GLOSSARY

All terms from official Google Cloud GenAI documentation. 55 entries.

ARCHITECTURE

Embedding

Vector representation of data (text, images) that captures semantic meaning for similarity search.

Multimodal

Capability to process and generate multiple types of content: text, images, audio, video.

Transformer

The neural network architecture underlying most modern large language models.

Vector Database

Database optimized for storing and querying high-dimensional vectors used in semantic search and RAG.

EVALUATION

BLEU

Automatic metric for evaluating machine-generated text against reference translations.

Perplexity

Metric measuring how well a model predicts text; lower is generally better.

ROUGE

Recall-Oriented metric for evaluating summaries against reference summaries.

LIMITATIONS

Data Bias

Prejudices in training data that can cause unfair or discriminatory model outputs.

Edge Cases

Rare or unusual scenarios that can reveal model weaknesses and lead to errors.

Hallucination

When a model generates plausible-sounding but incorrect or nonsensical information.

Knowledge Cutoff

The temporal limit of a model's training data; the model doesn't know events after this date.

MODELS

Gemini

Google's proprietary multimodal AI model family, accessed via API. Processes text, images, audio, and video.

Gemini Nano

Google's most compact AI model designed for edge deployment on smartphones and embedded systems.

Gemma

Google's open-weight model family, available for self-hosting and local deployment.

Imagen

Google's text-to-image generation model.

Veo

Google's video generation model.

PARAMETERS

Context window

Maximum tokens (input+output) a model can process in one API call.

Temperature

Parameter controlling output randomness; 0=deterministic, higher values=more creative/random.

Token

The basic unit of text that models process; can be a word, subword, or character.

Top-K

Parameter limiting token selection to the K most probable options.

Top-P

Nucleus sampling parameter that limits token selection to those with cumulative probability reaching P.

PRODUCTS

Agent Builder

Vertex AI tool for building conversational AI agents with playbooks and external tools.

Agentspace

Centralized platform to manage AI agents using company data, acting as enterprise AI assistants.

Google AI Studio

Consumer-focused interface for quick prototyping with Gemini, usable without full GCP setup.

Model Garden

Vertex AI feature providing access to various AI models including Google and third-party models.

NotebookLM

AI-first notebook grounded in your documents for research, summarization, and Q&A.

Vertex AI

Google Cloud's unified ML platform including GenAI services, model training, and deployment.

Vertex AI Feature Store

Service for managing, sharing, and serving ML features consistently across training and inference.

Vertex AI Pipelines

Tool for orchestrating and automating ML workflows.

Vertex AI Studio

Enterprise prototyping interface for testing and developing with generative AI models in GCP.

PROMPTING

Chain-of-thought

Prompting technique that encourages step-by-step reasoning, improving accuracy on complex problems.

Few-shot

Prompting with multiple examples to demonstrate desired format, style, or behavior.

Metaprompting

Dynamic, adaptable prompt creation and interpretation for flexible AI interactions.

One-shot

Prompting with a single example to demonstrate the desired format or behavior.

ReAct

Reason + Act framework combining reasoning and action steps for dynamic problem-solving with tools.

Role Prompting

Assigning a persona or role to the model to guide its response style and expertise.

Zero-shot

Prompting without examples; model uses only instruction to complete the task.

RESPONSIBLE AI

Accountability

Clear responsibility for AI outputs and decisions, essential for high-stakes applications.

Explainable AI

Tools and techniques for understanding why a model made specific predictions or decisions.

Fairness

Ensuring AI systems treat all users and groups equitably without discrimination.

SAIF

Secure AI Framework: Google's framework for building and maintaining secure AI systems.

Transparency

Making AI decision-making processes understandable and data handling practices clear.

SECURITY

Cloud DLP

Data Loss Prevention: service for identifying and protecting sensitive data (PII).

Data Poisoning

Attack where malicious data is injected into training sets to corrupt model behavior.

Model Theft

Unauthorized extraction or replication of a model's capabilities.

Prompt Injection

Attack attempting to manipulate model behavior through malicious input prompts.

TECHNIQUES

Fine-tuning

Adapting a pre-trained model to specific tasks or domains with additional training on task-specific data.

Grounding

Connecting model responses to external, verifiable data sources for improved accuracy and reduced hallucinations.

HITL

Human-In-The-Loop: Incorporating human oversight and intervention in AI systems for quality assurance and accountability.

Prompt Chaining

Creating complex workflows where the output of one prompt becomes input for the next, enabling multi-step reasoning.

RAG

Retrieval-Augmented Generation: pattern combining retrieval systems with generative models to provide current, accurate information.

TRAINING

LoRA

Low-Rank Adaptation: parameter-efficient fine-tuning method reducing trainable parameters.

Pre-training

Initial training phase where a model learns from vast amounts of data before task-specific fine-tuning.

RLHF

Reinforcement Learning from Human Feedback: training method using human preferences to align model behavior.

Supervised Fine-tuning

Adapting a model using labeled examples for specific tasks.

Ready to test your knowledge?

PRACTICE WITH FLASHCARDS →