One term, one clear answer.

AI and web topics explained with examples, limits and sources. Start with the definition, choose your reading depth or try a term yourself.

599 terms7 topics975 Term connections10 Oct 2026 last reviewed

Where would you like to start?

Seven topics with useful starting points

Learning paths

Recently reviewed

Accuracy↗
AI Fairness↗
AI Governance↗
API Endpoint↗
Automated Decision-Making↗
Bid Management↗

Connected knowledge

References to other terms

Claude6
Adafactor5
Agent Mesh5
AI Fairness5
Generative AI5
Accountability for AI Agents4

By letter

All terms · 599

All terms

AI fundamentals & models
Abductive Logic Programming (ALP)
ALP combines logical rules with possible assumptions that explain observations. Explanations must fit the modelled knowledge base and its conditions; several explanations may be possible.
AI fundamentals & models
Abductive Reasoning
Explanatory hypotheses are formed from observations and prior knowledge. Unlike a valid deduction from true premises, these conclusions are not certain.
AI fundamentals & models
Ablation Study
An ablation deliberately removes or disables a model component, input information or function and compares the result with the unchanged version. This helps estimate which parts actually contribute to the observed benefit.
Web design & development
Abstract Data Type
What matters for an ADT is what a data type can do — such as insert, remove or read — and which rules apply. Whether it uses an array, a linked list or another structure internally remains hidden behind the interface.
Web design & development
Abstraction
Software development hides implementation details that are not relevant to the current task and provides stable concepts, interfaces or modules. This lets you work with a system without knowing every internal step.
UX & accessibility
Accessibility
Digital accessibility aims to make websites usable for people with different abilities. WCAG describes assessment criteria; requirements under the German Accessibility Strengthening Act (BFSG) depend on the specific product or service.
Application & enablement
Accountability
Technical systems do not take organizational responsibility instead of people or organizations. Accountability requires clear responsibilities, documented decisions, suitable controls and the ability to trace errors or impacts.
Application & enablement
Accountability for AI Agents
Organizational accountability includes responsibilities for operations, data, ratings, results verification and incidents. It must be distinguished from legal liability and does not automatically answer legal liability questions.
AI fundamentals & models
Accuracy
In single-label classification, Accuracy is the number of correct predictions divided by all scored cases. Binary applies: (TP + TN) / (TP + TN + FP + FN). In rare classes, a high degree of accuracy can occur despite poor detection.
AI fundamentals & models
Action Language
A formal action language describes possible states, actions and state transitions with clearly defined meaning. Depending on the language, prerequisites, direct or indirect effects and queries about the consequences of action can be expressed.
AI fundamentals & models
Action Model Learning
When learning action models, models for AI planning are derived from observed states and action sequences. These describe, for example, when an action is possible and how it changes the state. This must be distinguished from learning a strategy of action.
Process automation
Action Schema
A scheme specifies under which modeled conditions an action is applicable and how it changes the state. By inserting concrete objects, an instantiated action is created; the real execution requires additional suitable interfaces and controls.
Process automation
Action Selection
In action selection, an agent determines the next action from possible actions. In reinforcement learning, this is done on the basis of a strategy, the so-called policy. Exploration is testing alternatives; exploitation uses actions already rated as favorable.
AI fundamentals & models
Activation Function
An activation function transforms the calculated signal of a neuron, for example its weighted input sum. It is usually non-linear so that several layers can depict complex relationships. Examples are ReLU, Sigmoid and GELU; the appropriate function depends on architecture and task.
AI fundamentals & models
Active Learning
Rather than requiring a large randomly labeled dataset, an active-learning method selects particularly informative examples and has an oracle, often a person, annotate them. Its aim is to achieve good model performance with less labeling effort.
AI fundamentals & models
Actor–Critic
The actor represents or improves the policy. The critic estimates value or advantage signals and supplies learning feedback used to adjust the policy. Many modern policy-gradient methods use variants of this principle.
Visibility & SEO
Ad Bidding
Bidding determines how much a campaign competes for visibility in an ad auction. Platforms allow manual or automatic bidding strategies. Automatic systems may adjust bids per auction based on destinations and signals. This can be distinguished from real-time bidding in programmatic advertising.
Visibility & SEO
Ad Exchange
Publishers provide advertising space, while various buyers or Platforms bid. At Real-Time Bidding, a single impression can be auctioned off in a very short time; standards such as OpenRTB support automated data exchange.
Visibility & SEO
Ad Rank
The ad rank is calculated dynamically at each auction. Among other things, Google cites bid, ad and landing page quality, thresholds, competition, context of search, and the expected impact of assets and ad formats as factors.
AI fundamentals & models
Adafactor
Adafactor scales parameter updates based on estimated second gradient moments. In the case of matrix parameters, it factors this estimate over rows and columns, thereby reducing the additional memory requirement compared to optimizers with complete moment estimates.
AI fundamentals & models
AdaGrad
In the common diagonal variant, AdaGrad accumulates squared gradients per parameter. Larger summed gradients reduce its effective step size.
AI fundamentals & models
Adam Optimizer
Adam combines exponentially weighted estimates of the first gradient moment and the second uncentered moment. With bias correction, parameter-adjusted updates for stochastic gradient optimization are created.
AI fundamentals & models
AdamW
In adaptive optimizers, classical L2 regularization is not generally equivalent to weight decay. AdamW therefore applies weight decay separately from the Adam update and makes regularization controllable separately.
AI fundamentals & models
Adaptive Algorithm
“Adaptive” is not a single algorithm class with a universal formula. This refers to methods that update their parameters on the basis of observed signals, such as adaptive filters, optimizers, controls or learning methods.
Application & enablement
Adaptive Learning
The system evaluates interactions, responses or competence levels and then adjusts, for example, order, difficulty, repetitions or assistance. Adaptive learning can be implemented rule-based or data-driven.
AI fundamentals & models
Adaptive Neuro-Fuzzy Inference System (ANFIS)
ANFIS combines fuzzy-if-then rules with learnable parameters. The documented MATLAB implementation trains Sugeno-Fuzzy inference systems using input-output data.
UX & accessibility
Adaptive User Interface
An adaptive user interface changes selected displays or controls based on context, device or usage patterns. The scope and rules of these adjustments depend on the system.
UX & accessibility
Adaptive Web Design
Adaptive web design provides different presentations specifically for particular device characteristics or usage contexts. Several layout variants are often used. Responsive elements can be added within these variants.
AI fundamentals & models
Admissible Heuristic
For a permissible heuristic, h(n)≦h*(n), i.e. no overestimation of the optimal remaining costs. Optimality statements for A* require additional prerequisites for search space, costs and implementation; in particular, the treatment of already visited states must be correct.
AI fundamentals & models
Adversarial Attacks
A classic case is adversarial examples: inputs are altered slightly in a targeted manner, so that a model misclassifies them. Depending on the taxonomy, the generic term adversarial attacks covers further forms of attack and should not be reduced to image disturbances.
AI fundamentals & models
Adversarial Robustness
A robust model should also react as correctly or reliably as possible when attacking inputs change in a targeted manner so that misclassifications arise. Robustness is always evaluated relative to a defined threat model.
Process automation
Agent Architecture
The architecture determines how the model, tools, context, state and control logic work together. This also includes task transfers, limits and auditing. A single language model does not automatically form a complete agent system.
Process automation
Agent Card
The Agent Card describes identity, capabilities, skills, supported interfaces and security requirements. Endpoints are part of the interface specifications; details depend on the A2A version.
Process automation
Agent Guardrails
Guardrails can check inputs, outputs, or appropriate tool calls. For risky actions, permissions and required approvals must apply before execution; parallel content checks cannot reliably prevent side effects that have already been triggered.
Process automation
Agent Handoff
In a handoff, active responsibility changes to another agent. Depending on the framework, context, history and structured handover data can be filtered or provided.
Process automation
Agent Loop
Agent runners typically perform multiple turns: the model decides on response or tool call, tool results flow back, and the cycle continues until a final result, error, or defined limit is reached.
Process automation
Agent Marketplace
Depending on the vendor, a marketplace can catalog full agents, templates, tools, or integrations. No cross-platform binding definition was demonstrated in the verified sources.
Process automation
Agent Memory
Memory can include conversation history, structured states, or externally stored facts. Technically, a distinction must be made between short-term context, session memory and permanent external memory.
Process automation
Agent Mesh
Agent Mesh refers to a networked agent architecture, depending on the provider. In the Solace example, communication takes place via an event infrastructure with task allocation; central coordination remains possible.
Process automation
Agent Observability
Among other things, model calls, turns, tool calls, handoffs, guardrails, errors, latencies and token consumption can be observed. The aim is not only monitoring, but also causal analysis via connected execution traces.
Process automation
Agent Orchestration
The orchestration determines which agent takes over which task, which information is passed on and when the workflow ends. It can use fixed rules, dynamic delegation or both.
Process automation
Agent Passport
The concrete meaning depends on the implementation. DigiCert describes a cryptographically verifiable proof that links agent identity to metadata, capabilities, policies, and human responsibility.
Process automation
Agent Payments Protocol (AP2)
AP2 describes how users, agents, merchants and payment services can exchange authorized purchase and payment actions. Cryptographically signed evidence documents mandate and authority. The Protocol does not replace a means of payment or the examination of concrete integration.
Process automation
Agent Reputation
Reputation bundles observed results or evaluations of other stakeholders. It can support the selection of cooperation partners, but it is neither a proof of identity nor a guarantee of future performance.
Process automation
Agent Sprawl
Agent Sprawl, for example, arises when teams independently introduce agents, temporary agents continue to run, or rights are never verified. What matters is the lack of control, not just the number.
Process automation
Agent Swarm
In the LLM context, a swarm of agents is often used for specialized agents with task transfers. It may be centrally coordinated. The name does not establish either decentralised swarm intelligence or higher quality.
Process automation
Agent Washing
Gartner uses the term for the renaming of assistants, RPA or chatbots without essential agent skills. In order to evaluate, actual task control, tool use and limits must be checked.
Process automation
Agent-to-Agent Communication
Agents can use interfaces to send messages, hand over tasks and return results. The general principle is distinct from the specific Agent2Agent protocol (A2A).
Process automation
Agent-to-Agent Protocol (A2A)
A2A describes the communication between agent applications. MCP serves to connect tools and data; the protocols can complement each other.
AI fundamentals & models
AgentBench
AgentBench evaluates models in multi-step environments. The original version includes eight areas, including databases, operating system tasks and web shopping. A benchmark result does not replace a test of your own system.
Process automation
Agentic AI
The term is used differently. In the narrower agent sense, a model controls the next step on the basis of intermediate results; a fixed workflow, on the other hand, follows predetermined paths. Both need defined boundaries.
Process automation
Agentic Analytics
A data agent can translate questions into data queries and explain results. Whether it plans further analysis steps independently depends on the system; a chat interface alone does not prove this ability.
Process automation
Agentic Checkout
The agent can provide shopping cart and order information. Purchase order, price, item and payment authorization must match the actual transaction. The specific processing depends on the shop and payment integration.
Web design & development
Agentic Coding
An agent reads project files, schedules changes, edits code, and uses tools like tests or a command line. Scope and rights are determined for the task.
Process automation
Agentic Commerce
Agents can compare offers and prepare or execute purchasing processes within granted powers. Research, ordering and payment approval are separate steps.
Data integration & search
Agentic RAG
At RAG, retrieved content is used for a response. An agent-driven process can additionally decide whether to search and check the search result or revise the request.
Process automation
AgentOps
This includes traceable processes, error analysis, quality measurement and controlled changes. Execution tracks show, for example, model and tool calls.
Application & enablement
AI Act Compliance
Requirements depend, inter alia, on the role, use and risk classification. Providers and operators have different obligations; not every AI system falls under the same requirements.
Process automation
AI Agent Development Framework
A framework supports, for example, tool connection, sequence control and state management. It provides a technical basis, but not a finished professional quality or secure approval policy.
Process automation
AI Agent Execution Trace
A trace assigns operations such as model and tool calls or handovers to a sequence. Individual operations are recorded as spans. What content is actually available depends on instrumentation and settings.
Process automation
AI Agent Identity Verification
Authentication verifies proof of identity; authorization then decides on permitted actions. For this purpose, agents can use their own managed identities or cryptographic proof.
Process automation
AI Agent Red Teaming
Red Teaming examines the entire agent system: for example, manipulated documents, tool calls, storage and transfers. The aim is to find undesirable actions and then test countermeasures.
Process automation
AI Agents
In the case of LLM agents, a language model controls actions and tool calls based on the goal and previous results. Rights, termination conditions and human trials shall be established in the Application.
AI fundamentals & models
AI Alignment
Alignment covers, among other things, target definition, evaluation and undesirable side effects. A system can successfully optimize a measured variable and still work past the actual goal.
Application & enablement
AI Art
AI can create or change images, sounds or other design elements. Selection, editing and artistic context can still be shaped by people.
Application & enablement
AI Audit
An audit examines with a defined scope and evidence whether agreed criteria are adhered to. This can include technical tests, document review and conversations.
Application & enablement
AI Avatars
In the video context, avatars can present texts with synthetic language and moving representation. Such an avatar does not have to be an interactive or self-acting agent.
Web design & development
AI Bill of Materials (AIBOM)
An AIBOM captures relevant components and source information, such as models and software dependencies. What data is included depends on the scheme used and available evidence.
Visibility & SEO
AI Bot Traffic Analysis
Requests can be evaluated for recognized bot, operator or requested content. Crawler traffic, human-triggered agent retrievals and visits via AI references are different signals.
Web design & development
AI Code Generation
In AI code generation, models generate code snippets, functions, or major changes from prompts and project context. Modern tools can additionally explain, complete, refactor or generate tests. Code generation is distinguished from an autonomous coding agent, which additionally controls files, commands and development tools independently.
Web design & development
AI Code Review
For example, a tool analyzes a pull request and suggests corrections. Its indications shall be checked; relevant errors may also remain undetected.
Web design & development
AI Coding Assistant
Assistance includes, for example, code suggestions, explanations and draft tests. Depending on the product, there are additional agent functions; the assistance itself does not require an autonomous execution.
AI fundamentals & models
AI Content Watermarking
Depending on the method, image, audio, video or text signals are changed during generation, so that a suitable detector can check a watermark. This must be distinguished from visible hints and metadata.
Process automation
AI Cost Attribution
Direct user charges and shared infrastructure shall be allocated on the basis of established rules. An assignment measures expenses; it does not automatically evaluate the benefits.
Web design & development
AI Debugging
For example, an assistant uses error messages and code to suggest causes or corrections. A plausible explanation is not yet proof; the error must be reproduced and the change checked.
Application & enablement
AI Ethics
AI ethics, for example, examines fairness, human oversight, privacy and possible harm. It helps to identify conflicts of goals and to make well-founded decisions.
Application & enablement
AI Fairness
It is socio-technical: metrics, legal situation, context, participation of those affected and process design belong together. Different criteria can contradict each other.
Web design & development
AI Gateway
An AI gateway bundles model requests and can provide logging, access control, rate limitation or fallback routes depending on the product.
Application & enablement
AI Governance
It includes inventory, responsibilities, risk decisions, data, vendors, evaluation, documentation, approvals, monitoring and incidents. Governance is more than compliance.
Application & enablement
AI Liability
Liability refers to the legal obligation to be liable for damage. In AI use, possible claims depend, among other things, on contractual relationships, breached obligations, fault and the applicable liability regime. In Germany, for example, §§ 280 and 823 BGB are possible bases; they do not depict all cases.
Process automation
AI Observability
For example, protocols, metrics and execution traces show model calls, runtimes and errors. Additional evaluations are necessary for substantive response quality.
Application & enablement
AI Risk Management
The process looks at the specific system, its use and possible impacts. Measures and remaining risks are documented and re-examined in the event of changes.
AI fundamentals & models
AI Safety
In the sense of safety, safety concerns, for example, the consequences of misconduct for people, physical assets and the environment. It covers development, testing and deployment limits; protection against attacks is a related security area.
Web design & development
AI SDK (Vercel)
The AI SDK unifies access to supported model vendors and provides building blocks for text generation, structured outputs, tool calls and user interfaces.
Application & enablement
AI Slop
This refers to digital content with little use or lack of quality, which is created with the help of AI. The designation is a quality critique and not a technical test method.
Application & enablement
AI Sprawl
sprawl can create duplicate tools, unclear accesses, and unallocated costs. The term describes an organizational development; agent sprawl is a special sub-area.
Application & enablement
AI Transparency
Transparency can make the purpose, intended areas of use and evaluation results of a model visible. It does not automatically mean that every single decision is fully explainable.
Application & enablement
AI-Assisted Browser
An AI-assisted browser combines web navigation with functions such as summarization, page questions and supported actions. What content the assistant is allowed to read and what steps it is allowed to take depends on the product and the permissions granted.
Application & enablement
AI-native browser
An AI-native browser combines web browsing with an integrated AI assistant.
Application & enablement
AI-Related Technostress
Technostress describes stresses in dealing with information technology. With AI introductions, additional tools, testing effort and changing processes can contribute to this. Copilot Fatigue is not a uniform diagnosis or standardized metric; the entry assigns the informal designation to the established technostress concept.
Web design & development
Aider Benchmark
The Aider benchmark evaluates models based on programming tasks in the Aider test setup. The Polyglot Benchmark contains 225 tasks in six programming languages. It must be distinguished from the older Python benchmark; results apply to the respective version and test configuration.
Process automation
Alexa for Shopping
According to Amazon, Alexa for Shopping combines Rufus product features with Alexa+. The assistant answers shopping questions, supports product comparisons and offers shopping actions depending on the available function. Amazon documents the renaming of Rufus on May 13, 2026.
Application & enablement
Algorithmic Discrimination
Automated systems can put people or groups at a disadvantage, for example through distorted data, unsuitable targets or features that indirectly map other properties. Whether a disadvantage is legally inadmissible depends on the use case and the applicable regulations.
AI fundamentals & models
Algorithmic Efficiency
Algorithmic efficiency describes the resource requirement of a method as a function of the input size. Time and memory complexity show how this demand is growing. Measured runtime additionally depends on implementation, hardware and data.
Application & enablement
Algorithmic Impact Assessment (AIA)
An algorithmic impact assessment examines how an automated decision-making system can affect people and their interests. For this purpose, the context of use, data, decisions, risks and countermeasures are evaluated. A concrete example is the AIA instrument of the Canadian government.
AI fundamentals & models
Algorithmic Probability
Algorithmic probability weights possible programs for a universal calculator according to their length. Shorter programmes are given greater weight; the contributions of suitable programmes are summarised. The concept forms a theoretical approach to inductive prediction.
AI fundamentals & models
Alibi
Alibi provides procedures for local and global explanations of machine learning models. Depending on the method, only inputs and outputs or internal model information are used. The project is described as source-available; license terms must be checked separately.
AI fundamentals & models
Alignment Tax
Alignment Tax refers to possible performance losses caused by alignment measures compared to a comparative version. The effect depends on model, procedure, task and measurement. It is neither an inevitable consequence of alignment nor a monetary levy.
AI fundamentals & models
Alpha–Beta Search
Alpha beta search accelerates the minimax evaluation of a game tree by limiting already known values. Branches that cannot improve the result are skipped. For the same tree under consideration, the minimax value is maintained.
Process automation
Amazon SageMaker Pipelines
Amazon SageMaker Pipelines links ML steps to a controlled workflow. This can include data processing, training, evaluation and deployment. Dependencies, parameters and execution history support traceable processes within the AWS environment.
AI fundamentals & models
Anchor Box
Anchor boxes are reference rectangles with defined sizes and aspect ratios at positions of a feature maps. An object detector predicts, among other things, displacements and size corrections. The anchor box is therefore not yet a recognized object boundary.
AI fundamentals & models
Anomaly Detection
Anomaly detection marks unusual data points or courses. In the case of outlier detection, training data can already contain outliers; in the case of novelty detection, learning is typically done on largely inconspicuous data. An anomaly is initially a conspicuity and not a proven error.
Visibility & SEO
Answer Engine Optimization (AEO)
AEO is used here as a descriptive term for better findability in response systems. Clear answers, traceable sources and technically accessible pages help to maintain such content. For Google’s AI search functions, the usual SEO basics apply; there are no special optimizations or a recording guarantee according to Google.
AI fundamentals & models
Ant Colony Optimization (ACO)
In ant colony optimization, artificial agents gradually construct solutions. Evaluated solutions affect virtual pheromone tracks that control later decisions. Metaheuristics is used for combinatorial problems and generally does not guarantee an optimal solution.
AI fundamentals & models
Anthropic
Anthropic develops AI models and applications, including Claude. The company identifies AI security and research as priorities. Anthropic is the provider; Claude designates a product and model family. Characteristics of individual offerings must be examined separately.
AI fundamentals & models
Anytime Algorithm
Anytime algorithms allow a balance between computing time and result quality. Interruptible variants provide a preliminary solution after an initial phase. Contract variants, on the other hand, require a predetermined time budget and do not have to deliver a useful result before it expires.
Process automation
Apache Airflow
Apache Airflow orchestrates workflows whose tasks and dependencies are defined as Python code. A dag describes tasks, execution conditions and further flow details. Airflow is suitable for time- or event-driven batch processes and is not a general real-time streaming system.
Web design & development
API Endpoint
For HTTP APIs, an operation is typically determined by URL or path and HTTP method. Entries, outputs, authentication and permissions are part of the interface agreement; other protocols can address differently.
Process automation
API Wrapper
An API wrapper encapsulates calls from another programming interface and offers simplified use. It can prepare parameters, answers or error handling. The approach is general and not limited to language models.
Process automation
Apify
Apify provides a platform on which Actors perform tasks such as web data extraction, browser automation or data processing. Actors accept structured input and can store results. There are existing actors and the opportunity to develop your own.
Web design & development
Application Programming Interface (API)
An API specifies how software can interact with an application, library or other component. It describes, for example, available functions, inputs and returns. Web APIs are one form; an API does not have to be accessible via the Internet.
AI fundamentals & models
Approximation Error
An approximation error describes the deviation between a target value or a target function and a selected approximation. The size depends on the method, presentation and error measure. It must be distinguished from additional rounding errors of the concrete calculation.
AI fundamentals & models
ARC-AGI-1
ARC-AGI-1 includes tasks with input and output grids. From a few examples, a system has to infer the transformation and apply it to a new input. The specific task format examines generalization under limited examples; a result alone does not prove general intelligence.
AI fundamentals & models
ARC-AGI-2
Like ARC-AGI-1, ARC-AGI-2 uses input and output grids in which a new transformation is to be derived from examples. The tasks have been made more demanding. Results are tied to tasks, test conditions and computing budget.
AI fundamentals & models
Area under the ROC Curve (ROC-AUC)
ROC-AUC is the area under the receiver operating characteristic curve. It summarizes how well predictive values rank positive cases higher than negative ones. AUC alone is more general and must always be named with the curve used.
AI fundamentals & models
Arena (formerly LMArena / Chatbot Arena)
Arena, formerly LMArena and Chatbot Arena, lets people compare two anonymized model outputs to the same input and choose a preferred response. Many pair comparisons result in rankings. These measure preferences within the respective test design and are not a universal quality judgement.
AI fundamentals & models
ARIMA Model
ARIMA combines autoregressive components, differencing and a weighted combination of past error terms (the moving-average component). The parameters p, d and q control these components. Seasonal extensions are possible, but must be explicitly modeled.
Process automation
Arize AX
Among other things, Arize AX captures traces of AI applications and supports their analysis, evaluation and comparison in experiments. Criteria and test data determine what is being evaluated. Arize AX can be distinguished from the independent open source solution Phoenix.
Process automation
Arize Phoenix
Arize Phoenix supports the study of AI applications through tracing, evaluation procedures, data sets and experiments. It can be operated locally or self-hosted and works with OpenTelemetry as well as OpenInference. Evaluation rules must fit the technical task.
Web design & development
Array
An array organizes elements via indices and can be designed differently depending on the programming language. In JavaScript, arrays are mutable array objects; their indices start from zero. Properties of a language must not be transferred across all arrays.
AI fundamentals & models
Artificial Intelligence (AI)
AI systems derive from inputs how they generate output for specific goals. Approaches can be data-based or knowledge- and rules-based. Autonomy and adaptability differ; AI does not equate to human thinking or exclusively to machine learning.
UX & accessibility
Atomic Design
A methodical design approach that systematically builds user interfaces from the smallest units: atoms, molecules, organisms, templates and pages.
AI fundamentals & models
Attention Mechanism
An attention mechanism calculates contextual weights for available representations and uses them to form an outcome. In translation models, suitable parts of the input text can thus be taken into account. Attention is a mathematical mechanism, not a human consciousness.
AI fundamentals & models
Attention Pooling
Attention-Pooling combines several representations into an overall description. A model determines weights and thus forms, for example, a weighted sum. In multiple instance learning, a group of individual examples can be classified together.
AI fundamentals & models
Attention Sink
In research, attention sinks are tokens that receive high attention weights from a language model without being correspondingly important to the content. StreamingLLM work observes this especially with initial tokens and uses their preservation in the KV cache for streaming processing.
Process automation
Attio
Attio combines configurable customer data with functions for workflows, evaluations and AI support. Which automations and integrations can be used depends on the specific offer. Data quality, responsibilities and access rights remain crucial.
AI fundamentals & models
Attributional Calculus
Attributional Calculus is a typed logic system that combines elements of propositional, predicate and multi-valued logic. It is intended to represent inductively learned hypotheses in human-oriented forms. This refers to a research approach to the representation of knowledge, not marketing attribution.
Visibility & SEO
Audience
Strategically, target group refers to a group of people with common needs or characteristics. In GA4, on the other hand, an audience is a rule-based and changeable group of detected users that meets defined conditions.
AI fundamentals & models
Audio Deepfake
Audio-deepfakes are created, for example, by text-to-speech or voice conversion and can imitate a real voice or expression. Not every AI-generated audio file is a deepfake: What matters is context, claimed authenticity and possible deception.
AI fundamentals & models
Audio Generation
Generative models learn structures of audio data and generate new waveforms or coded audio representations from them. Audio generation is the generic term; text-to-speech, music generation and sound effect generation are individual applications.
AI fundamentals & models
Audio Language Model
An audio language model transfers the principle of sequential prediction to audio. For this purpose, audio is often converted into discrete tokens, whose sequence is modeled and subsequently decoded again into an audio signal. It does not necessarily process words or understand language like a human.
Web design & development
Audit Logging
Audit logging typically captures actor, action, target object, time, outcome, and relevant context. The resulting entries form an audit trail. General application logging, telemetry and monitoring can pursue further goals.
Web design & development
Authorization
Authorization regulates rights: Who or what may perform which action on which resource? It must be distinguished from authentication, which verifies the identity. Authorization can be implemented via roles, attributes, policies or individual permissions.
UX & accessibility
Autocomplete
Upon autocomplete, an interface adds text entered inline or displays a suggestion list. `aria-autocomplete` describes this behavior for assistive technology, but does not implement it. The HTML attribute autocomplete, on the other hand, identifies input purposes and supports browser autofill.
AI fundamentals & models
Autoencoder
The encoder maps inputs to a latent representation, the decoder reconstructs the input therefrom. The information bottleneck can be caused by low dimensions, regularization, noise or other restrictions. A classic autoencoder is not automatically a generative model.
Process automation
AutoGPT
Today, AutoGPT is a platform with which AI agents can be created and executed from building blocks or natural language specifications. Agents can run workflows, respond to triggers, and integrate external tools. The product name refers to a concrete platform and is not a general technical term for autonomous AI.
AI fundamentals & models
Automata Theory
The automaton theory examines models such as finite automatons, basement automatons and Turing machines. It describes which inputs a model accepts and which calculations it can perform. It is not the same as business or industrial automation.
Process automation
Automated Decision-Making
Rules, statistics or AI provide recommendations or decisions. Partially automated support and exclusively automated decisions must be distinguished. Art. 22 GDPR sets limits for exclusively automated decisions with legal or similar significant effect; exceptions require prerequisites and protective measures.
AI fundamentals & models
Automated Driving
Automated driving involves different levels. In the SAE system, they range from Level 0 without driving automation to Level 5 with complete driving automation. Level 2 remains driver assistance with human monitoring; at higher levels, the system increasingly takes over more of the dynamic driving task within defined conditions.
AI fundamentals & models
Automated Machine Learning (AutoML)
AutoML bundles processes that automate parts of an ML workflow, such as model selection, hyperparameter optimization, feature processing or the search for model architectures. How far automation reaches depends on the system; data quality, goal definition and professional evaluation remain human tasks.
Process automation
Automated Planning
With automated planning, states, possible actions, conditions and goals are formally described. A planning procedure looks for a valid or as good an action sequence as possible. Planning is to be distinguished from mere scheduling and free text generation.
Process automation
Automation
Automation involves capturing, deciding, executing, monitoring or controlling a process with reduced manual intervention. It can be mechanical, industrial or software-based and does not necessarily require AI or complete autonomy.
Process automation
Autonomous Agent
For agents, autonomy means that the system has a certain room for manoeuvre and does not have to be controlled immediately for every single step. Classical agent research also connects agents with reactivity, goal orientation and interaction with their environment. Modern LLM agents are a possible form, not the definition of the term.
AI fundamentals & models
Autoregressive Model
Autoregressive models describe a quantity based on previous values. In generative AI, the principle is often used to gradually continue a sequence, for example by predicting a next token from the previous tokens. Statistical AR models and autoregressive language models use the same basic principle but are not the same model.
Application & enablement
Average Revenue per User (ARPU)
ARPU is calculated from the relevant revenue and the number of associated users in the same period. Which users will be included must be determined. In the subscription business, recurring proceeds and the associated subscribers are a possible reference basis; the specific definition shall be provided.
Visibility & SEO
B2B Marketing
B2B marketing is aimed at organizations and those involved in procurement or investment decisions. Purchasing processes can include multiple roles, longer decision paths, and professional requirements. B2B is a market relationship and not a particular marketing method.
AI fundamentals & models
Backpropagation
In the case of backpropagation, the influence of the model parameters on the loss function is determined backwards by the computational graph after a forward run. The basis is the chain rule of differential calculus. An optimization method such as Gradient Descent then uses these gradients to change parameters; backpropagation itself is not the optimizer.
AI fundamentals & models
Backtesting
In backtesting, a process is evaluated as if historical times were the present one after the other. For prognosis models, a rolling origin is often used: trained only with data that would have already been available at the time. Backtesting is particularly known from financial and time series applications.
AI fundamentals & models
Backtracking
Backtracking builds a solution step by step. If a choice leads to an impasse, the procedure jumps back to the last decision point and tries an alternative. Conceptually, this often corresponds to a depth-first search in the tree of possible partial solutions.
AI fundamentals & models
Bag-of-Words Model
In the Bag-of-Words model, a vocabulary is constructed and each document is displayed as a vector over that vocabulary. Values can express, for example, the occurrence, the number or a weighting of terms. The order of the words is largely lost.
AI fundamentals & models
Bagging (Bootstrap Aggregating)
Bagging creates multiple bootstrap samples from the same training dataset. A model is trained on each sample. For regression, predictions are typically averaged, combined for classification. The method can reduce the variance of unstable learning methods.
AI fundamentals & models
Bandit-Based Recommender Systems
Bandit-based recommendation systems treat recommendations as a sequence of decisions with feedback. Contextual bandits also take into account information about the situation, users or content. Central is the conflict of goals between exploiting known good options and exploring new or uncertain options.
AI fundamentals & models
Batch Normalization
In training, batch normalization calculates average and variance over a mini-batch, thus normalizing intermediate values and applying learnable parameters for scaling and shift. For the inference, stored running statistics are typically used. The procedure can stabilize training and enable higher learning rates.
Data integration & search
Batch Processing
In batch processing, data or jobs are first collected and then processed together according to schedule, quantity or trigger. This distinguishes them from real-time or streaming processing, where events are processed continuously.
AI fundamentals & models
Batch Size
In mini-batch training, the data set is processed in groups. The local batch size affects storage requirements and throughput. In the case of gradient accumulation, gradients of several such groups are merged before a parameter update; the effective batch size is then greater. Batch size and learning rate should be coordinated together.
AI fundamentals & models
Bayesian Optimization
Bayesian optimization learns from already tried-and-tested parameter combinations a surrogate model of the unknown target function. An acquisition function decides which point will be tested next to exploit promising areas and target uncertain areas.
AI fundamentals & models
Beam Search
Beam Search extends several possible partial sequences in parallel, but discards all but the best k candidates after each step. The width k is called Beam Width. In language models, the method is used for sequence decoding, but is not guaranteed to be globally optimal.
AI fundamentals & models
Bellman–Ford Algorithm
Bellman-Ford initializes distances and improves them repeatedly by relaxing all edges. After at most |V|-1 runs, the shortest paths are determined if no reachable negative cycle exists. An additional run can detect such cycles.
AI fundamentals & models
Benchmark
A benchmark specifies data, tasks, metrics, and often an evaluation protocol. In AI and machine learning, models are compared on defined capabilities. A benchmark only measures what its tasks and metrics actually cover.
AI fundamentals & models
Benchmark Contamination
Benchmark contamination occurs when a model has seen test data directly or indirectly before the actual evaluation. Then a high score can be based partly on recognition or reproduction instead of generalization. For web data and large language models, complete testing is often difficult.
Web design & development
BentoML
BentoML helps bundle model code, dependencies and service logic into reproducible deployment units. Services can be launched locally, containerized and operated in various infrastructures. The exact range of functions evolves with the versions.
AI fundamentals & models
BERT
BERT processes text with a Transformer encoder and learns contextual information from both directions of a sequence. The original model was pre-trained on large unlabeled sets of text and then fine-tuned for tasks such as classification, question answering or sentence relationships.
AI fundamentals & models
BERTScore
BERTScore maps tokens from candidate and reference text with a pre-trained contextual language model and compares their embeddings. From this, precision, recall and F1-like values are calculated. The metric can take into account semantically similar formulations better than pure n-gram metrics.
Data integration & search
BGE Embedding Models
BGE stands for BAAI General Embedding. BGE embedding models generate text vectors for semantic search and retrieval. The broader model family also includes rerankers that evaluate search hits. Supported languages, context lengths and retrieval types depend on the respective model.
Data integration & search
Bi-Encoder
A bi-encoder calculates representations for request and candidate separately. As a result, candidate vectors can be stored in advance and searched very quickly. This is more efficient than a cross-encoder that handles each request-candidate pair together, but can be less accurate when it comes to fine pair rating.
AI fundamentals & models
Bias in AI and Machine Learning
“Bias” has several related meanings in machine learning. Statistical bias describes a systematic deviation of an estimator or model. In the context of responsible AI, bias also refers to systematic biases in data or system behavior that can affect groups differently. The concrete meaning must therefore always be stated.
AI fundamentals & models
Bias–Variance Tradeoff
High bias models are often too inflexible and underfit; high variance models are highly responsive to specific training data and can overfit. In classically supervised learning problems, the bias-variance dilemma describes why model complexity must be chosen so that new data is generalized as well as possible.
Visibility & SEO
Bid Management
Bids can be controlled manually, rule-based or with platform algorithms. Google Ads Smart Bidding optimizes conversions or conversion value for each auction. Objectives such as CPA or ROAS are control variables; results depend, among other things, on tracking, competition and campaign conditions.
Web design & development
Big O Notation
In computer science, Big O is often used to describe how the time or memory requirement of an algorithm grows with the problem size. Formally, f(n)=O(g(n)) means that f is bounded above by a constant multiple of g from a sufficiently large input. Colloquially, Big O is often used inaccurately as an exact runtime class.
AI fundamentals & models
BIG-bench
BIG-bench (“Beyond the Imitation Game Benchmark”) brings together a large number of very different tasks from language, knowledge, mathematics, conclusions and other fields. The aim was to measure model behaviour more broadly than with individual standard benchmarks. BIG-bench Lite is a smaller subset.
Web design & development
Binary Search
Binary search compares the searched value to the middle element of a sorted sequence. Depending on the comparison, only the left or right half is searched further. As a result, the number of necessary comparisons only increases logarithmically with the input size.
AI fundamentals & models
BLEU Score
BLEU compares n-grams of a generated translation with one or more references, uses modified n-gram precision, and a length penalty for too short outputs. The metric was developed for corpus comparisons and does not measure factual fidelity or complete semantic equivalence.
Data integration & search
BM25 Ranking
BM25 calculates one relevance score per document for a search query. Frequent search terms in the document increase the score with decreasing added value; rare terms are weighted more strongly and the document length is normalized. BM25 works lexically and does not understand semantic similarity like an embedding model.
AI fundamentals & models
Boosting
In boosting, basic models are created sequentially instead of independently. Depending on the method, incorrectly predicted examples are weighted more strongly or negative gradients of the loss function are learned; in the case of quadratic loss, these are the residuals. The individual models are combined to form an overall model. AdaBoost and Gradient Boosting are different boosting families.
AI fundamentals & models
Bootstrapping
In the classical non-parametric bootstrap, the observed sample is used as an approximation for the unknown population. From this, many new samples of the same size are drawn. The distribution of a parameter calculated therefrom can be used to estimate uncertainty, standard errors or confidence intervals.
Process automation
Botsitting
Glean’s Work AI Index uses botsitting 2026 for work around AI use: submitting context, reviewing results, correcting errors, and cleaning up outputs. The report distinguishes between productive and unproductive effort. Botsitting can therefore also include meaningful human review and is not automatically avoidable rework.
Visibility & SEO
Bounce Rate
In GA4, the bounce rate is the percentage of non-engaged sessions. By default, a session is considered engaged if it lasts longer than 10 seconds, contains a key event considered, or contains at least two pages or pages. has screen views.
Process automation
Braintrust
Braintrust is a concrete developer tool, not a general AI term. It helps teams merge data sets, evaluators, experiments and production traces for LLM and agent applications to measurably compare changes.
Visibility & SEO
Brand Awareness
Brand awareness describes whether and how well people know or recognize a brand. It’s not the same as reach, click-through rate, or intent to buy. It can be measured directly, for example by means of surveys or suitable brand-lift methods.
Visibility & SEO
Brand Lift
Brand Lift tries to measure the incremental change in brand perception through advertising. Typical are surveys of exposed persons and a comparable group without ad contact. The exact methodology depends on platform and study; Google’s brand lift is a concrete implementation of this measurement principle.
Visibility & SEO
Brand Mention Share
Brand Mention Share compares relevant mentions of a brand to a defined comparison set. The exact count depends on the provider and the measurement method. With AI visibility, prompts, models, regions, time periods, repetitions and competitors should therefore be documented.
Application & enablement
Brand Safety
Brand Safety protects brands from unwanted associations, such as when ads are delivered alongside generally rated problematic content. It must be distinguished from brand suitability: brand suitability also takes into account the individual values and the specific positioning of a brand.
Web design & development
Breadth-First Search (BFS)
The breadth-first search first visits the starting node, then all directly adjacent nodes, then their still unvisited neighbors and so on. Typically, a queue is used for this. In unweighted graphs, BFS finds a path with minimal number of edges.
UX & accessibility
Breakpoints
Layout breakpoints define switching conditions, often based on the available width. CSS media queries can also take other media features into account. Suitable thresholds are guided by the content and how usable it is, rather than just by device models.
AI fundamentals & models
Brier Score
In the binary Brier score, the difference between the predicted probability and the actual result is squared for each prediction and then averaged. Lower values are better. The score evaluates probability predictions overall and is not isolated as a measure for calibration only.
Process automation
Browser Agent
A browser agent connects a decision model with browser tools. It processes information from the current website and performs derived actions such as navigation, input or data reading. Depending on the system, it uses a local or a provided cloud browser.
Process automation
Browserbase
Browserbase provides managed browser infrastructure through which applications or agents can start browser sessions, serve web pages and process web content. It is a specific product and not a general standard term.
Application & enablement
Business Continuity
Business continuity involves preparing for disruption, response and restarting critical performance. Business continuity management organizes risks, dependencies, emergency strategies, responsibilities, exercises and improvements. It is broader than pure IT recovery.
Data integration & search
Business Intelligence (BI)
Business Intelligence connects data from operational sources, prepares them and makes metrics accessible through reports, dashboards and analytics. BI does not describe a single product type and can be distinguished from advanced data science or AI, but can integrate it.
Visibility & SEO
Buyer Persona
A buyer persona translates insights about customers into a concrete work profile. It should be based on real interviews, CRM data, observations and resilient patterns rather than on fictitious demographic details. In B2B, it describes a person or role in the buying center and can be distinguished from the ideal customer profile at the company level.
AI fundamentals & models
Byte Pair Encoding (BPE)
BPE tokenization adapted for language models begins with small units and frequently occurring pairs are repeatedly merged. This creates a vocabulary of whole words, parts of words and smaller units. Rare or new words can thus be represented from known sections.
Application & enablement
C2PA Signing
A C2PA manifesto contains assertions on provenance and a claim referring to it. This claim is digitally signed with the private key of a signer. Validators can then check whether the signed data is unchanged and whether the credentials used are accepted according to the C2PA trust model.
Data integration & search
Cache-Augmented Generation (CAG)
Cache-Augmented Generation was proposed as an alternative for knowledge tasks with manageable, relatively stable document stock. Relevant content is loaded into the context before the request and its key value states are cached. This eliminates the need for a separate retrieval stage at runtime. CAG is not more suitable for every knowledge base than RAG.
UX & accessibility
Call to Action (CTA)
A call-to-action identifies a desired action and makes its next step findable. It can appear as a button, link or text prompt. CTA and button are not identical: a button is a UI component and can also trigger purely functional actions.
Web design & development
Canary Deployment
In a Canary deployment, old and new versions run in parallel. Only a small part of the traffic initially reaches the new version. If defined quality and error metrics are adhered to, the proportion is gradually increased; in the event of problems, the rollout can be stopped or reset.
Visibility & SEO
Canonical URL
A canonical statement indicates which URL of several duplicate or very similar pages should be preferred. This is often done via an HTML link element with rel="canonical". The indication is a strong signal, but not a command: Google may select a different canonical URL.
Web design & development
Cascading Style Sheets (CSS)
A styling language for the visual formatting of web pages, including layout, colours, typography and responsive design.
AI fundamentals & models
Causal Masking
In the case of causal masking, a position may take into account only itself and previous positions. Transformer decoders block the attention to later tokens. Together with correctly shifted inputs by one position, this prevents later outputs from being used in the autoregressive prediction.
Web design & development
Certificate Authority
A Certificate Authority uses digital certificates to confirm the assignment of a public key to an identity or name. Browsers and operating systems trust certain root CAs; this results in certificate chains. The CA is a part of the public key infrastructure and does not equate with the certificate itself.
AI fundamentals & models
Certified Robustness
With certified robustness, it is not only empirically tested whether known attacks fail. For a defined threat model, the stability of a prediction is ensured within a defined fault range. Depending on the procedure, deterministic or statistical guarantees with a specified confidence level apply. The statement only applies to the specifically certified conditions.
Application & enablement
Chain of Custody
A Chain of Custody documents the life cycle of a piece of evidence or digital artifact from capture to archiving or return. It records transfers, controllers, dates and purposes and is intended to support integrity and traceability. It is particularly well known from forensics and regulated testing processes.
AI fundamentals & models
Chain of Thought
In research, Chain of Thought refers to textual intermediate steps between task and response that can support complex inference. These visible justification steps are not equivalent to a complete or reliable insight into internal model states; modern systems can also use hidden internal reasoning.
Process automation
Chain-of-Agents (CoA)
Chain-of-agents was proposed as a multi-agent approach to long-context tasks. The input context is broken down into sections, several worker agents sequentially process partial information and pass on condensed results; a manager agent creates the final answer from this. The term here refers to a concrete research approach, not any arbitrary chain of agents.
Application & enablement
Chain-of-Thought Prompting
In the original Chain-of-Thought-Prompting, demonstration examples with question, intermediate steps and answer are given. This allows large language models to achieve better results for certain arithmetic, symbolic or logical tasks. The method is model- and task-dependent and does not equate with any form of “step by step”.
AI fundamentals & models
Change Point Detection
Change point detection searches for locations where the mean, variance, distribution, or other structural properties of a signal change. Methods can work offline on full data or online during a data stream. A detected change point is an indication of a structural change, not automatically of its cause.
AI fundamentals & models
Character Error Rate (CER)
CER is typically calculated as (substitutions + deletions + inserts) divided by the number of reference characters. It is similar to the Word Error Rate, but works at the character level and is therefore useful for languages without clear word limits or OCR/ASR comparisons. Lower values are better.
Process automation
Chatbot
A chatbot receives user input and generates responses or actions. Early systems were mostly rule-based; modern chatbots can combine language models, retrieval, tools and business logic. A chatbot is not automatically an autonomous AI agent.
Application & enablement
ChatGPT
ChatGPT is an OpenAI user interface and product platform. Depending on the tariff and mode, functions for text, images, research, files, programming and agent work with tools are available. ChatGPT is therefore not equated with a single GPT model.
Visibility & SEO
ChatGPT Ads
ChatGPT Ads enables advertisers to run campaigns that can appear in suitable usage situations for ChatGPT responses. According to OpenAI, ads are labeled, separated from answers and do not affect their content. In 2026, the offer was gradually expanded internationally and also in Europe.
Process automation
ChatGPT Work
ChatGPT Work can gather context, schedule work, and perform actions through files, tools, plugins, and supported browser workflows. The former ChatGPT Agent mode is no longer available, according to OpenAI; for longer multi-step tasks, Work became the current product path.
Data integration & search
Chunking
In chunking, documents are segmented in such a way that individual sections fit into search index, embedding model or context windows. Strategies range from fixed size with overlap to structure- or meaning-based chunking. Size and limits affect retrieval quality, cost and contextual fidelity.
AI fundamentals & models
Churn Prediction
Churn Prediction uses historical customer, usage, contract or service data to estimate churn probabilities. The model provides a risk value or classification; meaningful measures require additional root cause analysis, cost-effectiveness and a clean target definition.
Web design & development
CI/CD for Machine Learning
CI/CD for machine learning extends classic software pipelines with ML-specific checks: data and schema tests, pipeline tests, model evaluation, reproducible training artifacts and controlled deployment. Continuous training can be added as an additional process, but is not the same as CI/CD.
AI fundamentals & models
CIDEr
CIDEr evaluates how strongly n-grams of a generated image description match the consensus of multiple human reference descriptions. For this purpose, n-grams are weighted with TF-IDF, so that characteristic formulations count more than very common words. The metric evaluates text similarity to references, not automatically image truth or general language quality.
AI fundamentals & models
Classification Error Rate
In the case of classification, it often corresponds to one minus Accuracy, provided that the same set of cases and weighting is used.
AI fundamentals & models
Classifier-Free Guidance (CFG)
Classifier-Free Guidance does not require a separate classifier. The diffusion model is trained in such a way that it can work both with and without condition; in the generation, both predictions are combined. A higher guidance score can increase conditional fidelity but deteriorate diversity and image quality.
AI fundamentals & models
Claude
Claude refers to both the Anthropic model family and the associated applications. Depending on the product and version, Claude supports, among other things, long contexts, files, image comprehension, code, tool use and agentic working modes. The name should therefore always be specified with model version or product context.
Web design & development
Claude Code
Claude Code works in development environments such as Terminal and IDE with access to project files and allowed tools. It can plan changes, edit code, run tests, and edit development tasks over several steps. Which actions are possible depends on permissions and environment.
Process automation
Claude Computer Use
Anthropics Computer Use feature combines a model with tools for screenshots, mouse movements, clicks and keyboard input. This allows Claude to operate graphical interfaces, even if no suitable API is available. The function must be distinguished from general browser agents and classic RPA scripts.
Process automation
Claude Cowork
Claude Cowork was introduced for longer tasks where Claude collects information, edits files and uses tools. On September 16, 2026, Anthropic announced the merger of Cowork and Chat into Claude, initially as a rollout for Pro and Max. Available features continue to depend on tariff, platform, permissions and related tools.
AI fundamentals & models
Claude Haiku
Haiku describes the particularly fast and cost-efficient model class within the Claude family. It is typically suitable for high request volume, classification, extraction, and other tasks where speed and cost are weighted more heavily than maximum model performance. Concrete capabilities change with each version.
AI fundamentals & models
Claude Opus
Opus refers to the performance-oriented model class within the Claude family. It is offered for tasks where quality and problem solving are more important than minimal cost or latency. Features such as context, tool use and pricing are version-dependent.
Process automation
Claude Skills
Claude Skills bundle domain-specific work instructions, reference materials and, if necessary, executable tools into one structured unit. Claude can load appropriate skills as needed, making recurring procedures more consistent and modular. Skills are not a substitute for authorizations or professional quality assurance.
AI fundamentals & models
Claude Sonnet
Within the Claude family, Sonnet describes the balanced model class for demanding everyday, coding and agent tasks. It lies conceptually between the particularly fast haiku class and the performance-oriented opus class. Capabilities, context and prices are version-dependent.
Web design & development
ClearML
ClearML is a concrete open source-based tool and platform offering for the ML lifecycle. It captures experiments and artifacts, can orchestrate training jobs and pipelines, and supports model and computational resource management. It is not an ML model itself.
Web design & development
Clerk
Clerk provides pre-built components, SDKs and backend functions for login, registration, sessions, user profiles and organization functions. It is a specific authentication service and not the general technical term for authentication or authorization.
Visibility & SEO
Click-Through Rate (CTR)
The CTR is calculated as clicks divided by impressions times 100 percent. What counts as click and impression depends on the channel and measurement system. The metric shows a tendency to click, not automatically lead quality, sales or profitability.
Data integration & search
Clickstream Data
Clickstream data maps digital interactions as an event sequence, such as page views, clicks, searches, scrolls or transitions between views. Together with time stamps, session and context information, navigation paths and usage patterns can be analyzed. They show observed behavior but do not automatically explain motivation or intention.
Web design & development
Client-First
Client-First is a class and organisation system for Webflow projects. Its suitability depends on the team and project structure.
AI fundamentals & models
CLIP
CLIP was trained contrastively on image-text pairs: matching image and text representations should be closer together than inappropriate ones. This allows the model to compare or classify images based on natural language descriptions without classical task-specific training and serves as a building block in multimodal systems.
Web design & development
Cloudflare Workers
Cloudflare Workers runs application code in a distributed runtime close to users and data sources. Typical deployments are APIs, middleware, edge logic, authentication, caching and AI-related services. Workers is a concrete cloud product and not the general term edge computing.
AI fundamentals & models
Clustering
Clustering tries to find groups in data so that elements within a cluster are as similar as possible according to a chosen similarity measure. Methods such as k-means, hierarchical clustering or DBSCAN use different assumptions about the shape, density and number of groups. Clusters are mathematical structures and not automatically technically meaningful segments.
AI fundamentals & models
Cohen's Kappa
Cohen’s kappa compares the observed agreement with the agreement that would be expected randomly among the marginal distributions. A value of 1 means complete agreement; 0 is the random agreement expected according to the model. Negative values are possible. The interpretation depends heavily on class distribution and application.
AI fundamentals & models
Cohere
Cohere develops language, embedding and retrieval models as well as enterprise platforms such as North. In September 2026, Cohere signed a definitive agreement to merge with Aleph Alpha; the deal is still subject to regulatory approval.
Data integration & search
Cohere Embed
Cohere Embed converts text and – in current models – also images or mixed documents into vectors. These vectors can be used for similarity search, retrieval, classification and RAG. Since September 2026, Embed 5 is available in Pro and Fast variants.
Data integration & search
ColBERT
ColBERT generates contextual vectors for individual query and document tokens. Document representations can be calculated in advance; in the search, a so-called late interaction takes place, which can detect finer matches than a single document vector.
AI fundamentals & models
Cold-Start Problem
Many recommendation systems learn from past behavior. For new users or new products, these signals are initially missing, so that similarities and preferences can only be poorly estimated. Additional features, onboarding questions or hybrid procedures can help.
Data integration & search
Column Store
In the case of column-oriented storage, values of the same column are stored together. Analyses that aggregate a few columns across a very large number of data sets therefore have to read less data and often benefit from good compression. For single-row transactional accesses, other forms of storage are often more suitable.
Web design & development
Comet
Comet here refers to the ML platform Comet ML, not a browser of the same name. It logs training runs, parameters, metrics and artifacts and supports comparison and reproducibility. For LLM and agent applications, Comet also offers Opik with evaluation and production monitoring.
Application & enablement
ComfyUI
ComfyUI connects processing steps as nodes in a visual graph. This makes it possible to create, save and reuse generative image, video and other media workflows. Available functions depend on the models, nodes and execution environment; ComfyUI is not itself a single AI model.
AI fundamentals & models
Command R
Command R is a cohere model for dialogue-based and company-related tasks with a focus on retrieval-augmented generation and tool use. The current documentation indicates that Cohere now recommends newer Command models such as Command A for most new use cases.
Process automation
Composio
Composio provides toolkits and individual executable tools for numerous services. Agents can detect matching actions, authenticate user accounts, and perform actions across connected accounts, while permissions and authentication remain central security boundaries.
Process automation
Compound AI System
A compound AI system does not solve a task with a single model call, but through the interaction of several components. This can include LLMs, retrievers, databases, rules, tools, verifiers and orchestration logic.
AI fundamentals & models
Compute-Optimal LLM Training (Chinchilla)
DeepMind’s chinchilla work empirically showed that many large language models at the time used too many parameters and too few training tokens for their computing budget. For a fixed compute budget, a smaller model with more training data can achieve better performance. The concrete optimal scaling depends on assumptions and later research results.
Process automation
Computer Use
An AI model evaluates screenshots and suggests mouse or keyboard actions. A connected application runs them in the shared environment. Rights, controls and human approvals determine the room for manoeuvre.
AI fundamentals & models
Computer Vision
Computer vision includes tasks such as image classification, object recognition, segmentation, tracking and OCR. Modern systems often use deep learning, vision transformers or multimodal models to derive structured information from visual data.
Process automation
Computer-Use Sandboxing
Computer use actions are performed in a separate browser or desktop environment. Data access, accounts, rights and network destinations must be configured appropriately. Isolation limits impact but does not replace authorization and clearance controls.
AI fundamentals & models
Concept Drift
Concept drift changes the underlying pattern that a model has learned. The same entrance may later have a different meaning for the destination. This differs from pure data drift, in which only the distribution of inputs is shifted.
AI fundamentals & models
Conditional Generation
A conditional generative model does not generate random samples, but models outputs depending on additional information. Conditions can be, for example, a prompt, a category, a reference image or structured features.
Web design & development
Confidential Computing
According to the definition of the Confidential Computing Consortium, data is protected during processing in a hardware-based, attested TEE. The approach complements other protections; application security, key management and the concrete threat model remain relevant.
AI fundamentals & models
Conformal Prediction
Conformal prediction generates prediction sets or intervals. Under suitable assumptions, such as exchangeability of the data, coverage guarantees apply for finite samples. The usual guarantee is marginal over calibration and test data, not for each individual case or subgroup.
AI fundamentals & models
Confounding
In confounding, an observed relationship between two variables can be partially or completely explained by common causes. As a result, a statistical correlation cannot be interpreted as an effect without appropriate assumptions and causal analysis.
Process automation
Confused Deputy Problem in AI Agents
A service, plugin or agent has its own access rights. If the identity and purpose of a request are not linked to the authorisation, this intermediary may inadvertently use its rights for an unauthorised assignment.
AI fundamentals & models
Connectionist Temporal Classification Loss (CTC Loss)
CTC sums up the probabilities of all valid monotonic alignments between a longer input sequence and a shorter target sequence. A blank symbol and rules for merging repetitions allow for training without frame-accurate labels.
Application & enablement
Consent
GDPR consent must be voluntary, specific, informed and unambiguous. Those responsible must be able to prove it. A revocation must be at any time and as simple as the granting possible. Consent is one of several possible legal bases.
Application & enablement
Consent Orchestration
Consent Orchestration translates consent decisions into technical rules for connected data flows. The specific coverage is product-dependent: Tealium currently supports event activations; AudienceStream activations require separate implementation.
AI fundamentals & models
Consistency Model
Consistency models learn consistent mappings along a noise trajectory and can thus directly sample from noise to data. They have been developed as a faster alternative or complement to iterative diffusion sampling techniques and can be trained by distillation or directly.
AI fundamentals & models
Constitutional AI
Constitutional AI uses a collection of natural language principles as a constitution. In the original method, a model first generates answers, critiques and revises them on the basis of these principles and can then be further trained with AI-generated feedback.
AI fundamentals & models
Constitutional Classifiers
Constitutional Classifiers are protection components for model inputs and outputs. The classifiers are trained with synthetic data derived from a so-called constitution with rules on allowed and non-allowed content, and are intended to detect and block jailbreaks, among other things.
Application & enablement
Content Authenticity
Content Authenticity includes provenance and trust information about digital media. The C2PA standard defines content credentials with cryptographically verifiable manifests that can contain information about the origin and editing of an asset.
Application & enablement
Content Creation
Content creation involves more than the actual generation: topic selection, research, conception, production, editing, approval and publication are included depending on the workflow. Generative AI can support individual steps, but it does not automatically replace goal definition, fact-checking and editorial responsibility.
Application & enablement
Content Credentials (C2PA)
Content credentials store information about the origin and editing of a digital asset in C2PA manifests. These can include, among other things, creation and editing steps, tools used and cryptographic signatures. C2PA thus describes traceable provenance signals, not the automatic determination of whether a content is true or trustworthy.
Web design & development
Content Delivery Network (CDN)
A CDN distributes or caches content at many network locations. This allows requests to be answered from a nearby edge location, which can reduce latency and load on the origin server. Many CDN services add security features to performance.
Visibility & SEO
Content Distillation in Marketing
Here, “content distillation” refers to the editorial consolidation of extensive starting content for marketing purposes, not a standardized AI training method. Language models can provide summaries and format designs. Statements, conditions and sources must be checked on the original; a higher content ROI is not guaranteed.
AI fundamentals & models
Content Filter
Content filters often work in addition to the actual generative model. Classifiers check prompts and responses for defined risk categories or other unwanted content and trigger an action depending on configuration and severity.
Data integration & search
Content Fingerprinting
Content fingerprinting generates features or hashes from the actual content instead of just from filename or metadata. Depending on the method, modified versions of an image, audio or video can also be recognized as similar and compared against known fingerprints.
Visibility & SEO
Content Marketing
Content marketing seeks to build attention and trust not primarily through direct product advertising, but through helpful, relevant and consistent content. Content should be linked to target audience, customer journey and measurable business goals.
Visibility & SEO
Content Personalization
Websites, e-mails or product notes show different content depending on the segment or situation. Rules or models select suitable variants. Personalization can be based on voluntary information, behavior or context; it does not necessarily require AI.
Process automation
Content Pipeline
A content pipeline connects responsibilities and work steps of content creation. It can be manual, partially automated or largely automated. AI designs, clearances and CMS connections are possible components; publication must only take place after the specified checks.
Application & enablement
Content Policy
A content policy defines permitted, restricted and inadmissible content as well as responsibilities and audit processes. For AI-generated content, it can additionally set requirements for human review, labelling, originality and sources.
AI fundamentals & models
Content-Based Filtering
Content-based filtering represents content via features and forms a user profile based on previous interactions or explicit information. Objects whose characteristics are similar to the user profile are then recommended; data from other users is not absolutely necessary for this.
Process automation
Context Budget
A context budget specifies how much space individual input parts receive. Depending on the API, output or further processing must also be considered. It is a planning parameter and not a uniformly standardized parameter; the meaningful distribution depends on the task and the model.
AI fundamentals & models
Context Caching
A provider can cache common prompt prefixes or explicitly created context objects. A cache hit reduces costs or processing effort depending on the API. Minimum length, validity period and billing are dependent on the provider. Caching does not change model weights.
AI fundamentals & models
Context Compression
Extensive context is condensed for later model steps, for example by summary, token selection or special compacting elements. Details can be lost; task quality and obtained boundary conditions must be checked.
Application & enablement
Context Engineering
Context Engineering includes instructions, relevant documents, history, tool definitions and results. The goal is a helpful, up-to-date and trouble-free context. This includes retrieval, selection, summary, and separation of trusted instructions from external data.
AI fundamentals & models
Context Poisoning
Attackers manipulate documents, tool responses, or stored notes that an agent later uses. This can influence responses or actions. The attack on runtime context must be distinguished from the poisoning of training data and can work together with indirect prompt injection.
AI fundamentals & models
Context Rot
A large context window does not mean that all information is used equally reliably. Length, distractions, position and task can affect performance. Context Rot refers to observed quality losses, no temporal perishability of files and no universal fixed token limit.
AI fundamentals & models
Context Window
The context window is the model-dependent token capacity of a request. This includes inputs and outputs as well as internal reasoning tokens for corresponding models. It is not a permanent memory; large capacity does not guarantee error-free use.
AI fundamentals & models
Contextual Bandit
The system observes features of the situation, selects an available action and receives their reward or cost. It must use successful options while exploring alternatives. Unlike comprehensive reinforcement learning, the immediate decision is typically the priority.
Visibility & SEO
Contextual Targeting
For example, ads are displayed on suitable topics, keywords or placements. AI can classify content for this, but is not a defining component. Contextual targeting must be distinguished from personal behavioral targeting; concrete platforms can take further signals into account.
AI fundamentals & models
Continual Learning
A model is adapted over several learning phases. New updates can affect previous abilities, which is known as catastrophic forgetting. Methods such as replay or regularization are intended to limit this conflict. An ongoing chat is not automatically continuous learning of the model weights.
Web design & development
Continuous Batching
An inference server reschedules the batch during generation steps. Finished requests leave it and new requests are added. The entire batch therefore does not have to wait for the longest response. Higher throughput does not automatically mean shorter waiting times for every individual request.
AI fundamentals & models
Contrastive Learning
A model learns through positive and, depending on the procedures, negative pairs. For example, two views of the same image should be close together. Defining matching pairs, data augmentation and loss function are crucial; false negative pairs can interfere with learning.
AI fundamentals & models
ControlNet
ControlNet complements a pre-trained text-to-image diffusion model with trainable control components. A model trained to the respective specification uses, for example, edge maps or body poses. The control improves the structural binding, but does not guarantee an exact reconstruction.
AI fundamentals & models
Conversational AI
A dialogue system processes utterances, manages conversational context and generates answers or triggers appropriate steps. It can connect rule-based dialogue, language processing, LLMs and tools. Not every dialogue-capable system works autonomously or uses a generative model.
Data integration & search
Conversational Analytics
A system translates technical questions into data queries or analysis operations and explains results. Data context, key figure definitions and access rights are crucial. Dialogue-based BI is meant here; the term can also refer to the analysis of recorded conversations elsewhere.
Data integration & search
Conversational Search
The system connects follow-up questions with the original search intent, clarifies ambiguities and retrieves appropriate information. Results can be hit lists or source-based answers. A freely formulating chatbot without retrieval is not automatically a reliable search solution.
Visibility & SEO
Conversion Rate Optimization (CRO)
CRO examines obstacles on websites or apps and tests changes to content, operation or process. The conversion rate compares target actions to a clearly defined reference value. More deals only make sense if quality, sales and user interests are also taken into account.
AI fundamentals & models
Convolutional Neural Network (CNN)
Convolution layers apply common filters to different positions and create feature maps. Multiple layers can combine simple patterns into more complex features. CNNs are often used for images, but also for other spatial or temporal data.
Visibility & SEO
Copilot Search in Bing
Copilot Search in Bing summarizes information from the web to a generated answer and shows sources and other links. The feature is part of Bing and does not equate to the separate Microsoft 365 “Copilot Search” feature for corporate data.
Application & enablement
Copyright in AI-Generated Content
In Germany, a copyrighted work presupposes a personal intellectual creation. Humans can be creators; an AI system cannot. Whether human contributions to an AI-based result justify protection is to be examined on a case-by-case basis.
Application & enablement
Copywriting
Copywriting combines target group understanding, benefit arguments, brand voice and clear calls for action. Texts appear, for example, on offer pages, in advertisements or e-mails. Compared to content writing, the desired reaction is more important; good texts require research and revision.
AI fundamentals & models
Cosine Annealing
Learning rate planning reduces the step size of the optimizer from an initial value to a minimum over a defined training phase. Cosine Annealing describes the course of the curve. Warm restarts are a possible extension, but not part of every implementation.
Visibility & SEO
Cost per Acquisition (CPA)
CPA is calculated from the allocated costs divided by the number of defined acquisitions. The focus here is on attracting new customers. Platforms also use CPA for cost per action; this meaning must be distinguished from the already existing cost per conversion and comprehensive customer acquisition cost.
Visibility & SEO
Cost per Click (CPC)
The average CPC results from ad costs divided by click count. The CPC billing model pays for clicks. Maximum bid, actual click price and average CPC are different sizes; cheap clicks do not guarantee valuable deals.
Visibility & SEO
Cost per Conversion
Cost per conversion is typically calculated by dividing the total relevant cost by the number of conversions recorded. The significance depends heavily on what is defined as conversion and how cleanly it is measured.
Visibility & SEO
Cost per Mille (CPM)
CPM is calculated as cost divided by impressions times 1,000. Impressions are ad displays and not a thousand different people. Viewable CPM refers to visual impressions and must be distinguished from general CPM.
Web design & development
Creative Coding
For example, code creates or modifies graphics, animations, sound and interactive installations. Creative expression and exploration are in the foreground. Creative engineering is broader and inconsistent; for the creative programming context, creative coding is the more precise term.
Process automation
CrewAI
CrewAI organizes cooperating agents in crews with tasks and roles. Flows are used for sequence and status control and can integrate crews. The open source framework must be distinguished from additional commercial operational offerings; multiple agents do not make sense for every task.
AI fundamentals & models
Cross-Attention
Cross-attention connects two streams of information. In the classic encoder-decoder transformer, queries come from the decoder and keys as well as values from the encoder. In self-attention, they come from the same stream. The mechanism can also connect text and image information.
Data integration & search
Cross-Encoder
For example, a cross-encoder processes search query and document as a pair. As a result, their components can interact directly with each other. Unlike a bi-encoder, it typically does not generate independently indexable vectors per text. It is therefore often used for re-sorting a limited result set.
AI fundamentals & models
Cross-Entropy Loss
For a unique target class, cross entropy penalises a low predicted probability for that class. For multiclass classification, PyTorch CrossEntropyLoss expects unnormalized logits and processes the necessary normalization internally. Training loss and practical classification quality are different sizes.
Data integration & search
Cross-Modal Retrieval
The system maps content of different modalities – for example text, image, audio or video – to comparable representations. This allows a text query to find suitable images or an image to find similar texts. This is narrower than any arbitrary multimodal search.
AI fundamentals & models
Curriculum Learning
In curriculum learning, the order or sample distribution of training data changes during learning. Difficulty can be determined manually, heuristically or model-based. The process must be distinguished from mere one-off sorting and self-paced learning.
Web design & development
Cursor
Cursor is an AI-assisted code editor. It combines project context, natural-language instructions and suggested code changes. Agent features can carry out further development steps; the resulting changes must be reviewed.
Application & enablement
Custom GPT
A custom GPT combines reusable instructions, optionally provided knowledge and supported functions in ChatGPT. Creation and editing depend on the account and workspace. OpenAI has announced a replacement by plugins; existing GPTs can be used until the respective changeover.
Visibility & SEO
Customer Acquisition Cost (CAC)
CACs are typically calculated by dividing the considered marketing and distribution costs of a period by the number of new customers acquired. Exactly which costs are included must be documented so that comparisons remain meaningful.
Data integration & search
Customer Data Platform (CDP)
According to the current definition of the CDP Institute, a CDP bears responsibility for customer identity, structure of the customer data set and a uniform interface for data services. It can connect data from multiple sources; however, a full “360-degree view” does not automatically emerge.
UX & accessibility
Customer Journey
The customer journey includes actions, touch points, expectations and experiences before, during and after a decision. It is rarely strictly linear. A customer journey map is the visualized analysis of this process and not the journey itself.
Visibility & SEO
Customer Lifetime Value (CLV)
CLV may include expected turnover, contribution margin or net worth, depending on the model. Therefore, the period, margin, retention probability, discounting and included acquisition and support costs must be explicitly documented.
AI fundamentals & models
CutMix
With CutMix, a rectangular patch is inserted from one image into another. The training label is mixed according to the area proportion of both images. The method combines aspects of cutout and mixup, but retains visible image areas.
AI fundamentals & models
Cyclical Learning Rates (CLR)
Instead of remaining constant or only decreasing, the learning rate follows a cycle, approximately triangular or in variants with decreasing amplitude. Boundaries, cycle length and shape must be chosen or estimated with a learning-rate-range test.
Visibility & SEO
Dark LLM Traffic
When information about a visit's origin is missing, analytics may classify it as direct or unknown traffic. This can include visits from AI recommendations. The classification (direct) / (none) alone does not establish an AI origin; it also does not substantiate a percentage of such visits.
UX & accessibility
Dashboard
A dashboard compiles selected information visually and purposefully. It shows the current or periodically updated state of an area at a glance and refers to detailed analyses if necessary. It is not automatically real-time and not the same as a detailed report.
AI fundamentals & models
Data Augmentation
In data augmentation, training data is extended by transformations that generate meaningful variants for the task. In the case of classification, the label often remains the same; in the case of object recognition or segmentation, annotations must be appropriately transformed.
Data integration & search
Data Catalog
A data catalogue is a searchable metadata system. It can link descriptions, ownership, technical terms, classification, access routes, quality and lineage. The actual data is usually in other systems; the catalogue does not automatically grant access.
Data integration & search
Data Clean Room
A Data Clean Room is a technically and organizationally controlled environment for data matching and analysis. Access, use and output rules limit who is allowed to see which data or results. The concrete architecture can use pseudonymization, aggregation or privacy-enhancing technologies, among other things.
Data integration & search
Data Dictionary
A data dictionary contains technical and semantic metadata about fields, tables or interfaces. Possible indications are name, meaning, data type, format, unit, value range, key and relationships. The exact scope depends on the purpose.
AI fundamentals & models
Data Drift
Data drift changes the statistical distribution of characteristics P(X) between reference and operational data. It can be distinguished from Concept Drift, i.e. a change of P(Y|X), from Label Drift and from Training-Serving Skew. Drift is a test indication and not automatically a drop in performance.
Data integration & search
Data Enrichment
In data enrichment, existing data is augmented by additional attributes, references or metadata. Sources can be internal systems, open data or calculated features. The supplement can increase the utility value, but also introduce errors and risks.
Application & enablement
Data Governance
Data governance is the decision-making and responsibility framework for the effective, value-creating, secure and compliant handling of data. It defines roles, principles, decision-making rights and control. Operational data management implements these specifications in processes and systems.
AI fundamentals & models
Data Labeling
In data annotation, people or systems mark relevant properties in text, images, audio or table data. Labels can be created manually, semi-automatically or automatically. They represent an agreed interpretation and are not automatically objective ground truth.
Data integration & search
Data Lake
A data lake is a scalable repository for structured, semi-structured and unstructured data. Data can be available in original or processed form; structure and processing are often only defined for the respective use case. It must be distinguished from Data Warehouse and Lakehouse.
AI fundamentals & models
Data Leakage in Machine Learning
In the event of data leakage, unauthorized information affects training, feature generation or evaluation. As a result, a model in the test often seems unrealistically good. Typical forms are pre-processing before data splitting, target leakage and future information. This does not mean a data protection incident.
Data integration & search
Data Lineage
Data Lineage documents sources, movements, processing steps and dependencies of data up to their output. The term Provenance is often used more broadly for origin and context information; however, the delineation is not uniform everywhere.
Data integration & search
Data Management Platform (DMP)
A DMP collects and connects target group data from various sources, forms segments and can pass it on to advertising systems. First, second and third-party data can be used. Cookie or device identifiers are often pseudonymous and not automatically anonymous.
Data integration & search
Data Mesh
Data Mesh is a socio-technical organizational and architectural concept for analytical data. It combines four principles: domain responsibility, data as a product, a self-service data platform and federated computational governance. It is not a single product and not a mere platform.
Data integration & search
Data Mining
Data mining involves methods that derive structures, rules or predictive models from data. Typical tasks are classification, regression, clustering and association analysis. Finding patterns must be validated with suitable test data and interpreted professionally.
Web design & development
Data Parallelism
In the case of data parallelism, a calculation or a model is replicated on several processes or devices. Each replica processes different data partitions, then merges results or parameter updates. Synchronous gradient synchronization is a common but not mandatory procedure.
Data integration & search
Data Pipeline
A data pipeline is a coherent processing chain of sources, transformations and destinations. It can read, check, filter, enrich or store data. Batch and streaming are possible modes of operation; not every pipeline is ETL or modeled as DAG.
AI fundamentals & models
Data Poisoning
Data poisoning is an integrity attack on ML systems. Attackers change, supplement or weight data or labels in such a way that the general performance decreases, certain cases are treated incorrectly or a backdoor arises. The attack takes place before or during learning.
AI fundamentals & models
Data Preprocessing
Depending on the task, preprocessing involves cleaning, imputation, scaling, normalization, coding and transformation. Steps can take place before or within a model pipeline. Learned parameters must only come from the respective training data.
Application & enablement
Data Processing Agreement (DPA)
In the GDPR context, the data processing contract is a binding contract or other legal instrument in accordance with Article 28. It shall specify, inter alia, the subject matter, duration, nature, purpose, types of data, data subjects and rights and obligations. It is not a general data exchange agreement.
Web design & development
Data Structure
A data structure is the concrete organization and storage of related information, for example as an array, list, stack, queue, hash table or tree. Their characteristics determine effort, order and possible operations.
AI fundamentals & models
Data Validation for Machine Learning
ML data validation compares data with schema, value ranges, categories, completeness and distribution rules. It can detect anomalies, drift and training-serving skews. However, it does not confirm professional accuracy, fairness or model suitability.
Web design & development
Data Version Control (DVC)
DVC manages data and model artifacts via metafiles that can be versioned with Git. The actual files remain in a DVC cache and can be shared via local or external remotes. In addition, DVC supports pipelines and experiments; reproducibility also requires available code, dependencies and data.
UX & accessibility
Data Visualization
Data visualization encodes data by position, length, color, shape or surface. It is used for exploration or communication. Diagram type, scale, aggregation and context determine which statement is recognizable and how easily it can be misinterpreted.
Data integration & search
Data Warehouse
A data warehouse is a data storage optimized for analysis and reporting. Data from operational sources is integrated, cleaned, modeled and often historicized. Queries do not directly burden the operational systems. A warehouse can be operated locally or as a cloud service.
Data integration & search
Dataset Deduplication
Exact deduplication compares defined fields to equality. Fuzzy duplicate detection uses normalization, similarity, and entity resolution. Treatment can mean marking, linking, merging or removal; it must be distinguished from memory and block deduplication.
Application & enablement
Datasheets for Datasets
The documentation format proposed by Gebru and co-authors accompanies a dataset with questions about motivation, collection, composition, processing, use, distribution, maintenance and limitations. It is neither a universally mandatory standard nor a quality certificate.
AI fundamentals & models
DBSCAN
DBSCAN is an unsupervised clustering process. Points with enough neighbors within the radius eps become core points; reachable points can be border points, the rest are considered noise. min_samples specifies the required neighborhood density.
Process automation
Decision Support System (DSS)
A decision support system is an interactive information system that provides data, analysis, models or simulations for decision-making processes. It supports particularly partially structured problems, but does not necessarily make the decision itself.
AI fundamentals & models
Decision Theory
It combines possible actions, uncertain conditions, consequences and preferences into a comprehensible decision rule.
AI fundamentals & models
Decision Threshold
If a score is above or below the set limit, a decision is triggered. The appropriate value depends on error costs and context of use.
AI fundamentals & models
Decision Tree
Each inner node tests a feature; leaves provide class or numerical value. Training procedures select divisions according to a quality criterion.
AI fundamentals & models
Decoding Strategy
In autoregressive models, decoding determines which next token is chosen from the conditional probability distribution. Greedy Search and Beam Search choose more deterministic; sampling as well as top-k or top-p generate stochastic diversity.
AI fundamentals & models
Deductive Reasoning
A deductive argument is valid if its conclusion cannot be false when all its premises are true. Only a valid argument with actually true premises is sound and guarantees a true conclusion.
AI fundamentals & models
Deep Compression
The name refers to the combination of pruning, trained quantisation or weight sharing, and Huffman coding presented by Han, Mao and Dally in 2015. It does not mean model compression in general.
AI fundamentals & models
Deep Learning
Deep learning is a branch of machine learning. Models with multiple processing layers learn hierarchical features from data; there is no universal minimum number of layers that makes a model “deep”.
AI fundamentals & models
Deep Reinforcement Learning (DRL)
DRL combines reinforcement learning for sequential decisions with deep networks that approximate value functions, policies or state representations, for example.
Application & enablement
Deep Research
Deep research refers to a multi-stage research process with AI support. Functionality and source quality vary depending on the product.
AI fundamentals & models
Deepfake
The term often covers images, videos or audio generated or altered using deep learning. Not every synthetic file is a deepfake, and harmful intent is not a necessary part of the definition.
AI fundamentals & models
Deepfake Detection
Detection methods analyse artefacts, temporal patterns, metadata or model features and return scores or classes. They check for signs of manipulation, not automatically whether a statement is true.
AI fundamentals & models
DeepWalk
The algorithm, which was introduced in 2014, generates short random walks and learns from them using a skip-gram-like method, knot embeddings in which structurally similar knots are closer.
AI fundamentals & models
Demand Forecasting
Demand forecasting uses historical time series, influencing factors and assumptions to estimate expected demand together with uncertainty. It is an input to inventory, staffing or capacity planning, not the same as those tasks.
Visibility & SEO
Demand-Side Platform (DSP)
A DSP combines campaign goals, budgets and target group rules with available advertising inventory. Purchasing can take place via real-time auctions or other programmatic agreements. Good placement, exact target group approach and favorable prices are not automatically guaranteed.
Application & enablement
Demographic Parity
A binary classifier satisfies demographic parity when its positive decision rate does not depend on group membership. The criterion does not take ground truth or qualifications into account.
AI fundamentals & models
Denoising
Depending on the method, random or unwanted signal components in images, audio, measurement series or data representations are estimated and reduced. Denoising is not limited to neural networks.
Data integration & search
Dense Passage Retrieval
DPR trains two separate encoders (question and text section) in the same vector space. In contrast to pure keyword search, DPR also captures meaning relationships with different word choices.
Data integration & search
Dense Retrieval
An embedding model maps requests and documents as number vectors. A similarity measure, such as scalar product or cosine similarity, serves to order the hits. High similarity does not guarantee a factually correct answer.
AI fundamentals & models
Dependency Parsing
The method determines directed relationships (dependencies) between a head word and its dependent words. This explicitly maps the grammatical role (subject, object, modifier).
AI fundamentals & models
Depthwise Separable Convolution
The method combines a spatial convolution per input channel with a subsequent 1×1 convolution to mix the channels. Compared to a comparable standard convolution, it can save parameters and computing operations. The actual runtime also depends on hardware and implementation.
UX & accessibility
Design Grid
A system of columns and rows that serves as a design grid for the harmonious placement of elements.
Web design & development
Design Pattern
A design pattern explains a problem, the context and a possible structure of the solution, such as Observer or Factory. It is not a finished code and should only be used if it fits the specific problem.
AI fundamentals & models
Detection Transformer (DETR)
DETR considers object recognition as a direct set prediction problem. It eliminates the need for handcrafted components such as Non-Maximum Suppression (NMS) and Anchor Boxes and uses self-attention for global image contexts.
AI fundamentals & models
Detokenization
A decoder matching the model converts token IDs into text. Settings for special tokens and spaces affect the output. After normalization or other processing steps, the original text is not always faithfully reconstructable.
Web design & development
Devin (AI Software Engineer)
Devin from Cognition AI combines code generation, terminal access, browser and troubleshooting in one sandbox. It can analyze repositories, fix bugs, run tests and independently create pull requests.
Process automation
Dialogflow
Dialogflow supports the development of chat and voice agents. Depending on the product variant, intents, entities, defined dialog processes and generative functions are used, among other things. Connections to specialist systems require suitable interfaces and a tested integration.
Process automation
Dialogue Management
Dialogue Management tracks the state of the dialog, stores contextual information and decides on the next action of the system based on rules, policies or machine learning.
AI fundamentals & models
Difference-in-Differences (DiD)
The change of a target variable before and after a measure between two groups is compared. For a causal interpretation, it is necessary in particular that the groups would have developed in parallel in the relevant period without a measure. Inappropriate comparison groups and further changes can distort the result.
AI fundamentals & models
Differential Evaluation
Here, “differential evaluation” is used descriptively for a comparative test; it is not a clearly defined standard method. Both variants are evaluated under comparable conditions with the same tasks. In addition to average values, individual deteriorations and technical errors are checked.
Application & enablement
Differential Privacy (DP)
A randomized method fulfills a defined data protection guarantee if neighboring data sets have only limited different output probabilities. Frequently calibrated noise is used for this purpose. Parameters such as ε and, where applicable, δ and the protected unit determine the guarantee; absolute secrecy does not follow.
Visibility & SEO
Differential Privacy in Advertising
In a corresponding implementation, aggregated advertising key figures with controlled statistical noise are output. The guarantee depends on the mechanism and its data protection budget. Collection, identifiers and other processing steps shall be assessed separately.
AI fundamentals & models
Differentially Private SGD (DP-SGD)
DP-SGD typically limits the gradient contribution of each training example and adds calibrated noise. A privacy accounting records the accumulated loss over the training. The guarantee applies to a defined protected entity and defined parameters; it does not mean that any confidential information can never be issued.
AI fundamentals & models
Differentiation (in Machine Learning)
In machine learning, differentiation determines how much the error changes when a parameter is slightly varied. Through automatic differentiation (autodiff) in the backpropagation step, neural networks learn efficiently.
AI fundamentals & models
Diffusion Model
During training, the model learns to reverse a defined noise process. In the generation, an initial distribution is converted into data in several steps. Application, presentation form and sampling method determine whether, for example, images, audio or other data are generated.
AI fundamentals & models
Diffusion-based Language Model (Diffusion LLM)
Depending on the procedure, continuous text representations or discrete tokens are iteratively refined. Multiple positions can be edited in parallel; not every variant always updates the entire text at the same time. This distinguishes the approach from a purely left-to-right token generation.
Application & enablement
Digital Content Provenance
Standards such as C2PA link media to signed manifests describing origin, tools and edits. This can make integrity and provenance claims verifiable, but does not prove that the content is true.
Application & enablement
Digital Disruption
In common usage, digital disruption denotes significant market changes. Christensen’s theory of disruptive innovation is narrower: New providers open up underserved market segments or new markets and can later move on to become established providers. Not every rapid improvement meets this definition.
Process automation
Digital Farming
Digital farming uses, among other things, digital documentation, sensors, geodata and networked machines. This can support decisions and processes in agriculture and animal husbandry. Resource savings and returns depend on location, technology and implementation.
Application & enablement
Digital onboarding
Digital onboarding combines learning paths, checklists, documents and a question-answering assistant. It does not replace personal contact but secures knowledge that otherwise lives with individual employees.
Application & enablement
Digital Transformation
Digital transformation goes beyond the mere digitization of paper documents. It includes the fundamental modernization of working methods, customer interactions and value chains using modern cloud, automation and AI technologies.
Process automation
Digital Twin
A digital twin maps selected properties of a real object or system. Data can update the displayed state; models can support analysis and simulations. Update frequency and accuracy depend on the purpose and are not automatically real-time or error-free.
AI fundamentals & models
Dijkstra's Algorithm
Proceeding from the starting node, the not yet closed node with the smallest known distance is processed in each case. Assuming non-negative edge weights, shortest paths can thus be determined. The weights must match the desired size, such as distance or travel time.
AI fundamentals & models
Dilated Convolution (Atrous Convolution)
A dilated convolution increases the distances between the scanning positions of a filter. With the same filter size, a larger context is thereby detected without additional filter parameters. Whether the output resolution remains the same also depends on stride and padding.
AI fundamentals & models
Dimensionality Reduction
Linear methods such as PCA and non-linear methods such as t-SNE or UMAP produce low-dimensional representations. They can facilitate analysis and visualization. The benefit for a subsequent model must be verified; visible clusters are not an automatic proof of real groups.
AI fundamentals & models
Direct Preference Optimization (DPO)
DPO formulates the optimization problem of RLHF mathematically: Instead of training a separate reward model and optimizing it via reinforcement learning, DPO adapts the language model directly to paired preference data via a simple classification loss function.
Process automation
Directed Acyclic Graph (DAG)
A directed acyclic graph consists of nodes and directed edges. Acyclic means: If you follow the arrows, you never get back to the starting node via a closed path. Workflow systems use DAGs to derive permitted execution orders from dependencies.
Web design & development
Disaster Recovery
Disaster Recovery includes policies, tools, and automated backups to quickly restart business-critical systems, databases, and web applications after natural disasters, cyberattacks, or hardware defects.
Application & enablement
Disparate Impact
In US selection procedures, the ratio of the selection rates may indicate adverse effects. The four-fifths rule is a rule of thumb and does not definitively decide on unlawful discrimination. The German concept of indirect disadvantage according to § 3 para 2 AGG is related, but not identical.
Visibility & SEO
Display Advertising
Display advertising shows visual advertisements on websites and in apps. The placement can be done via advertising networks, direct bookings or programmatic systems. Goals can be reach, brand awareness or measurable actions; display advertising is not synonymous with programmatic purchasing.
AI fundamentals & models
Distribution Shift
Changes can affect inputs, targets or their relationship. As a result, a model can work worse, but does not have to lose quality with every change. Data distribution and actual result quality should be monitored separately.
AI fundamentals & models
Diversity in Recommender Systems
Classic recommendation systems tend to accumulate very similar hits (echo effect). Diversification algorithms balance relevance and diversity to show users new categories and avoid monotony.
Process automation
Document AI
In general, OCR, layout analysis and AI models support document content classification and extraction. Results can be output in a structured manner. A transfer to ERP or other specialist systems requires a separate integration; errors and missing information must be handled.
AI fundamentals & models
Double Machine Learning (DML)
ML methods estimate auxiliary variables, such as outcome and treatment models. Orthogonal estimation equations and cross-fitting can reduce sensitivity to estimation errors and overfitting. Causal identification and further model assumptions remain necessary.
AI fundamentals & models
DPO Training Pipeline
A DPO workflow includes the selection of a suitable starting model, curated pairs of preferences, the appropriate reference configuration, and training. Separate evaluation data examines quality and possible deteriorations. Data approval, configuration and results should be documented in a comprehensible manner.
AI fundamentals & models
DreamBooth
DreamBooth adjusts a pre-trained text-to-picture model based on fewer images of a specific subject. A special identifier in the prompts connects the motif with the model; for this, a new vocabulary entry does not necessarily have to be generated. Prior Preservation can support the preservation of general classes, but does not guarantee it.
AI fundamentals & models
DROP Benchmark
DROP is an English-language question-and-answer dataset for text sections. For example, some tasks require counting, adding or sorting. The benchmark measures performance on these test cases and proves neither general understanding nor reliability on operational documents.
AI fundamentals & models
Dropout
During training, randomly selected activations are set to zero. This can reduce overfitting; the benefits depend on the model and settings. Many implementations are already scaling in training, so that dropout does not make any further changes in the normal evaluation mode.
Web design & development
Dynamic Batching
An inference server bundles available requests into compute stacks. An optional waiting window can allow for larger stacks, but increases the waiting time. Throughput and latency depend on model, hardware and load.
Visibility & SEO
Dynamic Creative
Dynamic advertising tools use reusable templates or assets. The selection and combination can be based on product feeds, playout context or available target group signals. Responsive displays are a documented example; supported signals and functions differ between platforms.
Visibility & SEO
Dynamic Creative Optimization (DCO)
DCO generates variants from common advertising components and adjusts individual elements based on target group, context or measured performance. The goal is a better fit to the defined campaign goal. A certain click-through rate, conversion rate or return is not guaranteed.
Process automation
Dynamic Pricing
Rules or models adjust prices to data such as demand, inventory, or competitive prices. Companies set goals and limits; higher margins or optimal sales are not guaranteed.
Web design & development
E2B Sandbox
On request, E2B provides Linux VMs where agents can run code and tools. SDKs control start, commands and life cycle. Isolation reduces risks; permissions, network and sensitive data must be additionally secured.
Data integration & search
E5 Embedding Model
E5 was contrast trained with weakly monitored text pairs and refined on supervised tasks. The specific variant determines language, input format and context length. For multilingual-e5-large, query: and passage: are documented input processing.
AI fundamentals & models
Early Stopping
Often the loss of validation is monitored. Patience determines the number of epochs without sufficient improvement; min_delta sets the minimum change. The procedure can limit overfitting and unnecessary training, but it does not safely prevent them.
Application & enablement
Economics of AGI
Research explores how far-reaching AI automation could change tasks, productivity, income and capital-to-labour ratios. Assumptions about capabilities, dissemination and resources determine the results; this does not lead to a safe AGI forecast.
AI fundamentals & models
Edge AI
Models perform inference on devices or servers near the data source. This can reduce transmission times and data traffic. Offline operation requires locally available models and dependencies; security follows from the entire implementation.
Web design & development
Edge Computing
Calculation and storage take place close to data sources or users, for example on local devices, regional servers or distributed cloud nodes. This can reduce latency and bandwidth requirements; reliability depends on local dependencies and synchronization.
AI fundamentals & models
Effect Size
Effect measures such as mean value difference, Cohen's d, correlation or odds ratio describe the size and direction of an effect. The p-value evaluates data under a null hypothesis and is not a measure of effect size. Practical relevance and uncertainty must be assessed separately.
AI fundamentals & models
ElevenLabs (AI Voice Generation)
ElevenLabs offers generative audio features, including text-to-speech and voice cloning. Quality and naturalness depend on model, language, text and reference recording. Use of voices and generated content requires appropriate rights and conditions.
AI fundamentals & models
Elo Rating (in AI Evaluation)
Classic Elo updates relative ratings based on wins, defeats or draws and expected results. LLM evaluations use pair comparisons, but not every ranking calculates classic Elo: The Arena work of 2024 describes Bradley-Terry coefficients and confidence intervals.
Visibility & SEO
Email Marketing
Email marketing includes newsletters, offer mails and automated information channels. Purely transactional messages serve another main purpose. Segmentation and automation can improve communication; cost-effectiveness and deliverability must be measured.
Data integration & search
Embedding
The mathematical vector representation of texts, images or audios in multidimensional space to determine semantic similarities.
AI fundamentals & models
Embedding Model
Depending on the model, it maps text, images or other data in a vector space. Semantically similar content should be represented similarly according to the training goal. Model, version and task determine how distances or similarity scores are to be interpreted.
AI fundamentals & models
Encoder
It processes raw data or tokens into features or latent states that a decoder or downstream head uses. The representation does not have to be smaller than the input.
AI fundamentals & models
Encoder–Decoder Model
The encoder maps the input internally; the decoder gradually generates the output from it, often using Attention.
AI fundamentals & models
Energy-Based Model
Low energy stands for preferred or plausible combinations; probabilities can be derived from this, but do not always have to be calculated in a normalized manner.
Visibility & SEO
Engagement Rate
The formula depends on the platform. In GA4, dedicated sessions are divided by sessions. A session is engaged if it exceeds the configured engagement time limit (ten seconds by default), contains a key event, or contains at least two page or screen views. For web data streams, the time limit is adjustable.
AI fundamentals & models
Ensemble Learning
Bagging, boosting and stacking combine several basic models, for example to reduce variance or improve prediction quality. Models and errors do not have to be independent; the benefits are tested by independent validation.
Data integration & search
Entity Extraction
At NER, text spans such as people, organizations or places are marked and assigned to classes. Technical information such as contract numbers requires a suitable scheme, rules or training; general NER models do not automatically recognize them.
Data integration & search
Entity Linking
After recognizing a mention, the system decides on the basis of the context which unique data set is meant.
Data integration & search
Entity Resolution
Rules or probabilistic models compare features and decide whether different spellings or sources belong together.
Visibility & SEO
Entity-Based SEO
An editorial approach to SEO that describes topics, people, organisations and products unambiguously and connects relevant information. Structured data can supplement these descriptions. This establishes neither a separate confirmed ranking factor nor guaranteed search positions.
Web design & development
Envelope Encryption
A data encryption key (DEK) encrypts useful data; a key encryption key (KEK) encrypts the DEK. A KMS can manage KEKs centrally. Key rotation and re-envelopment must be deliberately planned.
AI fundamentals & models
Epistemic and Aleatoric Uncertainty
Epistemic uncertainty can decline with more appropriate data or better models; aleatory uncertainty stems from inherent noise or ambiguity.
Application & enablement
Equalized Odds
Within each actual target class, the prediction should be independent of protected group membership. The criterion requires ground truth and is only one fairness measure.
AI fundamentals & models
Error Analysis
Errors are categorized and prioritized by type, cause, subgroup, data source or process phase. It goes beyond a single overall metric.
Data integration & search
ETL
Source systems are read out; data is cleaned up, unified and checked; only then do they reach the data warehouse, for example.
Application & enablement
EU AI Act
The regulation distinguishes prohibited practices, high-risk systems, transparency obligations and rules for general purpose AI. Duties depend on the role, system and context of deployment.
AI fundamentals & models
Euclidean Distance
It is the square root of the sum of squared coordinate differences and depends on scaling and dimensions.
Web design & development
Eval-driven Development
Test cases and metrics are created early, changes are repeatedly checked and regressions are made visible before the rollout. The term is a practice, not a uniform standard.
AI fundamentals & models
Evaluation Framework
It connects goals, test cases, data sets, metrics, baselines, thresholds, human review, and approval rules. It is a process framework, not necessarily a software product.
AI fundamentals & models
Evaluation Harness
The technical environment loads models or systems, test cases and metrics, logs configurations and generates results for comparisons. Comparability requires controlled versions and conditions; the environment does not replace the evaluation concept.
Visibility & SEO
Event Tracking
An event model describes names and properties as well as the capture context, such as form sent or process failed. It requires a consistent taxonomy, data quality controls and a review of data protection and, if necessary, consent.
Web design & development
Event-Driven Architecture
Producers report state changes; brokers or streams distribute them to independent consumers. Processing can take place asynchronously and several times.
AI fundamentals & models
Evidence Lower Bound (ELBO)
The ELBO is at most as large as the logarithmic evidence. Their gap corresponds to the KL divergence between approximated and true posterior. Maximization optimizes an approximate target size; it does not guarantee exact evidence maximization and minimal prediction errors.
Visibility & SEO
Exit Rate
A common calculation is exits of a page divided by their page views times 100. The exact definition depends on the tool; GA4 documents in particular the absolute metric Exits, i.e. the last measured action of a session on a page or a screen.
AI fundamentals & models
Expected Calibration Error
Predictions are usually divided into bins according to their confidence. For each bin, the absolute difference between mean confidence and observed accuracy is calculated and weighted by the number of his cases. Result and statement depend on binning and variant.
Web design & development
Experiment Tracking
Experiment tracking assigns input, configuration, metrics and artifacts to a run. Together with data and code versions, it supports comparison and repetition; complete reproducibility does not automatically occur.
UX & accessibility
Explainability UX Patterns
Explanations should fit the user role, task and decision. For example, you can show reasons, comparative examples and uncertainty in an understandable order and check with users whether meaningful actions follow.
AI fundamentals & models
Explainable AI (XAI)
Explanations can be global or local and show features, examples, rules or uncertainty. Explanation, interpretability and transparency are related, but not identical.
AI fundamentals & models
Exploration–Exploitation Dilemma
Too much exploration costs short-term performance; too much exploitation prevents learning. Bandit and RL processes control this trade-off.
Data integration & search
Exploratory Data Analysis
Visualizations, distributions, relationships, missing values and outliers help to understand data quality and hypotheses. EDA does not confirm causality.
Web design & development
Exponential Backoff
After temporary errors, the interval typically grows exponentially and is often complemented by random jitter to avoid simultaneous repetitions.
AI fundamentals & models
Exponential Linear Unit (ELU)
For positive inputs, f(x)=x, otherwise f(x)=alpha·(exp(x)−1). At alpha>0, the negative output approaches −alpha. Negative activations can bring the mean closer to zero; training benefits depend on architecture and task.
AI fundamentals & models
Exponential Moving Average (EMA)
In the usual convention, EMA_t = α · x_t + (1 − α) · EMA_(t−1), with a smoothing factor between 0 and 1. A larger α weights the new value more strongly; initialization and convention used must be established.
AI fundamentals & models
Exponential Smoothing
Variations take into account level, trend and seasonality. Statistical ETS models provide not only point forecasts but also forecast intervals; suitable components and error assumptions must be selected.
Data integration & search
Extract, Load, Transform (ELT)
ELT loads source data into a target data platform before the professional transformation. There they are processed, for example, with SQL and dbt. Raw data management, access rights and retention are conscious architectural decisions; source data is not necessarily completely unchanged.
AI fundamentals & models
F1-Score
It is high when positive predictions are simultaneously precise and capture many actual positive cases. The exact averaging over classes must be given.
AI fundamentals & models
Faithfulness
In RAG, faithfulness typically checks whether statements can be derived from the retrieved sources; for explanations, the term can refer to fidelity to the model. The definition must be specified.
Web design & development
Fault Tolerance
Redundancy, isolation, repetition, replication and orderly degradation limit impacts. Fault tolerance is not the same as high availability or disaster recovery.
AI fundamentals & models
Feature Engineering
Expertise is used to clean, encode, aggregate or meaningfully combine raw data without injecting future information.
AI fundamentals & models
Feature Extraction
It can be rule-based, statistical or learned. For example, text, images or signals are converted into numerical characteristics; the number of dimensions does not have to become smaller.
AI fundamentals & models
Feature Importance
Methods measure different things, such as split gain, permutation loss, or local contribution values. Results are model- and data-dependent and not automatically causal.
Data integration & search
Feature Store
A feature store manages feature definitions and their deployment. Depending on the product, it includes transformations, versioning and access controls, and offline and online storage; consistent definitions can reduce training serving skews.
Visibility & SEO
Featured Snippet
For a featured snippet, Google displays the descriptive section before the usual page reference. The Google systems decide whether a page is suitable for the specific search query. A website cannot label itself as a featured snippet.
AI fundamentals & models
Federated Learning
Participants calculate local updates; often a central coordinator aggregates them into a common model. This avoids the central collection of raw data, but alone does not guarantee privacy or security.
Process automation
Feedback Loop
Feedback can stabilize or amplify. In AI systems, decisions often influence later data and thus future models.
AI fundamentals & models
Feedforward Neural Network
A feedforward network has no directed feedback cycles: information is processed from input to output. The multilayer perceptron is a typical variant; weights are often optimized with backpropagation during training.
AI fundamentals & models
Few-Shot Learning
The term includes procedures that learn from a few examples of a new class or task. In language models, prompting with few demonstrations is often referred to as such, without a change in weight.
AI fundamentals & models
Fine-Tuning
Subsequently adapting an existing model with specific data sets to specialize in specialist domains, tasks or brand tones.
Web design & development
Finite-State Machine
Inputs or events trigger defined changes of state; variants differ, for example, by output and determinism.
Web design & development
FinOps
According to the FinOps Foundation’s definition, updated in March 2026, FinOps is an operational framework and cultural practice for timely, data-based decision-making and shared cost responsibility. In addition to the cloud, the current framework also covers AI, SaaS, data platforms, licenses and data centers.
Data integration & search
First-Party Data
This includes, for example, purchases, service contacts, website events or voluntarily communicated preferences from direct customer relationships. As far as personal data are concerned, the collection and use must in particular be lawful, purpose-related and necessary. „Own data does not mean free use.
AI fundamentals & models
FlashAttention
The algorithm arranges computation steps so that less data is moved between GPU main memory and fast on-chip memory. It does not approximate the mathematical attention; small numerical deviations by computational order remain possible.
AI fundamentals & models
Flow Matching
A neural vector field is trained to transport data along conditioned paths from a simple initial distribution to data distribution; the training does not require a simulation of the complete process.
Web design & development
FLUX
FLUX refers to a family of image-generation models. Licences, open components and available functions differ between variants.
AI fundamentals & models
Focal Loss
The loss function developed for object recognition adds a factor to cross-entropy that reduces the contribution of well-recognized examples; an alpha weight can address class imbalance.
UX & accessibility
Focus Group
The group discusses a clearly defined topic; interactions provide motives, language and perceptions. The results are not statistically representative.
AI fundamentals & models
Forward Pass
With the forward pass, outputs are calculated from inputs and current parameters. In the usual training, loss calculation, backpropagation and an optimization step follow. Inference uses forward passes, often repeated with autoregressive generation.
AI fundamentals & models
Foundation Model
According to Stanford CRFM, the term encompasses large-scale, mostly self-supervised trained models that are versatile through prompting, fine tuning or other procedures. Errors and risks are transferred to many applications.
Visibility & SEO
Frequency Capping
An advertising platform limits impressions or views per assigned user and time window, depending on the campaign type and level. For example, in Google Ads display campaigns, only visible impressions count; incomplete identity assignment can affect the count.
Web design & development
Full-Stack Development
The term typically includes user interface, server logic, APIs, data storage, and deployment. It describes breadth, not automatically deep expertise in every area.
Web design & development
Fully Sharded Data Parallel (FSDP)
PyTorch FSDP is a concrete approach to data-parallel training, in which parameters, gradients and optimizer states are sharded and communicated as needed, depending on the phase.
Visibility & SEO
Funnel Analysis
A funnel analysis counts how many users reach defined steps in the intended order. In an open funnel, it is possible to start at later steps; a closed funnel begins at the first step. Abortions describe the observed behavior, but do not prove a cause.
AI fundamentals & models
Fuzzy Inference System
Inputs are assigned to membership levels, rules are evaluated and results are aggregated; defuzzification often follows. Mamdani and Sugeno systems are different.
Data integration & search
Fuzzy Matching
Unsharp matching compares strings using a similarity method. Power Query uses Jaccard similarity and an adjustable threshold in the Fuzzy Merge. A high score does not replace checking whether two records denote the same entity.
AI fundamentals & models
G-Eval
G-Eval is a 2023 presented framework for evaluating generated texts. A language model generates assessment steps and fills out an assessment form. The original work explores summaries and dialogues, especially with GPT-4. Judgments depend on task, criteria and model.
AI fundamentals & models
Gaussian Distribution
Its density is bell-shaped; many statistical methods use it as a model assumption. Real data does not have to be normally distributed.
AI fundamentals & models
Gaussian Mixture Model
Each component has a mean value and covariance; mixture weights indicate proportions. Parameters are often estimated with Expectation-Maximization.
Application & enablement
GDPR Compliance in AI
Data origin, legal basis, purpose, roles of stakeholders, transparency, data subject rights and technical protection measures shall be clarified. Article 22 refers to certain exclusively automated decisions with legal or similar significant effect; it is not a general transparency requirement for any AI.
AI fundamentals & models
GELU
GELU is defined exactly by GELU(x) = x · Φ(x); Φ is the distribution function of the standard normal distribution. The function weights inputs softly based on their value. Implementations can use numerical approximations.
Application & enablement
Gem
A Gem is a customized assistant in Google Gemini with reusable instructions for specific tasks. Supported features, knowledge materials and approvals depend on the account and current offer.
Application & enablement
General Data Protection Regulation (GDPR)
The GDPR regulates the processing of personal data in the EU. It enshrines principles such as data minimisation, purpose limitation, transparency and data subject rights (right of access, deletion) and threatens sensitive fines in the event of breaches.
AI fundamentals & models
Generalization
A model is generalized when learned patterns are transferred to independent examples under appropriate conditions. Good test performance does not guarantee robustness during distribution shift.
AI fundamentals & models
Generative Adversarial Network (GAN)
In classical GAN, a generator generates examples, while a discriminator is intended to distinguish training data from generated examples. The training goals are in competition. Good quality or variety is not guaranteed by this construction alone.
AI fundamentals & models
Generative AI
Generative models generate content such as text, image, audio, video or code from learned patterns. Expenditure can be factually incorrect or distorted. Check data origin, confidentiality and the terms of use applicable to your use.
Visibility & SEO
Generative Engine Optimization (GEO)
GEO is the preparation of digital content for findability and possible naming in generative response systems. A mention or positive citation cannot be guaranteed.
Visibility & SEO
Generative Search Engine Results Page
A search results page includes an AI-generated answer section, potentially with source links and additional traditional results. Such answer sections supplement or change how results are displayed; they do not necessarily replace all links.
Web design & development
Generative UI
An AI-based process selects, generates or updates matching UI components based on a task and its data. For example, tool results are displayed as a table or map. This can be based on predefined components; usability and accessibility must be tested.
Visibility & SEO
GEO-SFE (Structural Feature Engineering)
The research distinguishes document architecture, information sections and visual emphasis as three levels of structure. It investigates their influence on citations in the generative systems tested. The results are specific to the study and do not guarantee citations or ranking improvements for a particular website.
Visibility & SEO
Geo-Targeting
Geo-targeting uses geographic signals to select content or ads. Google Ads can include, among other things, user settings, device and behavior; target areas can be set as countries, regions or radii. Location signals are not a guarantee of an exact location.
Web design & development
GGUF
It stores tensors and typed metadata in an expandable, mmap-friendly file. GGUF can contain quantized tensors, but is not itself a quantization method.
AI fundamentals & models
GloVe
GloVe learns word vectors from a corpus’ global word-cooking statistics. In classical usage, each word receives a static vector; different meanings of the same word are not separated by the respective sentence.
AI fundamentals & models
Golden Dataset
A reference data set contains verified tasks and expected responses, labels or evaluation criteria. “Golden Dataset” is not a uniform quality certification: coverage, errors and timeliness must be checked separately.
Visibility & SEO
Google AI Overviews
Google supplements selected searches with a summary with further links. The function is a specific Google search offer; it does not appear on every request.
AI fundamentals & models
GPQA
Published in 2023, the dataset contains 448 questions created by professionals. It measures performance on this set, not general intelligence or scientific reliability.
AI fundamentals & models
GPT
The original GPT approach uses a transformer decoder for autoregressive language model pre-training and then adapts it to tasks. GPT is a model family designation and not a single fixed version. A GPT model is not synonymous with a full chatbot application.
AI fundamentals & models
Grad-CAM
Gradients of a target class weight activations of a convolution layer; this results in a coarse salience-like map. It is post hoc and not causal.
AI fundamentals & models
Gradient Accumulation
Gradients of multiple micro-batches are summed or averaged before the optimizer performs an update. Effective batch size and scaling must be documented correctly.
AI fundamentals & models
Gradient Checkpointing
Instead of storing all intermediate values for backpropagation, the training only retains checkpoints and recalculates missing activations. Memory decreases, computing time increases.
AI fundamentals & models
Gradient Clipping
In standard clipping, the entire gradient is scaled when its norm exceeds a limit; value clipping limits individual values. It treats symptoms of unstable updates.
AI fundamentals & models
Gradient Descent
The gradient shows locally the direction of the greatest increase; learning rate and variant determine the updating. Mini-batch-SGD uses estimates instead of the full gradient.
AI fundamentals & models
Graph Attention Network (GAT)
Masked self-attention aggregates neighboring features and can give different weight to relevant connections per task. Weights are not automatically causal explanations.
AI fundamentals & models
Graph Classification
Prediction unit is the entire graph. In GNN methods, node representations are combined to form a graph representation via a readout function. Graph features or graph cores are also possible. A continuous target value belongs to the graph regression.
AI fundamentals & models
Graph Convolutional Network (GCN)
The common Kipf-Welling variant approximates spectral graph convolutions and mixes features of connected nodes via normalized adjacency.
AI fundamentals & models
Graph Database
Graph databases store linked data for relationship and path queries. In the property-graph model, nodes and relationships can carry properties. RDF uses subject-predicate-object-triple. Data model and query language differ by system.
AI fundamentals & models
Graph Isomorphism Network (GIN)
GIN combines sum aggregation of neighboring representations with MLPs. Under the conditions of the original analysis, it can achieve the distinctive power of the one-dimensional Weisfeiler-Lehman test. This is not a guarantee to distinguish all non-isomorphic graphs.
Data integration & search
Graph Neural Network (GNN)
Many GNNs use Message Passing: Nodes collect neighbor information and update their state. Tasks include node, edge and graph predictions.
Web design & development
Graph Transformer
Models combine node and edge features, position or structure coding, and global or local attention. There is not a single binding architecture.
Data integration & search
Graph Traversal
Breadth-first search, depth-first search or domain-specific path rules determine order and termination. Traversal is an operation, not a database or a learning method.
Web design & development
Graphics Processing Unit (GPU)
Many computing units process vector and matrix operations simultaneously. This accelerates graphics, ML training and inference, as long as software and data flow fit.
Data integration & search
GraphSAGE
Instead of storing fixed embeddings per node, the model samples and aggregates local neighbors. As a result, it can be transferred inductively to unseen nodes with a suitable feature scheme.
AI fundamentals & models
Greedy Best-First Search
The priority is based only on the heuristic h(n), not on the previous path costs. This can be fast, but does not guarantee optimal paths or completeness without further ado.
AI fundamentals & models
Grouped-Query Attention (GQA)
The architecture lies between Multi-Head Attention with separate K/V heads and Multi-Query Attention with only one K/V head. The goal is faster decoder inference at lower KV cache.
AI fundamentals & models
Hallucination
The generation of plausible-sounding, but factually false, invented or unsupported statements by an AI model.
Web design & development
Headless CMS
Content management system that manages content decoupled from the frontend and delivers it to any output channel via APIs.
UX & accessibility
Hero Section
The top, immediately visible part of a web page, containing its central message and main call to action.
Web design & development
Higgsfield
Higgsfield is an AI creative service for image and video tasks. Available models, interfaces and terms must be checked with the provider for their current status.
Process automation
Human-in-the-Loop
Quality and governance principle, in which human experts review and approve decisions at critical control points.
Web design & development
HyperText Markup Language (HTML)
The standardised markup language for structuring text, images and elements on the web.
Web design & development
Ideogram
Ideogram is an AI image-generation service. The quality of text within an image depends on the subject, model and prompt.
AI fundamentals & models
Image Classification
An image classification model receives an image and predicts which class or classes it belongs to. Unlike object recognition, it does not locate objects with bounding boxes; unlike segmentation, it does not assign classes to individual pixels.
AI fundamentals & models
Image Generation
In image generation, a generative model generates visual content, for example using text prompts, reference images or masks. Diffusion models are a common approach. Image classification, on the other hand, assigns classes; generative image processing such as inpainting overlaps with image generation.
AI fundamentals & models
Image Segmentation
In image segmentation, not only does the entire image receive a label, but pixels or regions are classified. Semantic segmentation assigns pixel classes; instance segmentation additionally separates individual object instances of the same class.
AI fundamentals & models
In-Context Learning
The ability of a language model to learn new tasks or formats directly from examples provided promptly, without changing the model weights.
Web design & development
Inpainting
The targeted editing, replacement or addition of elements within a masked image area by AI.
Process automation
Instant Checkout in ChatGPT
Instant Checkout combines product search in ChatGPT with a direct purchase. Orders, payments and fulfillment remain with the merchant; technically, the integration is based on the Agentic Commerce Protocol. Product results are separate from ads and are selected organically, according to OpenAI.
Web design & development
JavaScript
JavaScript processes program logic and data, such as form logic and dynamic views. It runs in browsers and outside them in suitable runtime environments. Browser APIs and server-side functions must be distinguished.
AI fundamentals & models
Joint Distribution
It maps dependencies between variables; marginal and conditional distributions can be derived from this. Correlation is only a limited aspect of this relationship.
Data integration & search
Knowledge Base
For example, a knowledge base bundles articles, FAQs and work instructions. It can be used internally or publicly and serve as a source for a search or an assistant. Timeliness, responsibilities and access rights determine their reliability; an error-free AI response does not automatically follow from this.
AI fundamentals & models
Knowledge Distillation
In the teacher-student process, information from a teacher model or ensemble is used for training a student model, such as probability distributions. A smaller model can be cheaper to operate, but does not automatically reach the quality of the teacher.
AI fundamentals & models
Large Language Model (LLM)
A large language model is trained on extensive data and processes language based on learned statistical patterns. During text generation, it selects successive tokens. Liquid outputs can be factually wrong; the task and model variant determine the suitability.
Web design & development
Lazy Loading
Lazy loading defers the loading of suitable resources until they are needed. Images outside the initially visible area are a common example; loading may begin before they become visible. Delaying primary content that is needed immediately is usually unsuitable.
Visibility & SEO
llms.txt
llms.txt is a proposed text format that allows websites to find selected content for AI systems. A general support or ranking effect is not guaranteed.
AI fundamentals & models
Long-Context Processing Cost
Long inputs must be processed and taken into account during generation. The effort depends on architecture, infrastructure, caching and tariff. The raw designation Context-Window-Tax is not a standardized tax or metric. Calculation effort and possible quality losses shall be measured separately.
Web design & development
Lovable
Lovable supports the creation and revision of web applications based on natural language requirements. Frontend, backend connections and publication depend on the implementation and the integrations used. A generated design is only ready for production after suitable examinations.
Visibility & SEO
Machine Legibility
Machine legibility describes whether software can process content in a suitable format and evaluate its structure or meaning. Examples include semantic HTML and structured data. It also concerns systems outside AI and does not guarantee correct understanding.
Visibility & SEO
Marketing Attribution
Marketing attribution distributes the attributed contribution to a promotion, such as a request or a purchase, to contacts in the recorded user path. An attribution model can use rules or data-driven methods for this. Also observed AI references can be included; unrecorded contacts remain invisible. Attribution is not causal evidence.
Process automation
MCP Agent
Here, MCP agent refers to an AI agent that uses tools or data connected via an MCP client. MCP describes the connection, not the planning and execution of the agent. Access permission and control are in the respective application.
Process automation
MCP Registry
The official MCP Registry is a directory of metadata about MCP servers and their deployment. Other organisations may also maintain their own directories. Discoverability and name verification do not replace checks of code, security, permissions or suitability.
Process automation
MCP Server
An MCP server provides MCP clients with defined tools, resources or prompts. It can connect local or external systems for this purpose. Which accesses are actually allowed, must set hosts, servers and their operators separately.
UX & accessibility
Microinteraction
Microinteractions support a limited individual task, such as a copy confirmation or progress indicator. Triggers can come from users or the system. Feedback need not be exclusively visual; accessibility and reduced motion must be considered.
Process automation
Microsoft Copilot Studio Agent
With Copilot Studio, agents can be created for defined tasks and integrated into suitable channels. Sources of knowledge, tools and orchestration determine the skills. Copilot Agent does not refer to a universal architecture; functionality, permissions and billing depend on the specific Microsoft solution.
Web design & development
Midjourney
Midjourney is a service that generates images from inputs. Style, quality and usage rights depend on the current product and plan.
Web design & development
MLOps on the Edge (Edge MLOps)
Edge MLOps includes versioning, deployment and monitoring of models and runtimes on distributed devices. Compilation, quantization and OTA updates can be included; hardware, connections and restarts must be considered.
UX & accessibility
Mobile-First Approach
A design philosophy in which the design is first conceived for small mobile screens and then expanded for larger screens.
Process automation
Model Context Protocol (MCP)
Open communication standard for the structured connection of language models with local or external data sources and tools.
Process automation
Model Router
A model router forwards requests to selected models or deployments based on defined strategies. Criteria can be cost, latency, availability and task characteristics. The selection does not automatically guarantee the technically best result.
Process automation
Multi-Agent Orchestration
Structured coordination and division of labor of several specialized AI agents to solve complex overall tasks.
AI fundamentals & models
Multimodality
A multimodal system processes or generates more than one type of data, such as text and image, or audio and video. Which modalities can be combined depends on the model and interface; not every variant supports all types of media.
Application & enablement
Negative Prompt
A negative prompt specifies unwanted features for a generative model that supports this input. It influences the output but does not guarantee exclusion. Not all models or interfaces support this additional input.
AI fundamentals & models
Open-Source Models
According to the Open Source AI definition of the OSI, the prerequisites for use, investigation, modification and disclosure must be available. This includes appropriate code, model parameters and sufficient data information. Publicly available weights or a free downloadable model do not automatically meet these requirements; the decisive factor is a specific license and components provided.
Application & enablement
Payback Period
The amortisation period specifies when the cumulative returns cover the investment. In the case of the static calculation with constant annual net returns, the investment is divided by this amount. A dynamic calculation additionally takes into account the time value of money.
Process automation
Payment Infrastructure for AI Agents
In addition to a technical payment connection, an agent needs a verifiable order and limited powers. AP2 is a protocol approach to this; it is neither a currency nor even a payment network.
UX & accessibility
Progressive Disclosure
Important core information is prominently displayed, while in-depth details, settings or explanations only become visible during interaction (for example by clicking on 'Learn more' or accordion). This simplifies the user guidance.
Application & enablement
Prompt
A prompt is the instruction to an AI system. Well-written prompts include context, role, desired format and examples. Recurring tasks can be documented in a prompt library.
AI fundamentals & models
Prompt Caching / Prefix Caching
Prompt caching can cache recurring input parts. Whether time or cost decrease depends on provider, model, hit rate and tariff.
Application & enablement
Prompt Engineering
With clear instructions, appropriate context, examples and output specifications, results can be controlled in a more targeted manner. Quality is checked on representative tasks; errors remain possible.
AI fundamentals & models
Prompt Injection
Prompt injection involves placing instructions in untrustworthy content to influence the behavior of an AI system. Such attacks can circumvent protection rules, but do not override every system.
Application & enablement
Prompt Sandbox
A prompt sandbox is a descriptive term for a test environment for AI statements. There, response quality and error cases can be examined before use. Technical isolation, data access and protection measures must be set up and tested separately.
Process automation
Quality Gates
Quality Gates are stipulated terms and conditions under which a result is accepted, deferred or rejected. Testing can be done automatically or manually. In AI workflows, sources, expertise and approval can be included; even a passed gate does not guarantee freedom from errors.
Application & enablement
Quick Win
A Quick Win should enable early concrete success with limited effort. It can support a larger change process. Benefits, time requirements and acceptance must be measured on a case-by-case basis; immediate productivity gains are not a promise.
UX & accessibility
Responsive Web Design
A web design approach in which the layout flexibly adapts to the screen size of the device being used.
Data integration & search
Retrieval-Augmented Generation (RAG)
An RAG system retrieves appropriate information via a search component and passes it on to the language model as context. Among other things, keyword, vector or hybrid searches are possible. Mistakes and hallucinations remain possible.
AI fundamentals & models
Sampling Temperature
Temperature is a parameter of the probability distribution for token selection in supporting language models. Lower values tend to lead to less variation, higher ones to more variation. Neither factual fidelity nor identical results are guaranteed.
Visibility & SEO
Search AI Answers
AI answer blocks integrated directly into search interfaces that bring together information from different sources.
Web design & development
Semantic HTML / Markup
Semantic markup uses HTML tags such as <header>, <nav>, <main>, <article>, <section> and correct heading hierarchies (h1–h6). This gives browsers, search engines and screen readers the logical meaning of the content.
Web design & development
Server-Side Rendering (SSR)
Server-side rendering creates HTML on the server. This can support accessibility for crawlers and the first page setup, but it does not guarantee good loading times or a good search ranking.
Process automation
Service-Level Agreement for AI Agents
An SLA describes agreed performance targets and consequences in the event of non-performance. This must be distinguished from SLI as the measured variable and SLO as the target value. Agent-specific quality criteria must also be defined.
Visibility & SEO
Share of Answer
Share of Answer is a visibility metric with a provider-specific definition. One possible calculation is the proportion of tested AI answers that mention a particular brand; other definitions refer to all brand mentions. The denominator, question set, systems and time of measurement must therefore be documented.
Visibility & SEO
Share of Model (SoM)
Providers use Share of Model to describe proportions of brand presence in AI answers. The reference quantity may be answers or brand mentions within a defined sample. Values are not comparable without the exact provider definition, question set, models and measurement period.
Web design & development
Single Page Application (SPA)
A SPA first loads an HTML document and then updates views dynamically without reloading the full document for each change. It can have multiple URLs and routes; initial loading time and usability must be checked.
Visibility & SEO
Structured Data (Schema.org)
Structured data describes objects and properties in a machine-readable format. Schema.org is a widely used vocabulary; JSON-LD is one of the markup formats. Markup must match the visible content and does not guarantee rich results.
Visibility & SEO
Structured Product Data
Structured information describes, for example, product, price and availability. Product markup can be used for this on websites; data feeds are another means of delivery.
Web design & development
Subject Consistency
Subject consistency describes how well the characteristic features of a person, character or product are preserved across several generated images. References or model adaptation can help. This does not guarantee pixel-identical or completely identical representations.
Application & enablement
System Prompt
A system prompt gives a supporting AI application higher-level instructions on role, task framework and response behavior. Role priority and implementation depend on the model. It does not replace technical access control and is not an insurmountable security boundary.
Web design & development
Technical Debt in AI Systems
In addition to code, data dependencies, configurations and feedbacks can make subsequent changes more difficult. Rapid introduction can thus cause long-term follow-up costs.
Web design & development
Text-to-Video Generation
AI technology for generating moving video sequences directly from textual descriptions.
Web design & development
Throughput
Throughput measures processed work per unit of time, such as requests per second or generated tokens per second. In the LLM comparison, it must be clear whether input or output tokens and the overall system or a single request are measured. High throughput does not guarantee low response time or low overall costs.
AI fundamentals & models
Token
A token is a processing unit of the model. How text is divided into tokens and billed depends on the respective service and model.
AI fundamentals & models
Token Budgeting
Token budgeting plans and limits the scope of inputs, outputs and, where applicable, internal processing tokens. Context boundaries, pricing model and sufficient space for the task must be taken into account. Input length alone does not determine all costs.
AI fundamentals & models
Tool Use / Function Calling
A model answer identifies a tool and its arguments. The connected application checks permissions, executes the call, and returns the result. The model itself does not automatically have access to an API or database.
AI fundamentals & models
Training Epoch
In the usual mini-batch training, the training examples are run once per epoch. The number of parameter updates also depends on batch size and update strategy; data augmentation can change the specific input.
Application & enablement
Transparency Obligations under Article 50 of the AI Act
Article 50 of the EU AI Regulation regulates, among other things, information in the case of direct AI interaction as well as labelling of certain AI content. Provider and operator obligations differ. Machine-readable marking and visible disclosure are different requirements; scope, exceptions and transitional rules must be examined.
UX & accessibility
User Experience (UX)
A user's overall experience and perception when interacting with a digital application.
UX & accessibility
User Journey
The sequence of all touchpoints a user has with a brand on the way to their goal.
UX & accessibility
UX/UI
UX describes how an application feels and how easily tasks can be completed with it. UI describes the interface through which users interact with the application. Together they influence whether a system is accepted.
Data integration & search
Vector Database
A vector database manages embeddings and searches for similar entries based on a suitable distance or similarity measure. Additional metadata can enable filters. Speed and hit quality depend on data, index and configuration.
Web design & development
Vibe Coding
In the approach described by Karpathy, programs are created primarily via natural language instructions and repeated trials; the code generated is not necessarily understood in detail. This can be distinguished from systematic AI-supported development with code review.
Visibility & SEO
Virtual Influencers
Virtual influencers are digitally designed and used for contributions or brand communication. AI can help; however, a virtual figure is not automatically AI-driven or autonomous.
Visibility & SEO
Web Analytics
Web analysis collects and examines suitable usage data of a website, such as page views and defined events. Objectives are, for example, the understanding of visit routes and the examination of whether desired actions take place. The significance and scope depend on the measurement concept and data collection.
UX & accessibility
Web Content Accessibility Guidelines (WCAG)
The Web Content Accessibility Guidelines (WCAG) define technical and design criteria at conformity levels A, AA and AAA. They are based on four principles: perceptibility, usability, comprehensibility and robustness.
Web design & development
Webflow
Webflow is a visual website builder with CMS and publishing features. Export options and code quality depend on the current product and its implementation.
Web design & development
WebP and AVIF
WebP and AVIF are image formats for compressed web images. Both offer different quality and file-size options depending on the method used. The appropriate format depends on the image content, encoder, desired properties and browser support.
AI fundamentals & models
Weight Initialization
Initialization methods set start values and often take layer size and activation into account. Xavier/Glorot and He initialization should support appropriate scaling of signals and gradients. They do not prevent unstable training in all conditions.
UX & accessibility
Whitespace
Whitespace is empty space between and around design elements, regardless of its colour. It can support grouping, hierarchy and readability; its usefulness depends on the layout and context.
UX & accessibility
Wireframe
A schematic, structural draft of a web page without final colours and graphics, used to plan the arrangement of content.

Room for knowledge

Missing a term?

Tell me which term you would like explained clearly.

Suggest a term Go to knowledge search

Where would you like to start?

Data integration & search

Process automation

Application & enablement

AI fundamentals & models

Web design & development

Visibility & SEO

UX & accessibility

Every term in context

Learning paths

AI in five terms

Find knowledge at work

Website and visibility

Recently reviewed

Connected knowledge

By letter

All terms

Abductive Logic Programming (ALP)

Abductive Reasoning

Ablation Study

Abstract Data Type

Abstraction

Accessibility

Accountability

Accountability for AI Agents

Accuracy

Action Language

Action Model Learning

Action Schema

Action Selection

Activation Function

Active Learning

Actor–Critic

Ad Bidding

Ad Exchange

Ad Rank

Adafactor

AdaGrad

Adam Optimizer

AdamW

Adaptive Algorithm

Adaptive Learning

Adaptive Neuro-Fuzzy Inference System (ANFIS)

Adaptive User Interface

Adaptive Web Design

Admissible Heuristic

Adversarial Attacks

Adversarial Robustness

Agent Architecture

Agent Card

Agent Guardrails

Agent Handoff

Agent Loop

Agent Marketplace

Agent Memory

Agent Mesh

Agent Observability

Agent Orchestration

Agent Passport

Agent Payments Protocol (AP2)

Agent Reputation

Agent Sprawl

Agent Swarm

Agent Washing

Agent-to-Agent Communication

Agent-to-Agent Protocol (A2A)

AgentBench

Agentic AI

Agentic Analytics

Agentic Checkout

Agentic Coding

Agentic Commerce

Agentic RAG

AgentOps

AI Act Compliance

AI Agent Development Framework

AI Agent Execution Trace

AI Agent Identity Verification

AI Agent Red Teaming

AI Agents