex AI RAG Engine as a customizable orchestrator, and Vertex AI Vector Search as an optimized database for low-latency large-scale search. AlloyDB and Cloud SQL support embeddings for operational data, while BigQuery connects models to large analytical datasets.
Google Workspace Google Search grounding lets companion agents answer about current news from real-time search results built from the user prompt, customer context, and search data. Enabling it carries a tradeoff because Customer-Managed Encryption Keys, Data Residency, VPC Service Controls, and Access Transparency do not apply to those search requests.
Exam tip: A data store supplies private data for grounding, an extension reaches a live external API, and function calling runs a defined action or custom code.
Relevant services for agent tooling provide compute to run agents, storage to remember context, and prebuilt APIs to handle speech, language, and documents. The agent logic lives in a hosted environment, calls outward to tools through APIs, and uses returned results to decide the next action.
Cloud Run is a serverless execution environment that hosts agent logic without managing servers. An n8n agentic workflow on Cloud Run shows the pattern: traffic flows from the workflow outward to each connected app, and results flow back into the workflow. The visual node-based editor lets each node represent a step or service call, so reasoning leads directly to actions in mail, documents, or calendar.
Agent persistence separates unstructured file storage from structured memory. Cloud Storage is object storage for files, images, and datasets, while databases hold records and embeddings that agents query to remember context. The agent often pulls a raw document from Cloud Storage, sends it through processing, and writes embeddings into AlloyDB for PostgreSQL for RAG retrieval. Cloud SQL supports RAG on GKE, Spanner holds global multi-tier state, and BigQuery and Bigtable serve large-scale query workloads.
Specialized APIs let an agent work with speech, language, and documents without training. Speech-to-Text transcribes voice to text, and Text-to-Speech synthesizes text into human voice, forming a loop where user audio flows into transcription, text flows to the agent, and the agent reply flows into synthesis. Translation translates text for multilingual chatbots, Natural Language analyzes meaning, and Document AI extracts data from reports, invoices, and contracts to highlight errors, risks, or fraud. Conversational data agents let users ask about database data in natural language: data flows from the question to the agent, the agent translates it into SQL, the query executes, and the agent returns a summary or visualization.
Choosing between a pre-trained API and custom training starts with what already exists. First the team checks whether a pre-trained model such as Document AI, Vision AI, Natural Language API, or Video Intelligence API meets requirements, and if so uses it with no training. Only if none fits does the team consider custom training, choosing traditional training when high control over algorithms and hyperparameters is required and generative AI when faster go-to-market is the priority.
Google AI Studio and Vertex AI Studio are two entry points to Gemini models, divided by lifecycle stage and governance needs. Google AI Studio supports fast prototyping outside a Cloud project, while Vertex AI Studio supports the full lifecycle from customization to deployment inside Vertex AI on Google Cloud.
Google AI Studio is a browser-based tool for trying Gemini models quickly with minimal setup. A builder signs in, writes a prompt, adjusts temperature and output length, and receives a response within seconds. The prompt can then be turned into code that calls the Gemini API with an API key, so data flows directly from the application to the model and back. It does not manage Cloud projects, IAM permissions, virtual private networks, or long-term operations. Choose it to test an idea, compare prompts, explore multimodal inputs, and build a first demo.
Vertex AI Studio is the generative console inside Vertex AI for operating models at scale. A team works inside a Cloud project, so models, data, tuning jobs, and endpoints inherit ownership, billing, access controls, and network boundaries. It supports advanced stages in order: prototype with enterprise Gemini models, customize with supervised fine-tuning on private data, evaluate base and tuned models, deploy the approved version to Model Registry and a Vertex AI endpoint, then monitor quality and retune after performance drops. A common path is to start in Google AI Studio and rebuild the working prompt inside a Cloud project for governed deployment.
Enterprise integration explains the production choice. Vertex AI Studio runs with billing, enabled APIs, and IAM least privilege, plus customer-managed encryption, perimeters, residency, and audit transparency. Tools for customer experience agents (opens in a new tab) connect virtual agents to business systems, while conversational data agents over AlloyDB (opens in a new tab) and conversational data agents over Spanner (opens in a new tab) translate natural language into SQL. Agents move from draft for personal testing to published for broader use, and Agent Registry indexes discoverable agents through metadata and the Agent2Agent protocol.
Continue with this study guide: 2.5 Define the purpose and types of tooling for gen AI agents
Traditional AI trains on historical structured data such as tables and images to produce classifications, forecasts, predictions, and segments, while generative AI trains on broad patterns across text, images, video, and audio to create new summaries, conversations, code, media, and insights. The two approaches can combine when traditional AI output becomes part of a generative prompt, such as a churn forecast explained in natural language by a chatbot.
Use Gemini Enterprise when the starting point is unknown and search must span all enterprise data, and use Gemini Notebook Enterprise when answers must stay confined to a known set of sources. Gemini Enterprise is a broad search and agentic platform that searches many systems and uses agents to act across them with user-level access control and citations, while Notebook Enterprise is a focused research assistant that answers only from a curated notebook of documents, websites, and notes.
Use Google AI Studio for fast prototyping outside a Cloud project and use Vertex AI Studio for governed deployment inside Vertex AI on Google Cloud. Google AI Studio is a browser-based tool for trying Gemini models quickly with minimal setup without managing Cloud projects, Identity and Access Management (IAM) permissions, virtual private networks, or long-term operations, while Vertex AI Studio supports the full lifecycle from customization to deployment with ownership, billing, access controls, and network boundaries.
Retrieval-Augmented Generation (RAG) grounds a model by retrieving relevant passages first and adding them to the prompt before generation. Company documents are prepared in a search engine often as embeddings for similarity search, then when a user asks a question the system returns the most relevant passages, and those passages flow into the model prompt so the model generates an answer from that context.
Last updated