AI Glossary
Plain-language definitions of AI terms, concepts, and technologies — explained for every professional.
A
Agent Handoff
The moment one AI agent passes work, context, or responsibility to another agent in a multi-agent workflow.
Agent Memory
Agent memory is the broader system for deciding what an AI agent should remember, retrieve, update, and forget across interactions.
Agent2Agent (A2A) Protocol
The Agent2Agent protocol is a standard for agent-to-agent communication, delegation, and result exchange across systems, especially when one agent needs another agent to do part of the work.
Agentic AI
Agentic AI is the broader pattern of AI systems that can pursue goals with limited prompting between steps by combining models with tools, context, memory, and execution loops.
Agentic Loop
The iterative cycle where an AI agent calls a tool, inspects the result, and decides whether to call another tool or return a final answer based on the stop_reason field.
Agentic Workflow
A multi-step AI workflow where the system decides what to do next, uses tools, and adjusts based on results instead of returning one static answer.
AI Agent
An AI agent is a software system that can pursue a goal across multiple steps by choosing actions, using tools, observing results, and continuing until it reaches a stopping point.
AI Alignment
The field of research focused on ensuring AI systems behave in ways that are helpful, harmless, and aligned with human values and intentions — the challenge of making AI do what we actually want.
AI Benchmark
A standardized test or evaluation used to measure and compare the performance of different AI models — helping professionals understand which model is best for specific tasks.
AI Copilot
An AI assistant embedded directly in your workflow tools that provides real-time suggestions, completions, and help as you work — like having an expert looking over your shoulder.
AI Eval
AI evals are structured tests used to measure whether an AI model or workflow is accurate, reliable, and useful enough for a real task.
AI Governance
The frameworks, policies, and processes organizations use to manage AI systems responsibly — covering risk assessment, compliance, ethics, and accountability for AI decisions.
AI Guardrails
AI guardrails are the policies, validators, permissions, and workflow controls used to keep an AI system inside acceptable behavior.
AI Hallucination
AI hallucinations are confident-sounding outputs that are not properly grounded in evidence, current context, or reliable retrieval.
AI Memory
Information an AI system stores or reuses across steps or sessions so it can keep context, recall facts, and behave more consistently over time.
AI Orchestration
The process of coordinating multiple AI models, tools, and data sources into a unified workflow — connecting the pieces so they work together to accomplish complex tasks.
AI Overviews
Google Search results that use generative AI to summarize information and answer some queries directly on the results page.
AI Safety
The broad field dedicated to ensuring AI systems are developed and deployed in ways that minimize risks and harms — covering everything from preventing misuse to ensuring reliable behavior.
Anomaly Detection
A machine learning technique for finding unusual patterns, outliers, or rare events that differ from normal behavior.
API (Application Programming Interface)
A standardized way for software applications to communicate with each other — the mechanism that lets your apps connect to AI services like ChatGPT, Claude, and other models.
Attention Dilution
Quality degradation that occurs when a model processes too many items, instructions, or context elements at once.
Attention Mechanism
The core innovation inside transformer models that allows AI to understand which parts of the input are most relevant to each other — like a spotlight that highlights the most important words.
B
Bias in AI
Systematic errors in AI systems that produce unfair outcomes for certain groups — often reflecting biases present in training data or design decisions, with real consequences for hiring, lending, healthcare, and more.
Browser Agent
A browser agent is an AI system that interacts with websites and web apps as tools rather than only responding in chat.
C
Chain-of-Thought Prompting
A prompting technique that asks AI to show its reasoning step by step before giving a final answer, dramatically improving accuracy on complex problems.
Chatbot
An AI-powered conversational interface that can answer questions, complete tasks, and hold human-like conversations — the most common way people interact with AI today.
Chunking
The process of splitting a document into smaller pieces so an AI system can store, search, and retrieve the most relevant parts.
Classification Model
A machine learning model that predicts which category an input belongs to, such as spam vs. not spam or approved vs. rejected.
Claude Certified Architect
A professional certification from Anthropic that validates expertise in designing, building, and deploying production systems with Claude.
CLAUDE.md
A configuration file that provides persistent instructions, project context, and conventions to Claude Code across all sessions in a repository.
Computer Audition
The branch of AI that analyzes sound and audio signals, enabling systems to detect speech, music, events, voices, and patterns in recorded or live audio.
Computer Vision
The field of AI that enables computers to interpret and understand visual information from images and videos, powering everything from facial recognition to medical imaging.
Context Caching
A broader caching strategy that reuses valuable AI context such as prompts, documents, summaries, or memory so systems do not rebuild the same working state repeatedly.
Context Engineering
Context engineering is the work of deciding what the model sees, in what order, and with what supporting structure so the system behaves more reliably and usefully.
Context Window
The maximum amount of text an AI model can process in a single conversation — measured in tokens, it determines how much information the model can "see" and remember at once.
Context Window Management
Strategies for managing the finite token limit in conversations with a language model to maintain quality and avoid truncation.
D
Data Labeling
The process of tagging data with the correct answers or categories so machine learning systems can learn from examples.
Deep Learning
A subset of machine learning that uses neural networks with many layers to learn complex patterns — the technology behind image recognition, language models, and self-driving cars.
Document AI
A category of AI systems that read, understand, and extract useful information from documents such as PDFs, forms, invoices, and contracts.
Document Indexing
The process of preparing documents so an AI or search system can find the right passages quickly, accurately, and at scale.
Drift Detection
The process of monitoring an AI system for signs that its inputs, outputs, or performance are changing enough to require investigation or intervention.
E
Embedding (AI)
A way of converting text, images, or other data into lists of numbers that capture meaning — enabling AI to understand similarity, search through documents, and power RAG systems.
Entity Extraction
The task of pulling important names, amounts, dates, locations, and other key fields out of unstructured text.
Evals
A structured process for measuring how well an AI system performs on defined tasks, test cases, and failure scenarios before and after deployment.
Evaluation Harness
A repeatable testing setup for checking AI outputs against expected behavior so teams can compare prompts, models, and system changes safely.
Evaluator Model
A model or agent that scores, reviews, or critiques outputs so an AI system can filter weak results and improve quality before final delivery.
Explainable AI (XAI)
AI systems designed to explain their reasoning and decisions in ways humans can understand — critical for building trust, meeting regulations, and using AI responsibly in high-stakes domains.
F
Feature Engineering
The process of creating, selecting, or transforming the inputs a machine learning model uses so it can learn more effectively.
Few-Shot Learning
A prompting technique where you provide a few examples of the desired input-output pattern, helping the AI understand exactly what you want without any retraining.
Few-Shot Prompting
A prompting technique where you provide input-output examples in the prompt so the model learns the expected pattern before processing new input.
Fine-Grained Evaluation
A detailed testing approach that scores an AI system across specific sub-skills, edge cases, and failure types instead of using one broad quality score.
Fine-Tuning
The process of further training a pre-trained AI model on your specific data to customize its behavior, style, or knowledge for a particular task or domain.
Function Calling (AI)
The ability for AI models to generate structured requests to external tools and services — letting AI not just write text, but take actions like searching the web, querying databases, or calling APIs.
G
Generative AI
Generative AI refers to AI systems that create new content, such as text, images, audio, video, code, and structured outputs, rather than only analyzing or classifying existing data.
Grounding
The practice of tying AI outputs to trusted source material so answers are based on specific evidence instead of unsupported generation.
Guardrails
The rules, checks, and product constraints that keep an AI system operating within safe, reliable, and intended boundaries.
H
Hallucination (AI)
When an AI generates information that sounds confident and plausible but is factually incorrect, fabricated, or entirely made up — one of the most important risks to understand when using AI professionally.
Hub-and-Spoke Architecture
A multi-agent pattern where a central coordinator agent delegates tasks to specialized subagents and synthesizes their results.
Human-in-the-Loop
A system design approach where humans review, guide, correct, or approve AI outputs instead of letting the model operate without oversight.
Hybrid Search
A search approach that combines keyword search with semantic search so results are both precise and meaning-aware.
I
Image Segmentation
A computer vision task that divides an image into meaningful regions or objects so an AI system can understand exactly what is where.
In-Context Learning
A model behavior where the AI adapts to examples and instructions inside the prompt without changing the underlying model weights.
Inference (AI)
The process of using a trained AI model to generate outputs from new inputs — the phase where you actually use the model, as opposed to training it.
Instruction Tuning
A training step that teaches a model to follow human instructions more helpfully, clearly, and consistently.
J
Jailbreak
A technique used to bypass an AI system's safety rules or refusal behavior by phrasing prompts in a way that persuades the model to produce restricted output.
JSON Mode
A model setting or output pattern that makes AI return structured JSON so downstream systems can parse responses more reliably.
K
Knowledge Base
A structured collection of approved information that an AI system can search or retrieve from when answering questions or completing tasks.
Knowledge Graph
A structured network of real-world entities and their relationships — like a web of connected facts that AI can navigate to find accurate information and make logical connections.
L
Large Language Model (LLM)
A type of AI trained on vast amounts of text that can understand, reason about, and generate human-like language across virtually any topic or task.
Latency
The amount of time it takes a system to respond after a request is made.
LLM
A large language model is a system trained to predict and generate useful sequences of language from patterns in data.
Long Context
The ability of an AI system to accept and use large amounts of input at once, such as long documents, many files, or extended conversation history.
M
Machine Learning
A branch of AI where computers learn patterns from data and improve with experience, rather than being explicitly programmed with rules for every situation.
Mixture-of-Experts (MoE) Models
Mixture-of-experts models are architectures that keep many parameters available but activate only a small subset of them for each token or input, increasing capacity without paying the full dense-model inference cost every time.
Model Context Protocol (MCP)
An open standard that lets AI models securely connect to external data sources and tools — like a universal adapter that gives AI access to your files, databases, and applications.
Model Distillation
A technique for creating smaller, faster AI models by training them to mimic the behavior of larger, more powerful models — getting 90% of the quality at 10% of the cost.
Model Drift
The gradual decline in an AI system's performance when the real-world data or user behavior it sees no longer matches the conditions it was built or tuned for.
Model Routing
Model routing is the practice of sending different tasks to different models based on what the job needs most.
Multi-Agent System
A setup where multiple AI agents handle different roles, coordinate work, and pass results to each other to solve larger tasks.
Multimodal AI
Multimodal AI refers to systems that can work across more than one kind of input or output, such as text, images, audio, video, documents, and code.
N
Named Entity Recognition (NER)
A natural language processing task that identifies specific categories such as people, companies, dates, and locations inside text.
Natural Language Processing (NLP)
The branch of AI that enables computers to understand, interpret, and generate human language — the technology behind chatbots, translation tools, and voice assistants.
Neural Network
A computing system loosely inspired by the human brain that learns patterns from data — the fundamental building block of modern AI.
O
Object Detection
A computer vision task where an AI system identifies objects in an image and marks where they appear, usually with boxes and labels.
Open-Source AI
AI models and tools whose source code and model weights are publicly available for anyone to use, modify, and deploy — offering transparency, customization, and independence from any single AI provider.
Optical Character Recognition (OCR)
A technology that converts text in scanned documents, images, and photos into machine-readable text.
Orchestration Layer
The part of an AI system that coordinates models, tools, prompts, memory, and routing so the overall workflow runs in the right order.
P
Planning Agent
An AI agent focused on breaking a goal into steps, sequencing work, and deciding what should happen next before execution begins.
Prerequisite Gate
A programmatic check that blocks an AI agent from performing an operation until all required conditions are verified.
Pretraining
The large-scale training stage where a model learns broad patterns from massive amounts of data before it is customized for specific tasks.
Privacy-Preserving AI
An approach to building AI systems that minimizes exposure of sensitive data through techniques, controls, and product design choices that protect user privacy.
Prompt Caching
A technique that reuses previously processed prompt content so repeated requests can run faster or more cheaply when large parts of the input stay the same.
Prompt Chaining
A method where one AI prompt feeds the next so complex work is split into smaller, clearer steps instead of one oversized request.
Prompt Engineering
The practice of crafting effective instructions for AI systems to get useful, accurate, and precisely formatted outputs — the single most important AI skill for any professional.
Prompt Injection
A security attack where untrusted input manipulates an AI system's instructions, causing it to ignore the developer's intended rules or reveal information it should not expose.
Prompt Template
A reusable prompt structure with placeholders that helps teams generate more consistent AI outputs across repeated tasks.
Q
R
RAG
Retrieval-augmented generation is a pattern where an AI system gets relevant external knowledge at runtime instead of relying only on model training.
RAG (Retrieval-Augmented Generation)
A technique that combines AI text generation with real-time information retrieval from your own documents, producing more accurate, up-to-date, and source-grounded responses.
Reasoning Model
A reasoning model is an AI model optimized to spend more inference-time effort working through ambiguous or multi-step problems instead of returning the fastest plausible answer.
Reasoning Tokens
Tokens spent on an AI model’s internal reasoning process before it produces the final visible answer, often affecting quality, speed, and cost.
Recommendation System
An AI system that predicts and suggests items a user might like based on past behavior, preferences, and patterns — the technology behind Netflix suggestions, Amazon product recommendations, and Spotify playlists.
Red Teaming
A testing practice where people deliberately try to break an AI system, expose weaknesses, and uncover unsafe or unreliable behavior before real users do.
Regression Model
A machine learning model that predicts a continuous numeric value, such as price, demand, risk score, or time to completion.
Reinforcement Learning
A type of machine learning where an AI agent learns by trial and error, receiving rewards for good actions and penalties for bad ones — the method behind game-playing AI and robotics.
Reranking
A retrieval step that takes an initial set of search results and reorders them so the most relevant items appear first.
Retrieval Pipeline
The full sequence of steps an AI system uses to find, rank, and prepare information before generating an answer.
S
Semantic Caching
Semantic caching is a strategy that serves a previously computed answer when a new request is close enough in meaning to an earlier one, reducing cost and latency without requiring an exact prompt match.
Semantic Search
A search method that finds results based on meaning, not just matching keywords, so users can find relevant information even when they use different words.
Sentiment Analysis
An AI technique that automatically determines the emotional tone of text — positive, negative, or neutral — used for analyzing customer feedback, social media, and brand perception at scale.
Silent Error Suppression
An anti-pattern where tool implementations catch exceptions and return default values instead of surfacing the error to the model.
Speech-to-Text (STT)
AI technology that converts spoken audio into written text — powering meeting transcription, voice assistants, and accessibility features.
stop_reason
The field in a Claude API response that indicates why the model stopped generating, used to control agentic loop flow.
Structured Output
The ability to instruct an AI model to return data in a specific, machine-readable format like JSON, XML, or a defined schema — critical for building reliable AI-powered applications.
Structured Outputs
Structured outputs are AI responses constrained to a defined schema so downstream systems can validate and use them more reliably.
Subagent Isolation
The design principle that subagents in a multi-agent system operate with their own context and do not share memory with the coordinator or other subagents.
Supervised Learning
A type of machine learning where the model learns from labeled examples — like a student learning from an answer key — to make predictions on new data.
Synthetic Data
Artificially generated data that is created to resemble real data and used for training, testing, or protecting privacy.
Synthetic Environments for Robotics
Synthetic environments for robotics are simulated worlds used to train, test, and validate robotic systems before they operate in the physical world.
Synthetic Media
Media such as images, audio, video, or text that is generated or heavily transformed by AI systems rather than captured directly from the real world.
System Prompt
A high-priority instruction that sets the model’s role, behavior, tone, and rules before the user’s actual request is processed.
T
Temperature (AI Parameter)
A setting that controls how creative or predictable an AI model's responses are — lower temperature means more focused and consistent, higher temperature means more creative and varied.
Test-Time Compute
Test-time compute is the extra inference-time work a model does while answering a prompt, often to improve performance on harder problems at the cost of more latency or spend.
Text-to-Image AI
AI technology that generates images from written descriptions — type what you want to see, and the AI creates it. Powers tools like DALL-E, Midjourney, and Stable Diffusion.
Text-to-Speech (TTS)
AI technology that converts written text into natural-sounding spoken audio — enabling voice narration, accessibility features, and AI-generated podcasts from any text.
Throughput
A measure of how much work a system can handle over a period of time, such as requests per second or documents processed per hour.
Tokenization
The process of breaking text into smaller units (tokens) that an AI model can process — the first step in how language models read and understand your prompts.
Tokens and Context Windows
Tokens are the chunks an AI system processes, and the context window is the total working space available for instructions, history, and retrieved information.
Tool Description Routing
The mechanism by which Claude selects which tool to call based on the natural-language description provided in the tool definition.
Tool Use
The ability of an AI system to call external tools like search, calculators, databases, or APIs instead of relying only on its own generated text.
Tool Use and Function Calling
Tool use and function calling let AI systems call external tools, APIs, and actions instead of only responding in plain text.
tool_choice
An API parameter that controls whether Claude must use a specific tool, can choose any tool, or must respond without tools.
Transfer Learning
A technique where an AI model trained on one task is adapted to perform a different but related task — dramatically reducing the time, data, and cost needed to build useful AI.
Transformer Architecture
The neural network architecture behind modern AI models like ChatGPT and Claude — it processes text by understanding relationships between all words simultaneously rather than one at a time.
U
V
Vector Database
A specialized database designed to store and search AI embeddings — enabling fast semantic search where you find content by meaning rather than exact keywords.
Voice Cloning
A technique that creates a synthetic voice matching a real person's speech patterns, tone, and cadence using recorded audio samples.
W
Workflow Automation
The use of software and AI to move work through repeatable steps with less manual effort, fewer handoffs, and more consistency.
World Models
World models are AI systems or components that try to represent how an environment behaves so the system can predict what may happen next and plan actions more effectively.