Quick Navigation
Core AI Hierarchy and TerminologyGoogle Foundation ModelsGen AI Landscape and ML LifecycleAgent Platform and InfrastructureGemini Product SurfacesCustomer Engagement Suite and Developer ToolingPrompt Engineering TechniquesGrounding, RAG, and Model LimitationsFine-Tuning, Prompt Tuning, and Sampling ParametersResponsible AI, SAIF, and SecurityBusiness Strategy and Measuring Impact
Core AI Hierarchy and Terminology
- AI > Machine Learning > Generative AI
- Artificial intelligence is the broadest field (systems that mimic human intelligence). Machine learning is a subset that learns patterns from data. Generative AI is a subset of ML that creates new content (text, images, audio, video, code). Never treat these as interchangeable.
- Foundation Model
- A large model pre-trained on massive, broad datasets that can be adapted (via prompting, tuning, or fine-tuning) to many downstream tasks without training from scratch. Gemini, Gemma, Imagen, and Veo are all foundation models.
- LLM (Large Language Model)
- A foundation model specialized in understanding and generating text. All LLMs are foundation models, but not all foundation models are LLMs (e.g., Imagen is a foundation model but not an LLM).
- Multimodal Model
- A model that can accept and/or generate more than one type of data — text, images, audio, video — in a single request. Gemini is Google's flagship multimodal model family.
- Diffusion Model
- A generative model that creates images (and video) by starting from random noise and iteratively denoising it into a coherent output. Imagen and Veo use diffusion techniques; diffusion models do not generate text.
- Prompt Engineering vs Prompt Tuning
- Prompt engineering crafts the input text to guide a model's output — no change to the model itself, fast, no training data needed. Prompt tuning is a lightweight training technique that learns soft prompts to adapt model behavior — it does change the model's parameters (a small set of them).
- Three ML Approaches
- Supervised learning trains on labeled data (input-output pairs). Unsupervised learning finds patterns/clusters in unlabeled data. Reinforcement learning learns a policy through trial-and-error reward/penalty feedback from an environment.
Google Foundation Models
- Gemini
- Google's flagship, proprietary, multimodal foundation model family. Handles text, images, audio, video, and code. Offered in tiers trading off reasoning depth vs speed/cost (a top-tier reasoning model, a versatile Pro tier, and a fast/cheap Flash tier).
- Gemma
- Google's family of lightweight, open models derived from Gemini research. Freely downloadable, customizable, and deployable anywhere including on-premises. Represents Google's 'open approach' — choose Gemma when the scenario stresses openness, self-hosting, or full control.
- Imagen
- Google's text-to-image generation model (diffusion-based). Use when the scenario asks for generating still images from text descriptions.
- Veo
- Google's text-to-video generation model. Use when the scenario asks for generating video content (e.g., promotional clips) from text descriptions.
- Choosing a Foundation Model — Business Factors
- Match the model to the use case based on: modality needed, context window size, security requirements, availability/reliability, cost, performance, and fine-tuning/customization needs — a leadership trade-off, not just an accuracy question.
Gen AI Landscape and ML Lifecycle
- Five Gen AI Landscape Layers
- Infrastructure (compute/storage/networking, e.g., TPUs and data centers) → Models (foundation models like Gemini) → Platforms (tools to build with models, e.g., Vertex AI/Agent Platform) → Agents (autonomous task-completing systems) → Applications (end-user products).
- ML Lifecycle Stages
- Data ingestion → data preparation → model training → model deployment → model management (ongoing monitoring, versioning, retraining). Google Cloud provides tools at every stage.
- Create / Summarize / Discover / Automate
- Google's gen AI business use case framework. Create = generate new content (text/image/code/video). Summarize = condense large content into concise insights. Discover = find information/insights through search and analysis. Automate = streamline repetitive workflows.
- Structured vs Unstructured Data
- Structured data has a predefined schema (database tables, spreadsheets) and is easy to query. Unstructured data has no predefined format (documents, images, audio, video, emails) — it is the majority of enterprise data and where gen AI adds the most value.
- Labeled vs Unlabeled Data
- An independent axis from structured/unstructured. Labeled data has tagged outputs for supervised learning; unlabeled data has no tags and is used for unsupervised learning. Unstructured data (like images) can still be labeled; structured data can be unlabeled.
- Data Quality Dimensions
- Completeness, consistency, relevance, availability, cost, and format — the six dimensions Google's framework uses to evaluate whether data is fit for a gen AI initiative.
Agent Platform and Infrastructure
- Vertex AI / Agent Platform
- Google Cloud's unified, enterprise-ready AI platform where organizations access models, build agents, and manage gen AI at scale. Provides governance, security, and access to first-party, open, and third-party models. The umbrella for Domain 2 (35% of the exam).
- Model Garden
- A curated catalog (not a single model) within the Agent Platform offering first-party models (Gemini, Gemma, Imagen, Veo), open-source models, and third-party models for discovery, testing, and deployment. Embodies Google's open approach.
- AI Hypercomputer
- Google Cloud's AI-optimized infrastructure layer — combines custom-designed TPUs, GPUs, data centers, storage, and software to train and serve large models efficiently. The foundation of Google's scale/efficiency advantage.
- AutoML (Agent Platform)
- Low-code/no-code capability to train custom models on your own data without deep ML expertise — democratizes model creation for non-specialists.
- Feature Store
- Centralized repository for managing and serving ML features consistently — appears under continuous monitoring/model management, not content creation.
- Google's Four Differentiators
- 1) AI-first approach. 2) Enterprise-ready platform (responsible, secure, private, reliable, scalable). 3) Comprehensive AI ecosystem integrated across Google products. 4) Open approach offering first-party, open, and third-party models.
- Data Control and Democratization
- Google Cloud gives customers control over their own data (security, privacy, governance) while democratizing AI access through low-code/no-code tools, pre-trained models, and pre-built APIs — no deep ML expertise required.
- curl -X POST \ -H "Authorization: Bearer $(gcloud auth print-access-token)" \ -H "Content-Type: application/json" \ "https://us-central1-aiplatform.googleapis.com/v1/projects/PROJECT_ID/locations/us-central1/publishers/google/models/gemini-2.5-flash:generateContent" \ -d '{"contents":[{"role":"user","parts":[{"text":"Summarize this document"}]}]}'
- Example authenticated REST call to a Vertex AI Gemini model, using gcloud to mint the access token — the underlying mechanism behind Model Garden and Agent Studio calls.
Gemini Product Surfaces
- Gemini App / Gemini Advanced / Gems
- Gemini app = consumer-facing assistant. Gemini Advanced = unlocks the most capable models and larger context. Gems = user-created, customizable, reusable personas for recurring tasks — not full custom agents.
- Gemini Enterprise
- Business platform for multimodal enterprise search, Cloud NotebookLM API-style research, and building custom agents grounded in a company's own first-party data. Distinct from the consumer Gemini app.
- Cloud NotebookLM API
- A capability surfaced through Gemini Enterprise that lets organizations ground responses and generate research insights over their own document collections — recognize it as an enterprise research/grounding tool, not a standalone consumer product.
- Gemini for Google Workspace
- Gen AI embedded across Gmail, Docs, Sheets, Slides, and Meet to draft, summarize, and generate content within everyday productivity tools. Targets workforce productivity, not enterprise search or custom agents.
- Workspace vs Enterprise — The #1 Exam Trap
- Workspace = productivity inside Google's office apps. Enterprise = search, research, and custom agents over an organization's own data. Never pick Workspace when the scenario needs enterprise search or custom agent building.
Customer Engagement Suite and Developer Tooling
- Conversational Agents
- Customer-facing virtual agents (chatbots/voicebots) that interact directly with customers to resolve requests. Part of the Customer Engagement Suite.
- Agent Assist
- Real-time assistance for human contact-center agents — suggests responses and surfaces knowledge while a person handles the conversation. Does not talk to customers directly.
- Conversational Insights + CCaaS
- Conversational Insights analyzes contact-center interactions for trends/quality. CCaaS (Contact Center as a Service) is the cloud-delivered contact-center infrastructure. Both round out the Customer Engagement Suite alongside Conversational Agents and Agent Assist.
- Agent Search
- Enterprise search capability on the Agent Platform that lets agents/apps retrieve grounded answers from an organization's own data (and optionally external sources like Google Search). Google's primary search-based grounding offering.
- Agent Tooling: Extensions, Functions, Data Stores, Plugins
- Extensions and functions call external APIs/services so an agent can take action. Data stores supply grounding data. Plugins extend capabilities further. Together they let an agent go beyond text generation to retrieve data and act.
- Google AI Studio vs Agent Studio
- Google AI Studio = fast, browser-based prototyping of prompts/apps with Gemini — great for quick experimentation. Agent Studio (on the Agent Platform) = enterprise environment for building, testing, and deploying production agents with governance.
- Document AI vs Cloud Vision
- Document AI parses and extracts structured fields (invoice numbers, totals, dates) from scanned documents/PDFs; Cloud Vision analyzes general image content like objects, faces, and labels — match the API to document vs. general image tasks.
- Speech, Translation, Video Intelligence, and Natural Language APIs
- Speech-to-Text/Text-to-Speech convert between audio and text; Translation converts between languages; Video Intelligence analyzes video content; Natural Language analyzes text sentiment and entities — none require model training.
Prompt Engineering Techniques
- Zero-Shot Prompting
- The prompt includes no examples — just an instruction. The model relies entirely on pre-trained knowledge to respond.
- One-Shot Prompting
- The prompt includes exactly one example to demonstrate the desired format or behavior before the actual task.
- Few-Shot Prompting
- The prompt includes several (two or more) examples to guide format, tone, or classification behavior. Count the examples in the scenario to identify zero/one/few-shot.
- Role Prompting
- Assigns the model a persona or role (e.g., 'You are an expert financial analyst') to shape the tone, expertise level, and framing of its response.
- Prompt Chaining
- Breaks a complex task into a series of linked prompts, where the output of one prompt feeds into the next — useful for multi-step workflows.
- Chain-of-Thought (CoT) Prompting
- Asks the model to reason step by step before producing a final answer, improving accuracy on multi-step reasoning or math problems. Distinct from ReAct.
- ReAct (Reason + Act)
- Combines step-by-step reasoning with taking actions using external tools (e.g., search, calculators, APIs) in an interleaved loop. Do not confuse with plain chain-of-thought, which only reasons — it does not act.
Grounding, RAG, and Model Limitations
- Foundation Model Limitations
- Data dependency (only as good as training data), knowledge cutoff (unaware of events after training), bias and fairness issues, hallucinations (confident but false output), and edge cases. Every mitigation question assumes you can name the limitation first.
- Grounding
- Connects a model's responses to trusted data sources so answers are factual and current, reducing hallucinations and overcoming the knowledge cutoff. The most heavily tested Domain 3 concept.
- RAG (Retrieval-Augmented Generation)
- A specific grounding method that retrieves relevant documents/context at query time and feeds them to the model before it answers — makes responses factual and current without retraining the model.
- Three Grounding Data Sources
- First-party enterprise data (a company's own documents/systems), third-party data (licensed or partner data), and world data (broad public knowledge, e.g., via Grounding with Google Search).
- Google Cloud Grounding Offerings
- Prebuilt RAG with Agent Search, RAG APIs (build custom retrieval pipelines), and Grounding with Google Search (ground answers in live web results).
- Human-in-the-Loop (HITL)
- Requiring a person to review/approve model output before it takes effect — the right mitigation for high-stakes, high-risk decisions where errors are costly, even after grounding and prompt engineering are applied.
- Continuous Monitoring and Evaluation
- Ongoing practices to keep a deployed model reliable: automatic model upgrades, KPIs, security patches, versioning, performance tracking, drift monitoring, and Feature Store usage.
Fine-Tuning, Prompt Tuning, and Sampling Parameters
- Fine-Tuning
- Retrains a foundation model on a niche, labeled dataset to permanently change its behavior or specialized style. More effort/cost than prompting. Does not add new factual knowledge the way grounding does and does not guarantee factual accuracy.
- Prompt Engineering vs Fine-Tuning — Exam Trap
- 'Quick', 'no training data', 'just improve responses' → prompt engineering. 'Must adopt a specialized style/domain behavior prompting cannot achieve' → fine-tuning. If the problem is hallucinations or outdated facts, the answer is grounding/RAG, not fine-tuning.
- Temperature
- Controls randomness/creativity of output. Low temperature = more deterministic, consistent, factual output (good for FAQ bots). High temperature = more creative, varied output (good for brainstorming/creative writing).
- Top-P (Nucleus Sampling)
- Limits the model's next-token choices to the smallest set of tokens whose cumulative probability exceeds P — controls output diversity by cutting off unlikely tokens, working alongside temperature.
- Top-K
- Limits token selection to the K highest-probability tokens at each generation step, applied before top-p and temperature narrow the choice further; tested less often on the exam than temperature and top-p.
- Token Count / Output Length
- Sets the maximum length of the generated response. Affects output length only — has no effect on determinism or consistency (that is temperature's job).
- Safety Settings
- Configurable filters that block or flag harmful, unsafe, or policy-violating content in model output. Should never be disabled to 'improve' output quality.
- generation_config = { "temperature": 0.2, "top_p": 0.95, "max_output_tokens": 1024, } response = model.generate_content(prompt, generation_config=generation_config)
- Example Vertex AI Gemini API call showing temperature, top-p, and max output tokens configured together to produce more deterministic, length-bounded output.
Responsible AI, SAIF, and Security
- Five Responsible AI Principles
- Transparency (being open about how/when AI is used), privacy (protecting sensitive data), fairness (avoiding biased outcomes), accountability (owning AI decisions/outcomes), explainability (understanding why a model produced an output). Broader than bias alone.
- Transparency vs Explainability
- Transparency = openness about how and when AI is used in a process. Explainability = the ability to understand and articulate why a model made a specific decision. The exam treats these as distinct principles.
- Secure AI Framework (SAIF)
- Google's conceptual framework for securing AI systems against attacks and misuse across the ML lifecycle (data, models, infrastructure). A security framework — not the same as Responsible AI, which is about ethics/fairness.
- Anonymization vs Pseudonymization
- Anonymization irreversibly removes identifying information — cannot be re-linked. Pseudonymization replaces identifiers with reversible tokens — can be re-linked by authorized parties. Both are privacy techniques but are not identical.
- IAM and Security Command Center
- IAM (Identity and Access Management) enforces least-privilege access control over who can use which resources. Security Command Center is Google Cloud's centralized platform for detecting threats, misconfigurations, and vulnerabilities. Both support SAIF.
Business Strategy and Measuring Impact
- Steps to Implement a Transformational Gen AI Solution
- 1) Identify the business need. 2) Choose the right solution (match business requirements + technical constraints). 3) Integrate it into the organization (change management, training). 4) Measure impact (KPIs, ROI).
- Measuring Gen AI Impact
- Define KPIs tied to business outcomes (resolution time, cost savings, customer satisfaction, adoption rate) and track them over time — not just technical metrics like model accuracy or prompt volume.
- Types of Gen AI Solutions to Match to Needs
- Text generation, image generation, code generation, and personalized user experiences — a leader selects among these based on business requirements and technical constraints, not just novelty.
- Bias, Data Quality, and Fairness
- Poor data quality and unrepresentative training data lead to biased, unfair AI outcomes. Leaders are accountable for auditing data and outcomes for fairness before and after deployment.