Generative AI Leader Exam Glossary - 122 Terms

Search the terminology pack for Generative AI Leader. Use these definitions with the study guide and practice questions.

Download App Study Guide Free Practice Exam

A

accessible data: Data that is readily available, usable, and in the proper format for model training needs.
agent: A piece of software that learns how to best achieve a goal based on inputs and tools available to it.
Agent Assist: A tool that supports live human contact center agents.
AI: Abbreviation for artificial intelligence.
AI on the edge: Running AI solutions on infrastructure such as devices or servers closer to where the action is happening.
AI/ML: An acronym used in the text to refer to artificial intelligence and machine learning in the context of model risks and security.
artificial intelligence: The field of building machines that can perform tasks typically requiring human intelligence, such as learning, problem-solving, and decision-making.
AutoML: A way to create and train models with minimal technical knowledge and effort.
availability and reliability: Model-selection factors describing whether a model is consistently available and performs reliably under load, including considerations such as uptime guarantees, redundancy, and disaster recovery mechanisms.

B

Bias: A limitation in which models learn patterns from large datasets that may contain bias, and even subtle biases can be magnified in outputs.
BigQuery: A Google Cloud data service that Gemini for Google Cloud can use to analyze data.

C

chain-of-thought: A prompting technique that guides an LLM through a problem-solving process by providing examples with intermediate reasoning steps.
Cloud Video Intelligence API: A Google Cloud API that analyzes video content and extracts meaningful information for uses such as content recommendation, video search, and media analysis.
Cloud Vision API: A Google Cloud API that analyzes image content, tags images based on detected objects and text, identifies faces and landmarks, and supports content moderation and visual search.
Contact Center as a Service (CCaaS): An enterprise-grade contact center solution that is native to the cloud and can underpin customer engagement tools.
Content moderation: A human-in-the-loop use case in which user-generated content is moderated contextually so harmful material that algorithms might overlook can be caught.
Context window: The amount of text a model can consider at one time.
Conversational Agents: Tools that act as chatbots for customers.
Conversational Insights: A tool for gaining insights into communications with customers.
CoT: An acronym for chain-of-thought, a prompting technique that guides an LLM through a problem-solving process by providing examples with intermediate reasoning steps.

D

Data dependency: A limitation in which foundation model performance relies heavily on the quality and completeness of the training data; biased or incomplete data affects outputs.
data ingestion and preparation: The process of collecting, cleaning, and transforming raw data into a usable format for analysis or model training.
Data stores: Tooling that provides access to information.
deep learning: A subset of machine learning that uses artificial neural networks with many layers to extract complex patterns from data.
deterministic (traditional): Agents built with predefined paths and actions.
Document AI API: A Google Cloud API that extracts data from varied formats, automates data capture and document processing, and can summarize documents.
Document Translation API: A Google Cloud API that translates formatted documents while preserving the original layout.
Drift monitoring: The practice of watching for changes in a model's accuracy over time.

E

Edge cases: Rare and atypical scenarios that can expose a model's weaknesses and lead to unexpected results.
Extensions: Tooling that connects to external services via APIs.

F

Fairness: A responsible-AI concept concerned with assessing whether generative AI models treat outputs equitably.
Few-shot: A prompting technique where the model is given multiple examples to learn from.
Fine-tuning: A technique used to enhance a pre-trained or foundation model's performance for specific tasks or domains.
Foundation model: A pre-trained model whose performance depends heavily on data and that can be adapted for specific tasks or domains.
foundation models: Large AI models trained on massive datasets, or massive amounts of unlabeled data, that can be adapted to many tasks and serve as the basis of generative AI.
Functions: Tooling that defines specific actions or tasks.

G

Gemini: A Google generative AI model that supports multimodal understanding, advanced conversational AI, content creation, and question answering.
Gemini Enterprise: A Google offering that integrates customized search and conversation agents into internal websites or dashboards so they can access and understand data from various internal sources.
Gemini Nano: Google's most efficient and compact AI model, specifically designed to run on devices.
Gemma: A Google model that offers developers a user-friendly, customizable solution for local deployments and specialized AI applications.
Gems: Personalized AI assistants within Gemini that provide tailored responses based on specific instructions and can streamline workflows such as templates, prompts, and guided interactions.
gen AI: Abbreviation for generative AI.
gen AI-powered application: The user-facing part of generative AI that allows users to interact with and leverage AI capabilities.
generative: Agents defined with natural language using LLMs to give a real conversational feel to a chatbot.
generative AI: A type of AI that can create new content and ideas. In the text, it is described as an application of machine learning focused on creating new content and as capable of multimodal processing and generation.
Generative AI Leader: Google Cloud training and certification track for professionals who need to understand how generative AI can be applied in a business context, especially using Google Cloud tools and services.
Google AI Studio: A free Google AI environment meant for quick AI prototyping.
Google Cloud: The cloud platform whose tools and services are referenced as the environment for applying generative AI in business contexts.
Google Workspace: A suite of Google apps that can be integrated with Gemini to compose emails in Gmail, generate images in Slides, and summarize notes in Meet.
Grounding: The practice of connecting an AI model's output to verifiable sources of information.

H

Hallucinations: Outputs produced by AI models that are not accurate or not based on real information.
High risk decision-making: A high-stakes use case where human review can help safeguard accuracy and accountability for machine learning model outputs.
HITL: An acronym for humans in the loop, meaning human input and feedback are directly integrated into machine learning workflows.
Humans in the loop: A process in which human input and feedback are directly integrated into machine learning workflows.
hybrid agents: Agents that combine deterministic and generative capabilities.

I

IAM: Acronym for Identity and Access Management, used for controlling resource access.
Identity and Access Management: A control mechanism for managing resource access in secure AI systems.
Imagen: A text-to-image diffusion model that generates high-quality images from textual descriptions.
infrastructure: The core computing resources needed for generative AI, including physical hardware such as servers, GPUs, and TPUs, plus software needed to store and run AI models and training data.

K

Knowledge cutoff: The specific date up to which an AI model was trained, after which it may lack information about later events.

L

labeled data: Data that has associated tags, such as a name, type, or number.
large language models: A type of foundation model designed to understand and generate human language.
LiteRT: A Google tool that helps developers deploy AI models on edge devices.
LLMs: Abbreviation for large language models.

M

machine learning: A subfield of artificial intelligence in which machines learn from data to perform specific tasks.
Metaprompting: A prompting technique that uses prompting to guide an AI model to generate, modify, or interpret other prompts.
ML: Abbreviation for machine learning.
ML lifecycle: The end-to-end lifecycle of machine learning work, including data ingestion and preparation, model training, model deployment, and model management.
ML workflow: The end-to-end workflow for building, deploying, and managing machine learning solutions.
modality: The input and output data types a generative AI model works with, such as text, images, audio, or video. Model choice should align modality with the application's needs.
model: A complex algorithm trained on vast amounts of data that learns patterns and relationships, enabling it to generate new content, translate languages, answer questions, and more.
Model Builder: A Vertex AI option for training and using your own models, including fully custom training at scale or using AutoML.
model deployment: The process of making a trained model available for use.
Model Garden: A Vertex AI option for picking from existing Google, third-party, or open-source models.
model management: The process of managing and maintaining models over time.
model training: The process of creating an ML model using data.
multimodal: Describes generative AI applications that can process and generate different types of data, such as text, images, and code, simultaneously.

N

Natural Language API: A Google Cloud API that derives insights from unstructured text, including sentiment analysis, content classification, and entity extraction.
NotebookLM: A tool that lets users upload files and acts as a research assistant by summarizing key points, answering questions, and generating ideas while staying grounded in source material.
nucleus sampling: A sampling method in which the model considers the smallest set of tokens whose cumulative probability reaches the top-p threshold during text generation.

O

One-shot: A prompting technique where the model is given one example to learn from.
Output length: A setting that determines the maximum length of generated text.

P

Performance tracking: The practice of reviewing model metrics to check a model's performance.
platform: A layer that offers APIs, data management capabilities, and model deployment tools, bridging the gap between models and agents while simplifying infrastructure management.
Plugins: Tooling that adds new skills and integrations.
Post-generation review: A human-in-the-loop review step in which human review and feedback continue after deployment to help models improve and adapt.
Pre-generation review: A human-in-the-loop review step in which human experts validate machine learning outputs before deployment.
Prompt chaining: A technique for continuing a back-and-forth conversation with the AI, or maintaining context across prompts.
prompt engineering: The practice of creating effective prompts for generative AI models to maximize their value and tailor responses to specific needs.
prompting: The method of interacting with foundation models by providing instructions or inputs to guide them toward desired outputs.

Q

quality data: Data that is accurate, complete, consistent, and relevant.

R

RAG: An acronym for retrieval-augmented generation, a pattern in which an LLM retrieves relevant information from external sources using tooling, incorporates that information into the prompt, and then generates a response; it may optionally iterate on retrieval.
ReAct: An acronym for reason and act, a reasoning-loop technique that allows an LLM to reason and take action on a user query.
reasoning loop: An iterative process in which an agent observes, interprets, reasons, and acts, often using prompt engineering.
reinforcement learning: A machine learning approach in which learning occurs through interaction and feedback to maximize rewards and minimize penalties.
responsible AI: The practice of ensuring AI applications do not cause harm and are used in an ethical manner. The text says it should be considered throughout the entire AI lifecycle, from data preparation and model training to deployment and ongoing monitoring.
retrieval-augmented generation: A generation pattern in which an LLM retrieves relevant information from external sources using tooling, incorporates the retrieved information into the prompt, and then generates a response; it may optionally iterate on retrieval.
Role: A prompting technique that assigns a persona to the model to influence its style, tone, and focus.

S

Safety settings: Settings that filter out potentially harmful or inappropriate content from a model's output.
SAIF: Acronym for Secure AI Framework, which helps organizations manage AI/ML model risks and ensure security.
Sampling parameters: Settings that influence an AI model's behavior and allow customized results.
Secure AI Framework: A framework, abbreviated SAIF, that helps organizations manage AI/ML model risks and ensure security.
Security Command Center: A Google Cloud security tool used to provide security posture visibility.
Speech-to-Text API: A Google Cloud API that converts speech into text and can transcribe audio and video content.
structured data: Data that is organized and easy to search, often stored in relational databases.
supervised learning: A machine learning approach that trains models on labeled data to predict outputs for new inputs.

T

Temperature: A sampling parameter that controls the creativity or randomness of a model's word choices during text generation.
Text-to-Speech API: A Google Cloud API that converts text into natural-sounding speech and can create voice user interfaces and personalized communication.
Token count: A sampling parameter representing meaningful chunks of text, such as words and punctuation.
tools: Functionalities that allow an agent to interact with its environment, such as accessing and processing data or interacting with hardware.
Top-p: Also called nucleus sampling; a sampling parameter based on the cumulative probability of the most likely tokens considered during text generation.
Translation API: A Google Cloud API that translates text, documents, websites, audio, and video files.

U

unlabeled data: Raw, unprocessed information that has not been tagged and lacks meaning by itself, such as unorganized photos or streams of audio recordings.
unstructured data: Data that lacks a predefined structure and requires sophisticated analysis techniques.
unsupervised learning: A machine learning approach that uses unlabeled data to find natural groupings and patterns.

V

vector databases: Databases queried by a model in the illustrated prompt-and-output flow; the model sends a query to them and receives information used in the response.
Veo: A model that generates video content based on text descriptions or still images.
Vertex AI: Google Cloud's unified ML platform for building, training, deploying, tuning, and managing ML and generative AI applications.
Vertex AI MLOps: Vertex AI tools used by AI teams to better collaborate to monitor and improve their models.
Vertex AI Search: A search and recommendation solution for businesses.
Vertex AI Studio: A Google AI environment for building and deploying production-ready AI applications at scale.

Z

Zero-shot: A prompting technique where the model is asked to complete a task without any prior examples.

About These Definitions

These definitions are loaded from the shared release pack. Use them with the study guide and practice questions to connect vocabulary to exam scenarios.

Download App Read the full study guide Take the free practice exam