Generative AI Leader · Glossary
Generative AI glossary for the Generative AI Leader exam
· approx. 39 min read
Independent study aid – not by Google, not affiliated with Google.
This glossary explains 293 terms from the Google Cloud Generative AI Leader exam, each with a short definition, the matching Exam Guide section and – for products – the former name.
A
- Accountability
Clear responsibility for what an AI system does and decides. A prerequisite for fairness and trust.
- Accuracy (data quality)
Quality attribute: the data is factually correct; faulty data leads to wrong patterns.
- Active prompting
CoT variant in which the model may ask follow-up questions or request additional information.
- Adoption champion
A designated person who drives and coordinates gen AI adoption across the organization.
- Adversarial attack
Deliberately manipulated inputs or data intended to make a model misbehave.
- Agent (gen AI agent)
An application that pursues a goal by observing the world and acting with the tools available to it; the focus is on autonomous action.
- Agent (gen AI agent)
An application that pursues a goal by observing its environment and acting on it with tools. Its core is a reasoning loop.
- Agent Assist
Supports human contact center agents in real time with suggested responses, knowledge articles, live transcription (speech-to-text), translation and summaries.
- Agent components
A gen AI agent consists of a model (the brain), tools (interaction with the environment) and a reasoning loop (iterative decision-making).
- Agent Development Kit (ADK)
Open-source framework for creating agents and multi-agent systems in code. Model-agnostic.
- Agent Platform (Gemini Enterprise Agent Platform)
Google Cloud's unified AI/ML platform for building, training, tuning, deploying and managing models and gen AI applications.
- Agent Platform (Gemini Enterprise Agent Platform)
Google's unified, fully managed platform for creating, tuning, deploying and operating ML models, gen AI apps and agents.
- Agent Platform AutoML
Agent Platform service for training your own models (image, video, tabular) without deep ML expertise.
- Agent Platform AutoML
Trains custom models for image, video and tabular data with minimal ML expertise.
- Agent Platform Feature Store
Central repository for sharing, serving and reusing ML features. Ensures consistency between training and production.
- Agent Platform Model Registry
Manages model versions, changes and the model lifecycle.
- Agent Platform Pipelines
Orchestrates and automates ML workflows end to end, from data preparation to deployment.
- Agent Runtime
Managed runtime environment in which your own agents run at scale in production (with Sessions and Memory Bank).
- Agent Search
Google's search and recommendation solution on Agent Platform; out-of-the-box RAG that grounds answers in your own data, curated third-party data and Knowledge Graph data.
- Agent Search
Search and recommendations solution on Agent Platform. Delivers Google-quality search over your own data with grounding (prebuilt RAG), including for external customers.
- Agent Studio
Tool in Agent Platform (Cloud Console) for prompt design, testing, tuning and grounding with enterprise security and governance; intended for production.
- Agent Studio
Console tool in Agent Platform for prompt testing, tuning and low-code agent design. With enterprise security and cloud governance for production.
- Agents layer
Layer of the gen AI stack that provides capabilities and actions (search, book, recommend) and sits between Applications and Platform.
- AI Hypercomputer
Google Cloud's supercomputing architecture of integrated hardware and software for AI workloads at the largest scale.
- AI Principles (Google)
Guidelines published by Google in 2018 for the responsible development and use of AI.
- AI-first approach
At Google, AI is at the core of product strategy, backed by decades of research.
- Anonymization vs. pseudonymization
Anonymization irreversibly removes the link to a person. Pseudonymization replaces identifiers and can be reversed with a key.
- API Library (Google Cloud)
Catalog of all Google APIs (Maps, Workspace, YouTube ...) that agents can use as tools.
- Apps Script
Low-code platform (JavaScript) for automating Google Workspace and creating Gemini add-ons.
- AppSheet
No-code platform that lets business users create business apps, including with Gemini features.
- Artificial intelligence (AI)
Umbrella term for machines that perform tasks that normally require human intelligence.
- Artificial neural network (ANN)
Computational model made of interconnected "neurons" arranged in layers; the foundation of deep learning.
- Audio Overview
Podcast-style audio summary of the sources, generated by Gemini Notebook.
- Audit trail / traceability
Traceable logging of how and why a system made a decision. Often mandatory in regulated industries.
- Augmentation (RAG step)
Second RAG step: the retrieved information is added to the prompt together with the user's question.
- Augmentation (vs. automation)
AI supports people, who interpret and decide; the counterpart to automation, where AI takes over repetitive tasks entirely.
- Augmentation vs. automation
Augmentation supports people, who keep making the decisions. Automation takes over repetitive or rule-based tasks entirely. Augmentation is often the intermediate step.
- Authentication vs. authorization
Authentication verifies who someone is. Authorization verifies what they are allowed to do. IAM covers both.
- Automatic model upgrades
The provider continuously improves the models; users benefit automatically but must retest their workflows.
- Availability (data)
Accessibility attribute: the required data must be available in both training and production.
- Availability and reliability (model)
Model selection criterion: uptime guarantees, redundancy and disaster recovery for production use.
B
- Bias
Systematic distortion in data or model outputs; even subtle biases can be amplified in the outputs.
- Bias and fairness
AI picks up biases from data and processes and can make unfair decisions. Bias should be detected and mitigated; fairness is central.
- BigQuery
Google's serverless data warehouse for analytics and data preparation; BigQuery ML lets you train models using SQL.
- BigQuery universal catalog
Service for data governance and metadata management across the Google Cloud data ecosystem.
- Build-vs-buy ladder (solution choice)
Order from low to high effort: ready-made app → model plus prompting → grounding/RAG → fine-tuning → AutoML → custom model. Enter as early on the ladder as possible.
C
- CCaaS (Google Cloud Contact Center as a Service)
Cloud-native contact center solution with omnichannel support, CRM integration and managed infrastructure.
- Chain-of-thought (CoT)
Prompting technique in which the model reasons through intermediate steps. Good for internal reasoning.
- Chain-of-thought (CoT) prompting
Prompting technique that guides the model through a solution path using examples of explicit intermediate steps; improves reasoning and traceability.
- Chirp
Google's foundation model for audio and speech recognition in many languages.
- Cloud Run
Serverless platform for stateless containers. Good for tools with their own dependencies.
- Cloud Run functions
Serverless, event-driven functions that serve as specialized agent tools.
- Cloud SQL / Spanner
Google Cloud's relational database services for structured data; Spanner scales globally.
- Cloud Storage
Google Cloud's scalable object storage, typically for unstructured data such as images, PDFs or audio.
- Cloud Storage
Scalable object storage for unstructured data, e.g. as a source for search and RAG.
- Cloud Video Intelligence API
Pre-trained API that analyzes video content, e.g. for search, recommendations and media analysis.
- Cloud Vision API
Pre-trained API for image analysis: objects, text, faces, landmarks, moderation.
- CMEK (customer-managed encryption keys)
Google encrypts the data, the customer manages the keys. Usually for compliance reasons.
- Completeness
Quality attribute: sufficient data volume and coverage for reliable predictions.
- Comprehensive AI ecosystem
Gen AI is integrated into many Google products (Search, Workspace, Gemini app, Google Cloud), so AI arrives in familiar tools and adoption becomes easier.
- Connectors
Links from Gemini Enterprise to enterprise systems for reading data and triggering actions.
- Consistency
Quality attribute: uniform formats and labels across all sources; inconsistency confuses the model.
- Content moderation (HITL)
Humans review user-generated content in context and catch harmful material that algorithms miss.
- Content moderation policy
Rules and review processes defining which generated content is permitted.
- Context window
The amount of information (in tokens) a model can take into account at once; a larger window increases costs.
- Continuous monitoring and evaluation
Ongoing monitoring of model quality, drift, security and business impact across the entire lifecycle.
- Continuous training (CT)
Automated, regular retraining in a pipeline (MLOps); worthwhile mainly when data changes frequently, because every retraining run costs compute.
- Conversational agent
Agent for dialogue (answering questions, chat, access to information); technically simpler but directly visible to customers.
- Conversational Agents
Chatbots and voicebots that talk to customers on their own. Deterministic, generative or hybrid, created with CX Agent Studio.
- Conversational Insights
Analyzes all customer conversations for topics, sentiment and quality, including Generative FAQ.
- Conversational vs. workflow agents
Conversational agents hold dialogs (answering questions). Workflow agents carry out multi-step processes (onboarding, fulfillment).
- Cost (data)
Accessibility attribute: acquiring and labeling high-quality data can be very expensive.
- Cost (gen AI)
Costs arise for training, deployment and usage. Billing is by tokens, characters, requests or compute time. Larger models and context windows cost more.
- Create / Summarize / Discover / Automate
The four main uses of gen AI: generating content, condensing it, finding information and automating tasks.
- Creative matrix
Matrix-based ideation method: gen AI products are crossed with business goals to identify use cases.
- Custom training
Your own model with full control over architecture and training on Agent Platform.
- Customer data control
Customer data belongs to the customer; prompts and data in enterprise services are not used to train Google's models.
- Customer Engagement Suite
Google's suite for AI-powered customer interaction: Conversational Agents, Agent Assist, Conversational Insights, CCaaS.
D
- Data accessibility
Data must be available, affordable and in a suitable format; this includes holding the necessary rights.
- Data dependency
Limitation: model performance depends heavily on the quality and completeness of the training data.
- Data governance
Rules and processes for the quality, access, protection and use of data.
- Data ingestion
First lifecycle phase: collecting and ingesting suitable data (Pub/Sub, Cloud Storage, Cloud SQL, Spanner).
- Data poisoning
Manipulation of training data to corrupt model behavior.
- Data preparation
Lifecycle phase: cleaning, transforming, formatting and labeling raw data (BigQuery, universal catalog).
- Data quality
Fitness of the data for the task: accurate, complete, representative, consistent, relevant; the standard depends on the context.
- Data store
Structured or unstructured knowledge base (websites, databases, documents) that agents and RAG access.
- Data stores
Agent tool that provides access to information (knowledge bases, websites, databases). The foundation for RAG.
- Databases (Cloud SQL, Spanner, Firestore)
Managed databases in which agents store structured data, user data and progress.
- Deep learning (DL)
Subfield of ML that uses multi-layer neural networks; the engine behind gen AI.
- Democratization of AI
Tools that need little or no coding, models that come already trained, and ready-made APIs, so that non-developers can also create AI solutions.
- Deterministic agent
Agent with fixed rules and paths. The same input yields the same output; no foundation model.
- Diffusion model
Generative model that creates images, audio or video through iterative refinement, e.g. Imagen.
- Document AI API
Pre-trained API that extracts structured data from documents and automates processing.
- Document Translation API
Translates formatted documents while preserving the layout.
- Drift (drift monitoring)
Change in data or relationships over time that degrades model performance; monitored with Model Monitoring.
- Drift / skew
Production data diverges from the training data and model performance drops. Requires continuous monitoring.
E
- Edge AI / edge computing
AI runs on devices or servers close to the data source; benefits are low latency, privacy and offline capability.
- Edge cases
Limitation: rare, atypical situations expose model weaknesses and lead to unexpected errors.
- Edge computing / AI on the edge
AI runs on devices close to the data source. Real-time, privacy-friendly, works offline.
- Embeddings
Numerical vectors that represent the meaning of text, images and more, enabling semantic search.
- Enterprise-ready AI platform
Platform with enterprise-grade security, privacy, governance, reliability and scalability.
- Explainable AI / explainability
The ability to justify, in a traceable way, why a model arrives at a result.
- Extensions
Agent tool that connects external services via APIs and standardizes the interaction.
F
- Fairness
Goal and evaluation standard for equitable outcomes across groups; fairness assessments have inherent limits.
- Feature Store (Agent Platform Feature Store)
Central repository for sharing, serving and reusing ML features; ensures consistency between training and production.
- Few-shot prompting
Prompt with several examples from which the model picks up pattern, format and tone.
- Fine-tuning
Further training a pre-trained model on a task-specific dataset; expensive, which is why it is the last step after prompting and grounding.
- First-call resolution (FCR)
KPI: share of customer issues resolved at the first contact.
- First-party enterprise data
The company's own data (documents, databases, policies) in which the model is grounded.
- Format (data)
Accessibility attribute: data must be in a format the model can process.
- Foundation model
Large model pre-trained on vast, diverse data that can be adapted to many tasks; the basis of gen AI.
- Foundation model as the agent's "brain"
The model is the agent's reasoning and decision-making component. Tools are its hands.
- Fully managed
Google runs the infrastructure, scaling and monitoring. The customer only configures.
- Functions
Agent tool that defines specific actions or tasks for the reasoning system to choose from.
G
- Garbage in, garbage out
Principle: poor input data inevitably leads to poor results, including in RAG knowledge bases.
- Gemini
Google's closed, multimodal foundation model family for understanding, conversation, content creation and Q&A.
- Gemini Advanced
Advanced Gemini tier with the most capable models for coding and reasoning, priority access to new features and enterprise protections.
- Gemini API
Programming interface to Gemini models. Integrates into any application via HTTP.
- Gemini app
Google's gen AI chatbot (web and mobile) for writing, summarizing, translating and generating images.
- Gemini app (vs. Gemini model)
The Gemini app is Google's chatbot (the application); Gemini is the underlying model.
- Gemini Code Assist
AI pair programmer offering code suggestions, code generation and explanations.
- Gemini Enterprise
Central, secure agent platform for employees: search across all company data, assistants with connectors, custom agents, Notebook Enterprise.
- Gemini Enterprise for Customer Experience (CX)
Current umbrella name for the customer service suite (CX agents, Agent Assist, CX Insights, CCaaS).
- Gemini for Google Cloud
AI assistant in Google Cloud tools (Code Assist, Cloud Assist, BigQuery, Looker, Security ...).
- Gemini for Google Workspace
Gen AI directly in Gmail, Docs, Sheets, Slides, Meet and Vids, e.g. via the side panel.
- Gemini Nano
Google's most compact Gemini model for on-device execution; benefits: privacy, speed, offline use.
- Gemini Nano
Smallest Gemini model for on-device use: private, fast, offline.
- Gemini Notebook
Research assistant that grounds exclusively in uploaded sources (summaries, Q&A, audio overviews).
- Gemini Notebook
AI research partner grounded exclusively in uploaded sources. Provides summaries, Q&A, quizzes and Audio Overview.
- Gemini Notebook API
Programmatic access to Gemini Notebook (Enterprise) for embedding notebooks into business processes.
- Gemini Notebook Enterprise
Enterprise edition with compliance, admin features, IAM and data kept in your own cloud project.
- Gemini Pro vs. Flash / Flash-Lite
Model variants: Pro for complex tasks (slower, more expensive), Flash/Flash-Lite for fast, low-cost responses.
- Gemini side panel
Side panel in Workspace apps through which Gemini summarizes and generates content in context.
- Gemma
Family of lightweight open Google models (open weights) for local deployment and extensive customization.
- Gems
Personalized Gemini assistants with fixed instructions; deliver personalized answers, simplify workflows and keep contexts separate.
- Gems
Personalized Gemini assistants with their own instructions and files. For recurring tasks; each chat is isolated.
- Gen AI applications (layer)
Top, user-facing layer (front end) through which people use gen AI, e.g. the Gemini app or Workspace with Gemini.
- Gen AI landscape
Five-layer model: Infrastructure, Models, Platform, Agents, Applications.
- Generalization
A model's ability to apply what it has learned to new data not seen during training.
- Generation (RAG step)
Third RAG step: the LLM generates the answer from the augmented prompt, ideally with sources.
- Generative agent
Agent based on a foundation model. Flexible and natural-language driven, but may give varying answers.
- Generative AI (gen AI)
Subfield of ML that creates new content such as text, images, audio, video or code instead of only analyzing it.
- Generative FAQ
Feature in CX Insights that identifies frequent customer questions and how effective the answers are.
- Google AI Studio
Free web tool for rapid prototyping with Gemini and the Gemini Developer API; subject to usage limits.
- Google AI Studio
Free web tool for rapid prototyping with Gemini and the Gemini Developer API. For beginners and experimentation.
- Google Search (grounding with)
Connects Gemini to current web information, overcoming the knowledge cutoff.
- GPU
General-purpose parallel processor that is also used for AI training and inference.
- GPU / TPU
Specialized processors for parallel AI computation; TPUs are Google's chips purpose-built for AI.
- Grounding
Connecting model output to verifiable, specific sources of information; umbrella term that includes RAG.
- Grounding with Google Search
Connects Gemini to current public web information (world data), e.g. for questions about events after the knowledge cutoff.
H
- Hallucination
Model output that sounds plausible but is false or made up; grounding reduces it, but it can never be ruled out entirely.
- Human in the loop (HITL)
Human input and feedback are integrated directly into ML workflows, especially for sensitive or high-stakes decisions.
- Human in the loop (HITL)
Human judgment is deliberately built into the ML process, e.g. for moderation, sensitive areas and reviews before or after generation.
- Hybrid agent
Combines deterministic control with generative flexibility. Google's recommended approach for CX agents.
- Hypercomputer
Supercomputing cluster of many nodes for training and running large models (Google: AI Hypercomputer).
I
- IAM (Identity and Access Management)
Central access control in Google Cloud: identities, roles, permissions, auditing.
- IC (individual contributor)
Specialist without management responsibility, close to workflows and pain points. Drives the bottom-up part of the strategy.
- Identity and Access Management (IAM)
Google Cloud service for authentication and authorization: roles, permissions, auditing.
- Imagen
Google's text-to-image diffusion model for high-quality images from text descriptions.
- In-context learning
The model learns from examples in the prompt without its weights changing.
- Infrastructure (layer)
Bottom layer: compute (GPU/TPU), storage, networking and the software for training and running models.
- Interpretability
How well people can understand how a model works. Related to explainability.
K
- Key performance indicator (KPI)
Metric for the business impact of gen AI; approach: define metrics, collect data, iterate.
- Knowledge cutoff
The date up to which a model has seen training data; it knows nothing about later events.
- KPI (key performance indicator)
Metric used to tie the impact of a gen AI solution to business goals.
L
- Labeled data
Data with an assigned tag or target value (name, type, number); the basis for supervised learning.
- Large language model (LLM)
Foundation model for understanding and generating language (including code); all LLMs are foundation models.
- Least privilege (principle of)
Grant only the minimum permissions required and review them regularly.
- Limitations of foundation models
Data dependency, knowledge cutoff, bias, fairness, hallucinations and edge cases.
- LiteRT (Lite Runtime)
Google's runtime for running ML models efficiently on edge and mobile devices.
- LiteRT (Lite Runtime)
Runtime that lets ML models run efficiently on mobile and edge devices.
M
- Machine learning (ML)
Subfield of AI in which systems learn from data instead of being explicitly programmed.
- Mandiant
Google's threat intelligence and incident response expertise for proactive protection, including of AI systems.
- Metaprompting
Prompting that has the model generate, improve or interpret other prompts.
- ML lifecycle
Phases: data ingestion, data preparation, model training, model deployment and model management, each with Google Cloud tools.
- MLOps
Practices and tools for operating, automating, versioning and monitoring ML models.
- MLOps
Practices and tools for operating ML models reliably throughout their lifecycle.
- Modality
The data type a model processes or generates (text, image, audio, video); the first criterion in model selection.
- Model
A mathematical structure learned from data that turns inputs into outputs.
- Model deployment
Lifecycle phase: making a trained model available for predictions (Agent Platform endpoint).
- Model Evaluation
Assesses and compares model performance.
- Model Evaluation (Agent Platform)
Service for assessing and comparing model performance (performance tracking).
- Model Garden
Catalog in Agent Platform for discovering, customizing and deploying Google, open and third-party models as well as pre-trained APIs.
- Model Garden
Central catalog of more than 200 models (Google, open, third-party) for discovering, testing, tuning and deploying.
- Model management
Lifecycle phase: versioning, monitoring, maintaining and retraining models over time.
- Model Monitoring
Service that detects performance degradation, input skew and drift and triggers updates or retraining.
- Model Monitoring
Monitors production models for performance degradation, skew and drift and triggers retraining.
- Model Registry
Central management of model versions, changes and lifecycle.
- Model selection factors
Criteria for choosing a model: modality, context window, security, availability/reliability, cost, performance, fine-tuning/customization, ease of integration.
- Model training
Lifecycle phase: learning model parameters from data (Agent Platform: custom training or AutoML); for foundation models, usually tuning.
- Models (layer)
Layer of the stack containing the models, the "brain" of agents and applications.
- Multi-agent system
Several specialized agents work together toward a complex goal; one agent can serve as a tool for another.
- Multi-agent system
Several specialized agents work together. One agent can be a tool of another.
- Multi-directional strategy
Combination of leadership's top-down vision and employees' bottom-up ideas.
- Multimodal (foundation) model
Model that processes or generates several data types at once; multimodality means combining modalities.
- Multimodal CoT
CoT variant that combines text with images or video to improve reasoning.
- Multimodal search
Search across text, images and other formats in multiple enterprise systems, grounded in company data.
- MVP (minimum viable product)
The smallest production-ready version that you ship and improve iteratively.
N
- Nano Banana
Marketing name of the Gemini image models for image generation and editing (e.g. Nano Banana Pro); alongside Imagen, which the Exam Guide names, Google's second line of image models.
- Natural Language API
Pre-trained text analysis: sentiment, entities, classification.
- Natural language processing (NLP)
Subfield of AI concerned with understanding, processing and generating human language; LLMs are the dominant technique today.
- Needs (scale, customization, ...)
Requirements that drive the choice of solution: scale, customization, user interaction, privacy, latency, connectivity, accuracy, explainability.
O
- One-shot prompting
Prompt with exactly one example for guidance.
- Open approach
Openness to open-source frameworks, open models and third-party models. Result: freedom of choice without lock-in.
- Output length (max output tokens)
Sampling setting for the maximum length of the generated response.
P
- Performance tracking
Ongoing checks of model metrics to detect quality loss early.
- PII (personally identifiable information)
Data that can be used to identify a person. Requires special protection.
- Pipelines (Agent Platform Pipelines)
Orchestration and automation of ML workflows from data preparation to deployment.
- Plan for gen AI integration (6 steps)
Vision → high-impact use cases → capabilities → organizational change → measure value → responsible AI.
- Plan for impact
Define metrics → collect and analyze data → iterate and improve.
- Platform (layer)
Layer with APIs, data management, deployment and MLOps tools between models and agents; at Google: Agent Platform.
- Plugins
Agent tool that adds new skills and integrations with other platforms.
- Post-generation review
Form of HITL: continuous human review and feedback after deployment.
- Pre-generation review
Form of HITL: experts review ML outputs before deployment and catch errors or bias.
- Prebuilt RAG with Agent Search
Ready-made, managed RAG solution (data connection, search, grounding, citations) requiring little effort.
- Privacy
Protection of personal data, e.g. through anonymization and pseudonymization. Preventing models from leaking training data.
- Prompt / Prompting
Input (text, file, image) used to steer a foundation model and trigger an output.
- Prompt chaining
Step-by-step prompts that build on one another, or a conversation in the same thread, to reach a refined result.
- Prompt engineering
The art and science of designing effective prompts; does not change the model.
- Prompt engineering (significance)
The fastest and cheapest lever for better outputs; must be retested continuously with new model versions.
- Prompt tuning
Parameter-efficient tuning: only learnable "soft prompts" are trained, while the base model stays frozen.
- Pub/Sub
Google Cloud messaging service for ingesting real-time data streams (publish/subscribe).
R
- RAG (retrieval-augmented generation)
Grounding method: retrieve relevant information, insert it into the prompt, then generate the answer.
- RAG APIs
Building blocks (parsing, embeddings, Vector Search, ranking, grounded generation, check grounding) for your own, flexibly controlled RAG pipeline.
- RAG APIs
Building blocks (RAG Engine, Vector Search, grounding APIs) for your own RAG pipelines with more control.
- RAG Engine
Managed service on Agent Platform for creating your own RAG pipelines.
- ReAct (Reason + Act)
Prompting framework in which the LLM reasons in a Think, Act, Observe, Respond cycle and acts with tools; reduces hallucinations by drawing on external information.
- ReAct (reason + act)
Agent framework: reason, act (tool), observe, respond, in a loop. For tasks that involve external interaction.
- Reasoning loop
Iterative process in which an agent observes, reasons, chooses a tool, acts and evaluates the result.
- Reasoning loop
The iterative core of an agent: observe, reason, choose a tool, act, evaluate the result, repeat.
- Recommendations (Agent Search)
Personalized product or content recommendations for more engagement and discovery (media, commerce).
- Red teaming
Deliberate attacks on your own system, including with a second model, to find vulnerabilities.
- Reinforcement learning (RL)
Learning through interaction with an environment and feedback (reward/penalty) in order to maximize the total reward.
- Relevance
Quality attribute: the data fits the task the model is meant to perform.
- Representativeness
Quality attribute: the data covers all relevant groups and situations; if it does not, bias results.
- Resources (people, money, time)
Organizational constraints of a gen AI project: expertise, budget, timeline.
- Responsible AI
Ensuring that AI is used ethically and does no harm, across the entire lifecycle.
- Retrieval (RAG step)
First RAG step: retrieving relevant information with tools (data stores, vector databases, search engines, knowledge graphs).
- Retrieval-augmented generation (RAG)
Grounding architecture: retrieve relevant information (retrieval), insert it into the prompt (augmentation), generate the answer (generation), optionally iterate.
- Reusable prompts / prompt templates
Saving proven prompts as templates with placeholders and reusing them.
- RLHF
Reinforcement learning from human feedback: post-training a model based on human ratings.
- ROI (return on investment)
Ratio of benefit to cost. A key criterion for prioritizing use cases.
- Role prompting
Assigning the model a role or persona to steer the style, tone and focus of its answers.
- Roles (business leaders, developers, AI practitioners)
Business leaders use ready-made apps. Developers create agents and integrations. AI practitioners customize and operate models.
S
- Safety settings
Settings that filter potentially harmful or inappropriate content from the output (categories and thresholds).
- SAIF (Secure AI Framework)
Google's framework of security standards and best practices for AI systems. A guide, not a product.
- Sampling parameters
Settings that control model behavior: token count, temperature, top-p (top-K), safety settings, output length.
- Saved Info
Gemini remembers specific information about the user that has been deliberately saved, without using the entire chat history.
- Saved info (Gemini)
Persistent information stored in Gemini that the model takes into account consistently, without using the whole chat history.
- Secure AI
Protecting AI systems against attacks, misuse, data leakage and manipulation.
- Secure-by-design infrastructure
Infrastructure with security designed in from the ground up (e.g. automatic encryption at rest).
- Security across the ML lifecycle
Security measures in every phase: collecting, preparing, training, deploying, managing.
- Security Command Center (SCC)
Central dashboard for security posture, vulnerabilities and threats in Google Cloud.
- Security patches and updates
Ongoing security updates to keep AI systems secure and up to date.
- Self-consistency
CoT variant: generate several solution paths and choose the most consistent one.
- Semantic search
Searching by meaning rather than exact keywords; the core of retrieval in RAG.
- Semi-/self-supervised learning
Hybrid forms in which the model partly generates its own learning signals from unlabeled data; typical in pre-training foundation models.
- Semi-structured data
Data with tags/keys but no rigid table schema (JSON, XML, logs).
- Sensitive Data Protection
Service that discovers sensitive data (PII) and de-identifies it, e.g. through masking or anonymization.
- Serverless
The provider fully manages the servers, scales automatically and bills by usage.
- Six strategy factors
Strategic focus, exploration, responsible AI, resourcing, impact, continuous improvement, each from the leadership and the IC perspective.
- Solutionism
The mistake of deploying a technology without first clarifying the business problem.
- Speech-to-Text API
Pre-trained API that converts spoken language into text (calls, meetings, video).
- Structured data
Organized, easily searchable data, often in relational databases or tables.
- Supervised learning
Learning from labeled examples to predict outputs for new inputs (classification, regression).
T
- Technique ladder
Memory aid: prompt engineering first, then grounding/RAG, fine-tuning last; HITL and monitoring throughout.
- Temperature
Sampling parameter for randomness or "creativity": low = focused and repeatable, high = varied.
- Text-to-Speech API
Pre-trained API that converts text into natural-sounding speech.
- Third-party data
External, curated or licensed data from third-party providers in which the model is grounded.
- Token
Unit of text (roughly 4 characters); the unit of measure for context windows, traffic and billing.
- Token
Small unit of text (~4 characters). Measure of context size and the billing unit for LLMs.
- Token count
Number of tokens processed; models have a token limit, and tokens determine costs.
- Tool (agent tool)
Function through which an agent interacts with its environment: retrieving data, taking actions.
- Top-K
Sampling parameter: selection only from the K most likely tokens.
- Top-p (nucleus sampling)
Sampling parameter: selection from the most likely tokens until their cumulative probability reaches p.
- TPU (Tensor Processing Unit)
Accelerator designed in-house by Google and optimized for AI.
- Traditional AI model
Model trained for a single purpose on specific data (e.g. a spam filter); a specialist rather than a generalist.
- Training-serving skew
Difference between the data distribution in training and in production; detected by Model Monitoring.
- Transformer
Neural network architecture introduced by Google in 2017; the foundation of today's LLMs (the "T" in GPT).
- Translation API
Pre-trained API that translates text, documents, websites, and audio and video.
- Transparency
Disclosing that and how AI and data are used, so that affected people can understand the system.
U
- Unlabeled data
Raw data without a tag or target value; the basis for unsupervised learning and for pre-training foundation models.
- Unstructured data
Data without a predefined structure, such as free text, emails, images, audio and video.
- Unsupervised learning
Learning from unlabeled data to find groups, patterns or outliers (clustering, anomaly detection).
- Use-case prioritization
Rank use cases by business value/ROI, feasibility, cost and risk. Pick a first one with low risk but visible benefit.
V
- Vector database
Database that stores embeddings and searches them by similarity; a typical RAG component.
- Vector Search
Service for semantic similarity search over embeddings. A building block for RAG.
- Veo
Google's model for generating video from text or still images.
- Versioning
Tracking and managing model versions, using Model Registry.
W
- Workflow agent
Agent that automates multi-step processes, e.g. order fulfillment, customer onboarding or research.
- Workload monitoring tools
Logging, monitoring and observability tools for operating AI systems securely and detecting anomalies.
- World data
Publicly available world knowledge, e.g. via Google Search or Google's Knowledge Graph.
Z
- Zero-shot prompting
Task without examples; the model relies solely on its prior knowledge.
Read next: Google Cloud Generative AI Leader · Product map