AI Glossary

Activation Function

A function used in a neural network to determine whether and how strongly a neuron should be activated. Common examples include ReLU, Sigmoid, and Tanh.

Active Learning

A machine learning approach where the model identifies which data points would be most useful to label, allowing humans to focus their labeling effort on the most informative examples.

Adversarial Attack

An attack designed to manipulate an AI model by providing specially crafted inputs that cause the model to make an incorrect prediction or produce an undesirable output.

Adversarial Example

An input intentionally modified in a subtle way to cause an AI model to produce an incorrect result.

Agent / AI Agent

An AI system that can perceive information, reason about a goal, make decisions, and take actions, often by using tools or external systems.

Agentic AI

AI systems designed to operate with a degree of autonomy by planning, making decisions, using tools, and executing multiple steps toward a goal.

AGI — Artificial General Intelligence

A hypothetical form of AI capable of performing a broad range of intellectual tasks across different domains at a level comparable to or beyond humans.

AI — Artificial Intelligence

The field of creating computer systems capable of performing tasks that normally require human intelligence, such as reasoning, learning, perception, language understanding, and decision-making.

AI Alignment

The field concerned with ensuring that AI systems behave according to human intentions, values, goals, and safety requirements.

AI Assistant

An AI-powered system designed to help users perform tasks such as answering questions, generating content, analyzing information, or interacting with software.

AI Automation

The use of AI to automate tasks or workflows that traditionally require human intervention, particularly tasks involving language, perception, reasoning, or decision-making.

AI Benchmark

A standardized dataset, task, or evaluation procedure used to measure and compare the performance of AI models.

AI Bias

A systematic tendency in an AI system to produce results that unfairly favor or disadvantage certain groups, individuals, or outcomes.

AI Copilot

An AI assistant designed to work alongside a human by providing suggestions, generating content, analyzing information, or assisting with tasks while the human remains involved in the process.

AI Governance

The policies, processes, roles, controls, and standards used to manage AI systems responsibly throughout their lifecycle.

AI Hallucination

A situation where an AI model generates information that is incorrect, fabricated, or unsupported by the available evidence, while presenting it as if it were factual.

AI Inference

The process of using a trained AI model to generate a prediction, classification, or response from new input data.

AI Lifecycle

The complete process through which an AI system is designed, developed, evaluated, deployed, monitored, maintained, and eventually retired.

AI Model

A computational model trained to perform one or more AI tasks, such as classification, prediction, generation, or language understanding.

AI Safety

The field concerned with preventing AI systems from causing unintended, harmful, or unsafe outcomes.

AI-as-a-Service (AIaaS)

A model in which AI capabilities are provided through a cloud service or API rather than being built and operated entirely within an organization’s infrastructure.

Algorithm

A defined sequence of instructions or computational steps used to solve a problem or perform a task.

Algorithmic Bias

Systematic bias introduced or amplified by an algorithm, potentially resulting in unfair or inaccurate outcomes.

Annotation / Data Annotation

The process of adding labels, tags, or other information to data so that it can be used to train or evaluate an AI model.

API — Application Programming Interface

A defined interface that allows different software systems to communicate and exchange data or functionality.

API Call

A request sent from one software system to an API in order to retrieve data, execute an operation, or obtain an AI-generated result.

API Gateway

A service that acts as an entry point between clients and backend services, often handling authentication, routing, rate limiting, logging, and security.

Architecture

The overall technical structure of an AI system, including its models, databases, APIs, applications, data pipelines, infrastructure, and integrations.

Artificial Neural Network (ANN)

A machine learning model inspired by the structure of biological neural networks and composed of interconnected computational units called neurons.

Attention Mechanism

A mechanism that allows a model to assign different levels of importance to different parts of an input when processing information.

Augmented Generation

A generation approach where an AI model uses additional information retrieved from external sources to improve the relevance or accuracy of its output.

Autonomous Agent

An AI agent capable of independently deciding and executing actions within a defined environment and set of constraints.

Automation

The use of technology to perform tasks or processes with reduced human intervention.

AutoML — Automated Machine Learning

Tools and techniques that automate parts of the machine learning development process, such as model selection, feature engineering, and hyperparameter optimization.

Availability

The degree to which an AI system or service is operational and accessible when users need it.

Average Precision

An evaluation metric commonly used in information retrieval and classification tasks to measure the quality of ranked results, particularly across different recall levels.

Backpropagation

A training algorithm used to calculate how much each parameter in a neural network contributed to an error, allowing the model to update its parameters and improve its predictions.

Bag of Words (BoW)

A text representation method that represents a document based on the words it contains, typically ignoring word order and grammar.

Batch

A group of data samples processed together by a machine learning model during training or inference.

Batch Size

The number of training examples processed together in one iteration during model training.

Batch Processing

A method of processing data in groups at scheduled or predefined intervals rather than processing each item immediately.

Bayesian Network

A probabilistic graphical model that represents relationships between variables and uses probability to reason about uncertain events.

Beam Search

A search algorithm commonly used in generative AI to generate sequences by keeping several of the most promising candidate sequences at each step instead of considering only one.

Benchmark

A standardized dataset, task, or evaluation procedure used to measure and compare the performance of AI models or systems.

Benchmark Dataset

A dataset specifically designed to provide a standardized basis for evaluating and comparing the performance of AI models.

Bias

A systematic tendency in an AI model or dataset that can cause predictions or outputs to consistently deviate from the desired or expected result.

Bias-Variance Tradeoff

A fundamental machine learning concept describing the balance between a model being too simple and systematically inaccurate (high bias) and being too sensitive to its training data (high variance).

Binary Classification

A machine learning task in which a model assigns an input to one of two possible categories, such as positive or negative.

BLEU — Bilingual Evaluation Understudy

An evaluation metric originally designed for machine translation that compares generated text with one or more reference texts based primarily on matching word sequences.

Bootstrap Sampling

A statistical technique that repeatedly samples observations from an existing dataset, with replacement, to estimate uncertainty or assess the stability of a model or metric.

Bot

A software system that performs automated tasks, often repeatedly and without direct human intervention. AI-powered bots may use language models to understand and generate natural language.

Bounding Box

A rectangular region used in computer vision to identify the location of an object within an image.

Business Rules

Explicit rules or constraints that define how a system should behave in specific business situations. In AI systems, business rules can be combined with AI outputs to enforce predictable behavior.

Calibration

The process of adjusting a machine learning model so that its predicted probabilities accurately reflect the actual likelihood of an outcome.

Candidate Generation

The process of generating a set of potentially relevant items, documents, products, or results before a model ranks or filters them.

Capability

A specific task or function that an AI system is able to perform, such as summarization, classification, image generation, reasoning, or information retrieval.

Chain-of-Thought (CoT)

A prompting or reasoning approach in which a model is encouraged to work through a problem step by step before producing an answer.

Chunk

A smaller piece of a larger document or dataset created so that it can be processed, retrieved, indexed, or embedded more efficiently. Chunks are commonly used in RAG systems.

Chunking

The process of splitting a document or other large input into smaller pieces, usually called chunks, before indexing or processing it.

Classification

A machine learning task in which a model assigns an input to one or more predefined categories or classes.

Classifier

A machine learning model designed to assign data to one or more predefined categories.

Cloud AI

AI capabilities delivered through cloud infrastructure, allowing organizations to use models, computing resources, storage, or AI APIs without operating all of the underlying infrastructure themselves.

Clustering

An unsupervised machine learning technique that groups data points according to similarities without requiring predefined labels.

Context

The information provided to an AI model alongside a request to help it generate a relevant and accurate response. In an LLM system, context can include the user’s query, conversation history, retrieved documents, system instructions, and other relevant information.

Context Length

The maximum amount of information that an AI model can process as input within a single interaction, usually measured in tokens.

Context Window

The total amount of input and output information that an AI model can handle within a single interaction, subject to the model’s token limit.

Continuous Learning

A machine learning approach in which a model is continuously or periodically updated using new data as it becomes available.

Computer Vision

A field of AI focused on enabling computers to interpret and understand visual information such as images, videos, and other visual data.

Concept Drift

A change over time in the relationship between input data and the outcome a machine learning model is expected to predict, which can cause model performance to decline.

Confidence Score

A numerical value representing how confident a model is in a prediction, classification, or other output. The meaning and reliability of the score depend on the model and its calibration.

Content Generation

The use of AI models to automatically create content such as text, images, audio, video, or code.

Content Moderation

The process of identifying, filtering, classifying, or managing content according to predefined safety, legal, ethical, or platform-specific rules.

Conversational AI

AI technology designed to interact with users through natural language, typically using text or voice.

Cosine Similarity

A mathematical measure used to determine how similar two vectors are by comparing the angle between them. It is commonly used in semantic search and vector databases.

Cost per Token

The price charged by an AI model provider for processing a certain number of input or output tokens.

CPU — Central Processing Unit

A general-purpose processor used to execute instructions and perform computations. CPUs can run AI workloads but are generally less efficient than GPUs for large-scale parallel machine learning operations.

Cross-Validation

A model evaluation technique in which data is divided into multiple subsets so that a model can be trained and evaluated several times on different portions of the dataset.

CUDA

A parallel computing platform and programming model developed by NVIDIA that allows software to use NVIDIA GPUs for general-purpose computation, including AI and machine learning workloads.

Custom Model

An AI model specifically adapted, trained, fine-tuned, or configured for the needs of a particular organization, use case, or domain.

Customer Data

Information provided by or generated from customers and used by an AI system, such as documents, messages, transactions, behavioral data, or user queries.

Customer Support AI

An AI system designed to assist customer service operations by answering questions, retrieving information, classifying requests, summarizing interactions, or assisting human support agents.

Cybersecurity AI

The use of AI techniques to detect, prevent, investigate, or respond to cybersecurity threats and incidents.

Data Augmentation

The process of creating additional training data by modifying or transforming existing data, such as rotating images, changing text, or adding synthetic examples.

Data Drift

A change over time in the statistical characteristics or distribution of the data used by an AI system, which can negatively affect model performance.

Data Governance

The policies, processes, roles, and controls used to ensure that data is properly managed, secure, accurate, accessible, compliant, and used appropriately.

Data Labeling

The process of assigning labels or categories to data so that it can be used to train or evaluate a machine learning model.

Data Leakage

A situation where information that should not be available to a model during training or evaluation is accidentally included, potentially producing artificially high performance.

Data Pipeline

A sequence of automated processes used to collect, transform, validate, move, and prepare data for storage, analysis, or use by an AI system.

Data Privacy

The protection and appropriate handling of personal, confidential, or sensitive data throughout its collection, processing, storage, and transmission.

Data Quality

The degree to which data is accurate, complete, consistent, relevant, reliable, timely, and suitable for its intended purpose.

Data Retrieval

The process of finding and returning relevant information from a dataset, database, document collection, or knowledge base.

Data Scientist

A professional who uses statistics, machine learning, programming, and data analysis to extract insights from data and develop data-driven models or solutions.

Data Source

An origin from which data is collected or retrieved, such as a database, API, document repository, website, sensor, application, or external dataset.

Data Validation

The process of checking data against predefined rules or requirements to ensure that it is accurate, complete, consistent, and suitable for processing.

Database

A structured system used to store, organize, manage, and retrieve data electronically.

Deep Learning

A branch of machine learning that uses neural networks with multiple layers to learn complex patterns from large amounts of data.

Deep Neural Network (DNN)

A neural network containing multiple layers between its input and output layers, allowing it to learn increasingly complex representations of data.

Deployment

The process of making an AI model or AI-powered application available for use in a production environment.

Deployment Pipeline

An automated or semi-automated sequence of steps used to test, validate, package, and deploy software or AI models into different environments.

Deterministic System

A system that produces the same output when given the same input and conditions. This contrasts with many generative AI systems, which can produce different outputs for the same prompt.

Diffusion Model

A type of generative AI model that learns to create data, particularly images, by gradually transforming random noise into a meaningful output.

Dimensionality Reduction

The process of reducing the number of variables or dimensions in a dataset while attempting to preserve the most important information.

Distillation

A model compression technique in which a smaller model, called the student, is trained to reproduce the behavior or knowledge of a larger model, called the teacher.

Distributed Computing

A computing approach in which a workload is divided across multiple machines or computing resources that work together.

Document Corpus

A collection of documents used as a source of information for tasks such as search, retrieval, training, or RAG-based question answering.

Document Retrieval

The process of identifying and retrieving documents or document sections that are relevant to a user’s query.

Domain Adaptation

The process of adapting a machine learning model so that it performs effectively on data from a specific domain that may differ from the data used during its original training.

Domain-Specific Model

An AI model designed, adapted, or optimized to perform tasks within a particular industry, subject area, or specialized domain.

Drift Detection

The process of monitoring data or model behavior to identify significant changes that could indicate data drift, concept drift, or declining model performance.

Docker

A platform used to package applications and their dependencies into isolated containers, making them easier to develop, test, deploy, and run consistently across environments.

Dynamic Prompting

An approach in which prompts are automatically constructed or modified based on factors such as user input, retrieved information, conversation history, or application state.

Early Stopping

A training technique that stops model training when performance on a validation dataset stops improving, helping prevent overfitting.

Edge AI

The use of AI models directly on devices or at the edge of a network, such as smartphones, cameras, vehicles, or IoT devices, rather than sending all data to a centralized cloud system.

Embedding

A numerical representation of data, such as text, images, or audio, in the form of a vector that captures meaningful relationships and similarities between data points.

Embedding Model

An AI model that converts data such as text or images into numerical vectors called embeddings, which can then be used for tasks such as semantic search, similarity matching, and retrieval.

Embedding Space

A mathematical space in which data is represented as vectors, where the relative positions of vectors reflect similarities or relationships between the underlying data.

End-to-End AI System

An AI solution that covers the complete workflow from input data and processing through model inference to the final output or business action.

Entity Extraction

The process of identifying and extracting specific entities from unstructured data, such as people, organizations, locations, dates, products, or monetary values.

Entity Recognition

The process of automatically identifying specific types of entities within text, commonly known as Named Entity Recognition (NER).

Environment

A specific infrastructure or configuration in which an application or AI system runs, such as development, testing, staging, or production.

Epoch

One complete pass through the entire training dataset during the training of a machine learning model.

Error Analysis

The systematic examination of a model’s incorrect or undesirable outputs in order to identify patterns, diagnose problems, and determine how performance can be improved.

Evaluation Dataset

A dataset used to measure the performance of an AI model or system against predefined criteria.

Evaluation Metric

A quantitative measure used to assess the performance, quality, accuracy, or reliability of an AI model or system.

Evaluation Pipeline

A structured process used to automatically or systematically test an AI system against predefined datasets, scenarios, and evaluation metrics.

Explainability

The degree to which humans can understand how an AI system arrived at a particular prediction, decision, or output.

Explainable AI (XAI)

Methods and techniques designed to make AI system behavior and outputs more understandable and interpretable to humans.

Exploratory Data Analysis (EDA)

The process of analyzing and visualizing a dataset to understand its structure, patterns, relationships, anomalies, and potential quality issues before building a model.

External Knowledge Base

A repository of information outside the AI model itself that can be accessed by an AI system to provide additional knowledge or context, for example through retrieval or RAG.

Extraction

The process of identifying and retrieving specific information or structured data from unstructured or semi-structured sources such as documents, websites, images, or emails.

ETL — Extract, Transform, Load

A data integration process in which data is extracted from one or more sources, transformed into a suitable format, and loaded into a target system such as a database or data warehouse.

Event-Driven Architecture

A software architecture in which components communicate and trigger actions in response to events, such as a new document being uploaded, a payment being completed, or a user submitting a request.

Experiment Tracking

The process of recording experiments, datasets, model versions, parameters, metrics, and results so that machine learning experiments can be reproduced and compared.

Expert-in-the-Loop

An AI workflow in which a human expert reviews, validates, corrects, or approves AI-generated outputs as part of the system’s process.

Expert System

A computer system designed to reproduce decision-making in a specific domain using explicitly defined rules, knowledge, and inference mechanisms.

Exposure Bias

A problem in sequence generation where a model is trained using the correct previous outputs but must rely on its own potentially incorrect outputs during inference.

F1 Score

A classification evaluation metric that combines precision and recall into a single score using their harmonic mean. It is particularly useful when both false positives and false negatives matter.

False Negative (FN)

A case where an AI system fails to identify something that is actually present or positive. For example, a fraud detection system classifies a fraudulent transaction as legitimate.

False Positive (FP)

A case where an AI system incorrectly identifies something as present or positive when it is actually absent or negative. For example, a fraud detection system flags a legitimate transaction as fraudulent.

Feature

An individual measurable property, characteristic, or input variable used by a machine learning model to make a prediction or classification.

Feature Engineering

The process of selecting, transforming, or creating input features to improve the performance of a machine learning model.

Feature Extraction

The process of transforming raw data into a smaller or more useful set of representations that can be used by a machine learning model.

Feature Store

A centralized system used to manage, store, and serve machine learning features consistently for training and production inference.

Federated Learning

A machine learning approach in which models are trained across multiple devices or organizations without requiring the underlying raw data to be centralized.

Fine-Tuning

The process of further training a pre-trained AI model on a smaller, task-specific or domain-specific dataset to adapt its behavior or performance to a particular use case.

Foundation Model

A large, general-purpose AI model trained on broad datasets that can be adapted to perform many different tasks, often through prompting, fine-tuning, or other techniques.

Few-Shot Learning

An approach in which an AI model performs a task after being given only a small number of examples demonstrating the desired behavior.

Few-Shot Prompting

A prompting technique in which a small number of examples are included in the prompt to guide the model toward a desired output format or behavior.

Feedback Loop

A process in which outputs, user interactions, corrections, or performance results are collected and used to improve an AI system or its future behavior.

Feedback Data

Data collected from users, experts, systems, or evaluation processes that indicates whether an AI output was useful, correct, relevant, or acceptable.

Fine-Grained Control

The ability to precisely control specific aspects of an AI system’s behavior, permissions, outputs, or configuration.

Fine-Grained Authorization

An access-control approach that allows permissions to be defined at a detailed level, such as by user, role, document, resource, action, or tenant.

Flan-T5

A family of instruction-tuned language models developed by Google Research based on the T5 architecture and trained to follow natural-language instructions across many tasks.

Foundation Model Provider

An organization that develops and provides access to large general-purpose AI models through APIs, cloud platforms, downloadable models, or other services.

Function Calling

A capability that allows an AI model to identify when an external function or tool should be used and generate structured arguments for that function.

Function Calling API

An API mechanism that allows an application to expose predefined functions or tools to an AI model so that the model can request their execution.

Functional Requirements

Requirements describing what an AI system must be able to do, such as answering questions, retrieving documents, classifying requests, generating reports, or calling external tools.

Factuality

The degree to which an AI-generated output is factually accurate and supported by reliable information or evidence.

Factual Consistency

The degree to which generated content remains consistent with the source information or reference material it is expected to represent.

Framework

A software framework that provides reusable components, abstractions, and tools for building AI or machine learning applications.

Front-End

The part of an AI application that users directly interact with, such as a web interface, chatbot interface, dashboard, or mobile application.

GAN — Generative Adversarial Network

A type of generative AI architecture composed of two neural networks, a generator and a discriminator, that compete with each other to produce increasingly realistic synthetic data.

Generative AI

A category of AI systems capable of generating new content such as text, images, audio, video, or code based on learned patterns and user inputs.

Generative Model

An AI model designed to learn the underlying patterns of data and generate new data that resembles the data it was trained on.

Generator

In a generative adversarial network, the generator is the model responsible for creating synthetic data intended to resemble real data.

Generative Pre-trained Transformer (GPT)

A family of language models based on the Transformer architecture that are pre-trained on large amounts of data and can generate and process natural language.

Generalization

The ability of an AI model to perform well on new, unseen data rather than only on the examples used during training.

General-Purpose AI

An AI system designed to perform a broad range of tasks rather than being optimized for only one narrowly defined use case.

GPU — Graphics Processing Unit

A processor designed to perform large numbers of parallel computations efficiently. GPUs are widely used for training and running modern AI models.

GPU Cluster

A group of interconnected GPUs or GPU-equipped machines working together to provide the computational resources required for large-scale AI workloads.

Gradient

A mathematical quantity that indicates how a model’s parameters should change to reduce its error during training.

Gradient Descent

An optimization algorithm used to train machine learning models by iteratively adjusting model parameters in the direction that reduces the model’s error.

Gradient Explosion

A training problem in which gradients become extremely large, causing unstable updates and potentially preventing a neural network from training successfully.

Gradient Vanishing

A training problem in which gradients become extremely small as they are propagated through a neural network, making it difficult for earlier layers to learn effectively.

Ground Truth

The trusted or reference answer, label, value, or outcome used as the basis for evaluating an AI model’s predictions or outputs.

Ground Truth Dataset

A dataset containing validated reference examples used to train, test, or evaluate an AI system.

Guardrails

Rules, controls, filters, or technical mechanisms designed to constrain an AI system’s behavior and reduce unwanted, unsafe, or non-compliant outputs.

Guardrail Model

A model or AI component used to detect, filter, classify, or block inputs or outputs that violate predefined safety, security, quality, or policy requirements.

Graph Neural Network (GNN)

A type of neural network designed to operate on graph-structured data, where entities are represented as nodes and relationships between them are represented as edges.

Graph Database

A database designed to store and query entities and the relationships between them, often represented as nodes and edges.

GraphRAG

A RAG approach that uses graph structures and relationships between entities to retrieve and organize information before providing context to a language model.

Grounded Generation

A generation approach in which an AI model produces responses based on specific, retrieved, or otherwise verifiable information rather than relying only on its learned parameters.

Groundedness

The degree to which an AI-generated response is supported by the information or evidence provided to the model.

Generative AI Evaluation

The process of measuring the quality, accuracy, relevance, safety, factuality, and other characteristics of outputs produced by generative AI systems.

Governance

The framework of policies, responsibilities, processes, controls, and decision-making mechanisms used to manage an AI system responsibly throughout its lifecycle.

Governance Framework

A structured set of policies, roles, procedures, controls, and standards used to ensure that AI systems are developed and operated according to organizational, legal, security, and ethical requirements.

Hallucination

A situation in which an AI model generates information that is incorrect, fabricated, or unsupported by reliable evidence while presenting it as if it were factual.

Hallucination Rate

A metric or measurement indicating how frequently an AI system produces unsupported, incorrect, or fabricated information under defined evaluation conditions.

Hard Negative

An example that is highly similar to the correct answer or target but is nevertheless incorrect. Hard negatives are often used to improve retrieval and classification systems.

Hard Prompt

A prompt that contains explicit instructions, constraints, formatting requirements, examples, or rules designed to strongly control an AI model’s behavior.

Hardware Acceleration

The use of specialized hardware, such as GPUs, TPUs, or AI accelerators, to perform AI computations more efficiently than on general-purpose processors alone.

Hidden Layer

A layer of neurons located between the input and output layers of a neural network. Hidden layers transform information and enable the network to learn complex patterns.

Hidden State

An internal numerical representation maintained by a neural network while processing information. It captures information learned or inferred from the input and intermediate computations.

Hierarchical Classification

A classification approach in which categories are organized into multiple levels, such as a broad category followed by increasingly specific subcategories.

Hierarchical Retrieval

A retrieval strategy that searches information at multiple levels of organization, such as document, section, paragraph, and sentence, to improve retrieval efficiency or relevance.

Human-in-the-Loop (HITL)

An AI system design in which humans participate in the workflow by reviewing, correcting, validating, approving, or otherwise influencing AI outputs or decisions.

Human-on-the-Loop

An AI system design in which humans supervise an automated system and can intervene when necessary, rather than manually reviewing every individual output.

Human-in-the-Loop Evaluation

An evaluation approach in which human experts assess AI outputs according to predefined criteria such as accuracy, relevance, safety, or usefulness.

Human Feedback

Information provided by humans about the quality, correctness, relevance, or acceptability of AI outputs, which can be used for evaluation, improvement, or training.

Human Feedback Dataset

A dataset containing human evaluations, preferences, corrections, rankings, or annotations of AI-generated outputs.

Hyperparameter

A configuration value set before or during model training that controls aspects of the learning process, such as learning rate, batch size, or number of training iterations.

Hyperparameter Optimization

The process of systematically searching for hyperparameter values that produce the best model performance according to a chosen evaluation metric.

Hybrid AI

An AI approach that combines different techniques, such as machine learning, generative AI, symbolic rules, knowledge bases, or traditional software logic, within the same system.

Hybrid Search

A retrieval approach that combines multiple search methods, commonly keyword-based search and semantic vector search, to improve the relevance of retrieved information.

Hybrid RAG

A RAG architecture that combines different retrieval methods, such as vector search, keyword search, metadata filtering, or structured database queries, before providing information to a language model.

Hypothesis Testing

A statistical method used to determine whether observed differences or relationships in data provide sufficient evidence to support a specific hypothesis.

Inference

The process of using a trained AI model to generate a prediction, classification, recommendation, or response from new input data.

Inference Engine

A software component responsible for running a trained AI model on new inputs and producing predictions or outputs.

Inference Pipeline

The sequence of processes through which input data is received, prepared, processed by an AI model, and transformed into a final output.

Inference Latency

The amount of time required for an AI system to process an input and return an output.

Inference Cost

The cost associated with running an AI model to process inputs and generate outputs, including compute resources and, for API-based models, provider usage fees.

Inference-Time

Referring to the stage when a trained AI model is being used to process new data and generate outputs, as opposed to the training stage.

Input Token

A token contained in the information sent to a language model, including the user’s prompt, system instructions, conversation history, or retrieved context.

Instruction Tuning

The process of further training a language model on examples designed to teach it how to follow natural-language instructions more effectively.

Instruction-Tuned Model

A model that has been specifically trained or fine-tuned to follow natural-language instructions and perform tasks described by users.

Interpretability

The degree to which humans can understand how an AI model processes information and produces its predictions or outputs.

Interoperability

The ability of different AI systems, applications, platforms, or data sources to communicate and work together effectively using compatible interfaces, formats, or standards.

IoT — Internet of Things

A network of physical devices equipped with sensors, software, and connectivity that allows them to collect, exchange, and process data. AI can be used to analyze this data or control IoT systems.

Image Classification

A computer vision task in which an AI model assigns an image to one or more predefined categories.

Image Generation

The use of generative AI models to create new images based on text prompts, reference images, or other inputs.

Image Recognition

The process of using AI to identify objects, patterns, people, scenes, or other information contained within an image.

Image Segmentation

A computer vision task in which an AI model assigns a label to individual pixels or regions of an image to identify specific objects or areas.

Imbalanced Dataset

A dataset in which some classes or categories contain significantly more examples than others, potentially causing a machine learning model to perform poorly on minority classes.

Imitation Learning

A machine learning approach in which a model learns how to perform a task by observing demonstrations or examples of desired behavior.

Incremental Learning

A learning approach in which a model is updated progressively as new data becomes available instead of being retrained entirely from scratch.

Information Retrieval (IR)

The process of finding and retrieving relevant information from a large collection of documents, databases, websites, or other information sources in response to a query.

Information Retrieval System

A system designed to identify and return information relevant to a user’s query, often using keyword search, semantic search, vector search, or a combination of retrieval methods.

Information Extraction

The process of automatically extracting structured information from unstructured or semi-structured data such as documents, emails, websites, or text.

Input Validation

The process of checking and validating user or system inputs before they are processed by an AI application, helping prevent invalid, malicious, or unexpected inputs.

Isolation

The technical separation of data, processes, users, or resources so that one system, tenant, or user cannot improperly access another’s information or resources.

Isolation Forest

An unsupervised machine learning algorithm commonly used for anomaly detection by identifying observations that are easier to isolate from the rest of the data.

Jailbreak

An attempt to bypass an AI system’s safety rules, behavioral restrictions, or system instructions by using specially crafted prompts or inputs.

Jailbreak Attack

An attack in which a user deliberately attempts to make an AI model ignore or circumvent its predefined safety policies, restrictions, or instructions.

JavaScript

A widely used programming language for web applications that can also be used to build AI-powered applications, interact with AI APIs, and implement frontend or backend functionality.

JSON — JavaScript Object Notation

A lightweight structured data format commonly used for exchanging information between applications and APIs. AI systems frequently use JSON for structured inputs, outputs, and tool calls.

JSON Schema

A formal specification that defines the expected structure, data types, and constraints of JSON data. It can be used to ensure that AI-generated structured outputs follow a predefined format.

Jensen-Shannon Divergence

A statistical measure of similarity between two probability distributions. It is related to Kullback-Leibler divergence and can be used to compare distributions in machine learning and generative modeling.

Jupyter Notebook

An interactive computing environment widely used by data scientists and machine learning teams to write code, analyze data, visualize results, and experiment with AI models.

JAX

An open-source Python library developed by Google for high-performance numerical computing and machine learning, particularly useful for automatic differentiation and accelerated computation on GPUs and TPUs.

Job Queue

A system that stores tasks waiting to be processed, allowing computationally expensive or asynchronous AI operations to be executed in an organized manner.

Job Scheduling

The process of determining when and where computational tasks should be executed, often used to manage AI training jobs, batch processing, or distributed workloads.

Joint Embedding

An approach in which different types of data, such as text and images, are represented in a shared embedding space so that their semantic relationships can be compared.

Joint Training

A training approach in which multiple objectives, tasks, or model components are trained together rather than independently.

Just-In-Time (JIT) Compilation

A compilation technique in which code is compiled into machine-level instructions during execution rather than entirely before execution. It can improve performance in certain AI and numerical computing workloads.

K-Means Clustering

An unsupervised machine learning algorithm that groups data points into a predefined number of clusters based on their similarity, typically by minimizing the distance between data points and their assigned cluster center.

Knowledge Base

A structured or unstructured repository of information that an AI system can use to retrieve facts, documents, or domain-specific knowledge.

Knowledge Graph

A structured representation of entities and the relationships between them, typically represented as nodes and edges, that allows an AI system to reason over connected information.

Knowledge Retrieval

The process of retrieving relevant information from a knowledge base, database, document collection, or other external source to support an AI system’s response.

Knowledge Representation

The process of representing information about entities, concepts, relationships, and rules in a structured form that can be processed or reasoned about by an AI system.

Knowledge Distillation

A model compression technique in which a smaller student model learns to reproduce the behavior or outputs of a larger teacher model.

Knowledge Injection

The process of providing an AI model with additional domain-specific information so that it can use that information when generating outputs or making decisions.

Knowledge Grounding

The process of connecting an AI model’s responses to specific external information or evidence so that its outputs are based on identifiable sources rather than only on its learned parameters.

Knowledge Graph RAG

A retrieval-augmented generation approach that uses a knowledge graph to retrieve relevant entities, relationships, and connected information before providing context to a language model.

Knowledge Cutoff

The point in time up to which the information contained in a pre-trained AI model’s training data is generally available to the model’s learned parameters. Information after this point may require external retrieval or updated training.

Key-Value Cache (KV Cache)

A memory mechanism used during language model inference to store previously computed key and value representations, reducing redundant computation when generating tokens sequentially.

Key Performance Indicator (KPI)

A measurable indicator used to evaluate whether an AI project, product, or system is achieving its defined business or operational objectives.

Kullback-Leibler Divergence (KL Divergence)

A mathematical measure of how one probability distribution differs from another. It is commonly used in machine learning to compare probability distributions and in some model training objectives.

Label

A category, value, or annotation assigned to a data sample and used as the target that a machine learning model is trained or evaluated to predict.

Labeling

The process of assigning labels or annotations to data so that it can be used to train, validate, or evaluate a machine learning model.

Large Language Model (LLM)

An AI model trained on large amounts of text data and designed to understand and generate natural language. LLMs can perform tasks such as question answering, summarization, translation, classification, and content generation.

Latent Space

A mathematical representation space in which an AI model encodes complex data into a more compact form that captures underlying patterns or relationships.

Latency

The amount of time between sending an input to an AI system and receiving its output.

Layer

A computational stage within a neural network that transforms the information received from the previous layer before passing it to the next one.

Learning Rate

A hyperparameter that controls how much a machine learning model’s parameters are adjusted during each training update.

Learning-to-Rank

A machine learning approach in which a model is trained to order or rank items according to their relevance to a query or objective.

Linear Regression

A supervised machine learning algorithm used to model the relationship between one or more input variables and a continuous numerical output.

Log Loss

An evaluation metric that measures the quality of probabilistic predictions by penalizing incorrect predictions, particularly when the model is highly confident in the wrong answer.

Logit

A numerical value produced by some machine learning models before it is transformed into a probability or final prediction.

Long Context

The ability of a language model to process a large amount of information within a single context window.

Long Context Window

A large token capacity that allows a language model to process extensive inputs such as long documents, large codebases, or long conversation histories in a single interaction.

Loss Function

A mathematical function that measures how far a model’s predictions are from the desired outputs. During training, optimization algorithms attempt to minimize this loss.

Low-Code AI

An approach to building AI-powered applications using visual interfaces, pre-built components, and limited amounts of traditional programming.

Low-Rank Adaptation (LoRA)

A parameter-efficient fine-tuning technique that adapts a large AI model by training a relatively small number of additional parameters rather than updating the entire model.

Language Model

An AI model trained to understand, predict, or generate sequences of language, such as words or tokens.

Language Model Evaluation

The process of assessing a language model’s performance using criteria such as accuracy, relevance, factuality, safety, instruction following, and task-specific quality.

Lexical Search

A search method that retrieves information based primarily on exact or related words and textual matches rather than semantic similarity.

LLM-as-a-Judge

An evaluation approach in which one language model evaluates the quality of outputs generated by another AI system according to predefined criteria or a rubric.

LLM Gateway

A software layer that provides a centralized interface for accessing and managing multiple language models or AI providers, often handling routing, authentication, logging, monitoring, security, and cost control.

LLM Orchestration

The process of coordinating language models, prompts, tools, retrieval systems, APIs, databases, and workflows to create a complete AI application.

LLM Routing

The process of selecting which language model should handle a request based on factors such as cost, latency, capability, context length, or task requirements.

LLM Serving

The infrastructure and software processes required to make a trained language model available for inference and handle requests from applications or users.

Machine Learning (ML)

A branch of AI in which systems learn patterns from data to make predictions, classifications, decisions, or generate outputs without being explicitly programmed for every individual case.

Machine Learning Model

A computational model trained on data to identify patterns and perform tasks such as prediction, classification, recommendation, or generation.

Machine Learning Pipeline

A sequence of processes used to collect, prepare, train, evaluate, deploy, and monitor a machine learning model.

Machine Learning Operations (MLOps)

A set of practices, tools, and processes used to develop, deploy, monitor, maintain, and manage machine learning models reliably in production.

Machine Translation

The use of AI to automatically translate text or speech from one language into another.

Metadata

Data that describes other data, such as a document’s title, author, creation date, source, language, category, or permissions.

Metadata Filtering

A retrieval technique that filters information based on metadata attributes before or during the search process, such as document type, date, department, user, or access permissions.

Model

A mathematical or computational representation learned from data or designed according to specific rules to perform a task such as prediction, classification, generation, or decision-making.

Model Architecture

The underlying structure and design of an AI model, including its layers, components, connections, and computational mechanisms.

Model Compression

A set of techniques used to reduce the size, computational requirements, or memory usage of an AI model while attempting to preserve its useful performance.

Model Drift

A decline in model performance caused by changes in the data, environment, user behavior, or relationships that the model was originally trained to represent.

Model Evaluation

The process of measuring how well an AI model performs according to predefined metrics, datasets, scenarios, or business requirements.

Model Fine-Tuning

The process of further training a pre-trained model on a specific dataset to adapt its behavior or performance to a particular task or domain.

Model Monitoring

The continuous observation of a deployed AI model to detect changes in performance, data quality, latency, cost, errors, or other operational indicators.

Model Serving

The infrastructure and software required to make a trained AI model available for inference and process requests from users or applications.

Model Validation

The process of evaluating a model on data that was not used to train it in order to estimate how well it is likely to perform on unseen data.

Multimodal AI

AI that can process or generate multiple types of data, such as text, images, audio, video, or other modalities, within the same system.

Multimodal Model

An AI model capable of processing, understanding, or generating information across multiple modalities, such as text and images.

Multi-Agent System

An AI system composed of multiple agents that can collaborate, communicate, delegate tasks, or perform different roles within a larger workflow.

Multi-Tenant Architecture

An architecture in which a single application or infrastructure serves multiple customers or organizations while keeping their data, permissions, and resources appropriately isolated.

Multi-Tenant RAG

A RAG architecture serving multiple customers or organizations where retrieval and access controls must ensure that one tenant cannot retrieve or receive another tenant’s confidential information.

Model Context Protocol (MCP)

An open protocol designed to standardize how AI applications and language models connect to external tools, data sources, and services.

Mixture of Experts (MoE)

A model architecture in which different specialized subnetworks, called experts, are selectively activated for each input rather than using the entire model for every computation.

Multilingual Model

An AI model trained to understand or generate content in multiple languages.

Mean Absolute Error (MAE)

An evaluation metric that measures the average absolute difference between predicted values and actual values.

Mean Squared Error (MSE)

An evaluation metric that measures the average squared difference between predicted values and actual values, placing greater emphasis on larger errors.

Monitoring

The continuous observation of an AI system’s performance, behavior, infrastructure, usage, cost, and reliability after deployment.

Moderation Model

An AI model designed to identify, classify, filter, or flag content that violates predefined safety, legal, quality, or platform policies.

Natural Language Processing (NLP)

A field of AI focused on enabling computers to understand, process, analyze, and generate human language.

Natural Language Understanding (NLU)

The ability of an AI system to interpret the meaning, intent, entities, and context of human language.

Natural Language Generation (NLG)

The process of generating human-readable language from structured data, instructions, retrieved information, or other inputs.

Natural Language Interface

An interface that allows users to interact with a software system using natural human language rather than predefined commands or traditional graphical controls.

Named Entity Recognition (NER)

An NLP task that identifies and classifies entities in text, such as people, organizations, locations, dates, products, or other predefined categories.

Neural Network

A machine learning model composed of interconnected computational units, typically organized into layers, that learns patterns from data.

Neural Language Model

A neural network trained to model patterns and probabilities in language, such as predicting the next token or generating text.

Neural Machine Translation (NMT)

A machine translation approach that uses neural networks to translate text from one language to another.

Neural Retrieval

An information retrieval approach that uses neural models or embeddings to identify semantically relevant information rather than relying only on exact keyword matches.

Negative Example

A training or evaluation example representing an incorrect, irrelevant, or non-target outcome.

Negative Sampling

A technique in which negative or incorrect examples are selected during training to help a model distinguish relevant examples from irrelevant ones.

Nearest Neighbor Search

A search method that identifies the data points most similar to a given query according to a defined distance or similarity measure.

Nearest Neighbor

A data point that is considered most similar to another data point according to a chosen distance or similarity metric.

Neural Embedding

A numerical vector representation generated by a neural model to capture semantic, contextual, or structural information about data.

Node

An individual element in a graph structure, representing an entity, concept, object, or other unit of information.

Node Classification

A graph machine learning task in which a model assigns categories or labels to individual nodes in a graph.

Noise

Irrelevant, random, erroneous, or unwanted information in data that can reduce the quality or reliability of an AI system.

Noise Reduction

The process of identifying and reducing irrelevant, erroneous, or unwanted information in data or model inputs.

Non-Deterministic Output

An AI output that can vary between executions even when the input and instructions remain substantially the same.

Non-Parametric Model

A machine learning model that does not assume a fixed functional form or a fixed number of parameters and can adapt its complexity to the data.

Normalization

The process of transforming data into a standardized numerical or structural form to improve consistency and, in some cases, model performance.

Null Handling

The process of detecting and appropriately managing missing or undefined values in datasets or application inputs.

Numerical Stability

The ability of a computational method or model to produce reliable results without being excessively affected by small numerical errors or rounding issues.

Neural Architecture Search (NAS)

An automated process that searches for effective neural network architectures according to predefined objectives such as accuracy, latency, or computational cost.

Next-Token Prediction

A language modeling task in which a model predicts the most probable next token based on the preceding context.

Nucleus Sampling

A text generation technique that restricts token selection to the smallest set of tokens whose cumulative probability exceeds a chosen threshold.

OCR — Optical Character Recognition

A technology that converts text contained in images, scanned documents, or PDFs into machine-readable text.

Object Detection

A computer vision task that identifies objects in an image or video and determines their locations, typically using bounding boxes.

Object Recognition

The process of identifying and classifying objects or entities within images, video, or other data.

Objective Function

A mathematical function that a machine learning algorithm seeks to optimize during training, such as minimizing error or maximizing a desired performance measure.

Observability

The ability to understand the internal state and behavior of an AI system by collecting and analyzing logs, metrics, traces, outputs, and other operational signals.

Online Learning

A machine learning approach in which a model learns incrementally from new data as it becomes available rather than being retrained only on large batches of historical data.

One-Shot Learning

A learning approach in which a model is expected to perform a task after seeing only one example of the desired behavior or category.

One-Hot Encoding

A method for representing categorical values as binary vectors in which one position represents the selected category.

Open-Source Model

An AI model whose model files, code, weights, or other components are made available under specific open-source or open-weight licensing terms.

Open-Weight Model

An AI model whose trained weights are publicly available, allowing users to run or adapt the model subject to its license, without necessarily providing all training data or training code.

Out-of-Distribution (OOD)

A situation in which input data differs significantly from the data distribution encountered during model training.

Out-of-Distribution Detection

The process of identifying inputs that are sufficiently different from the data a model was trained on and may therefore produce unreliable results.

Out-of-Domain Query

A user request that falls outside the domain, knowledge scope, or intended capabilities of an AI system.

Output Validation

The process of checking an AI-generated output against predefined rules, schemas, constraints, or quality requirements before it is used downstream.

Output Guardrail

A mechanism that evaluates, filters, transforms, or blocks AI-generated outputs when they violate predefined safety, quality, policy, or business requirements.

Output Token

A token generated by an AI model as part of its response.

Overfitting

A modeling problem in which a model learns the training data too closely, including its noise or specific patterns, and performs poorly on unseen data.

Optimization

The process of adjusting a model, system, or set of parameters to improve a defined objective such as accuracy, cost, latency, or another performance measure.

Optimizer

An algorithm used during model training to update model parameters in order to minimize a loss function or optimize a defined objective.

Orchestration

The coordination of multiple AI models, tools, data sources, services, or processing steps to execute a larger workflow.

Agent Orchestration

The coordination of one or more AI agents, tools, memory components, and workflows to accomplish a task or objective.

Ontology

A formal representation of concepts, entities, relationships, and categories within a specific domain.

Ontology-Based Retrieval

A retrieval approach that uses a domain ontology to identify and connect relevant concepts, entities, and relationships when retrieving information.

OpenAI API

An application programming interface that allows software applications to interact programmatically with OpenAI models and related capabilities.

Offline Evaluation

The evaluation of an AI system using a predefined dataset or test suite before or independently of real-world production usage.

Online Evaluation

The evaluation of an AI system using real-world or production interactions, often through monitoring, user feedback, experiments, or controlled tests.

Parameter

A numerical value learned or used by a model to determine how it processes inputs and produces outputs.

Parameter-Efficient Fine-Tuning (PEFT)

A family of techniques that adapts a pre-trained model by updating only a small portion of its parameters instead of retraining the entire model.

Parametric Model

A model that uses a fixed number of parameters to represent the relationship between inputs and outputs.

Perplexity

An evaluation metric commonly used for language models that measures how well a model predicts a sequence of tokens, with lower values generally indicating better predictive performance on the evaluated data.

Precision

A classification or retrieval metric representing the proportion of relevant or correct positive results among all results identified as positive.

Precision@K

A retrieval metric measuring the proportion of relevant results among the top K retrieved results.

Prediction

The output produced by a machine learning model when it processes an input.

Predictive Model

A machine learning model designed to predict an outcome, value, class, or probability from input data.

Pre-trained Model

An AI model that has already been trained on a large or general dataset and can subsequently be used directly or adapted to specific tasks.

Pretraining

The initial training phase in which a model learns general patterns, representations, or capabilities from a large dataset before being adapted to specific tasks.

Prompt

The input instructions, context, data, examples, or questions provided to an AI model to guide its behavior or output.

Prompt Chaining

A technique in which the output of one prompt or model call is passed as input to another prompt or model call as part of a larger workflow.

Prompt Engineering

The practice of designing and optimizing prompts to improve the quality, consistency, relevance, or behavior of AI model outputs.

Prompt Injection

An attack in which malicious or unintended instructions are introduced into an AI system’s inputs or retrieved content in order to influence the model’s behavior.

Prompt Template

A reusable prompt structure containing fixed instructions and variable elements that can be dynamically populated at runtime.

Prompt Versioning

The practice of maintaining and tracking different versions of prompts to compare, reproduce, and manage changes in AI system behavior.

Prompt Evaluation

The systematic testing of prompts to determine their effectiveness according to predefined quality, accuracy, safety, or business criteria.

Prompt Library

A centralized collection of reusable, documented, and versioned prompts used across an AI application or organization.

Prompt Optimization

The process of systematically modifying prompts to improve model performance, reliability, cost, latency, or another defined objective.

Prompt Injection Defense

A set of techniques used to reduce the risk that untrusted instructions will manipulate an AI model or override the intended behavior of an application.

Pipeline

A sequence of automated processing steps through which data, requests, or model outputs pass to complete a defined workflow.

Post-Processing

The processing applied to model outputs after inference, such as formatting, validation, filtering, transformation, or business-rule enforcement.

Post-Training

Training performed after a model’s initial pretraining to improve its usefulness, instruction-following, safety, domain adaptation, or task-specific performance.

Probabilistic Model

A model that represents uncertainty using probability distributions and produces outputs based on probabilistic relationships.

Probability Distribution

A mathematical representation describing the likelihood of possible values or outcomes.

Production Model

An AI model that is deployed and actively used by real users or applications in a production environment.

Production Monitoring

The continuous monitoring of an AI system after deployment to detect performance degradation, errors, data drift, safety issues, latency, or cost changes.

Parameter Count

The number of learned parameters contained in an AI model, often used as an approximate indicator of model scale and computational requirements.

Pooling

A neural network operation that reduces the spatial or dimensional representation of data while retaining important information.

Pipeline Orchestration

The coordination and execution of multiple processing steps, services, models, or tools within an automated AI workflow.

Policy Model

A model that determines which action an AI system or agent should take given a particular state, input, or context.

Policy Enforcement

The implementation of technical controls that ensure an AI system follows predefined security, safety, compliance, or business policies.

Preference Learning

A learning approach in which a model learns desired behavior from preferences or comparisons between different outputs or actions.

Preference Dataset

A dataset containing human or system preferences between alternative outputs, responses, or actions, often used to train or evaluate AI models.

Proprietary Model

An AI model developed and controlled by an organization whose model parameters, training data, architecture, or implementation are not fully publicly available.

Private AI

An AI system designed to operate within controlled infrastructure or environments with specific requirements for data privacy, security, access, or compliance.

Private RAG

A RAG system designed to retrieve information from private or controlled data sources while enforcing appropriate access and security controls.

Public Benchmark

A standardized evaluation dataset or task that is publicly available and used to compare AI system performance.

Python

A widely used programming language in AI and machine learning for data processing, model development, experimentation, automation, and application development.

Q-Learning

A reinforcement learning algorithm in which an agent learns the expected value of taking an action in a given state and uses those values to improve its decision-making.

Q-Value

A value representing the expected future reward of taking a particular action in a particular state under a given policy.

Quantization

A technique that reduces the numerical precision used to represent a model’s parameters, such as converting higher-precision values to lower-precision formats, in order to reduce memory usage and improve inference efficiency.

Quantized Model

An AI model whose parameters or computations have been converted to a lower numerical precision to reduce memory consumption, latency, or computational requirements.

Quantization-Aware Training (QAT)

A training approach that accounts for the effects of quantization during training so that the resulting model can maintain better performance after being quantized.

Question Answering (QA)

An AI task in which a system receives a question and produces an answer based on its learned knowledge, provided context, or external information sources.

Extractive Question Answering

A question-answering approach in which the system identifies and extracts the answer directly from a provided text or document.

Generative Question Answering

A question-answering approach in which the system generates an answer rather than simply extracting a span of text from the source.

Question Answering System

An AI system designed to interpret questions and return relevant answers using language models, retrieval systems, databases, or other information sources.

Query

A request for information, search, computation, or action submitted to an AI system, database, search engine, or other software system.

Query Expansion

A retrieval technique that modifies or enriches a user’s query with additional terms, concepts, or related information to improve search results.

Query Rewriting

The process of transforming a user’s original query into a clearer or more effective query before sending it to a search or retrieval system.

Query Routing

The process of directing a query to the most appropriate model, tool, knowledge source, agent, or processing workflow.

Query Understanding

The process of interpreting a user’s query to identify its intent, entities, constraints, context, and other information needed to process it correctly.

Query Embedding

A vector representation of a query generated by an embedding model so that it can be compared with other vector representations during semantic retrieval.

Query-Time Retrieval

The retrieval of relevant information dynamically when a user query is received rather than relying only on information included during model training.

Query Performance

The effectiveness and efficiency of processing a query, commonly measured through factors such as relevance, accuracy, latency, and resource consumption.

Quality Assurance (QA)

A systematic process for verifying that an AI system meets predefined functional, quality, reliability, safety, and business requirements.

Quality Evaluation

The systematic assessment of AI outputs or system behavior against predefined quality criteria, evaluation datasets, metrics, or human judgments.

Quality Gate

A predefined set of quality criteria that must be satisfied before an AI system, model, feature, or release can proceed to the next stage of development or deployment.

Quality Metric

A measurable indicator used to assess a specific aspect of an AI system’s performance, such as accuracy, relevance, factuality, latency, or user satisfaction.

Quality Threshold

A predefined minimum or maximum value that an AI system must meet for a particular quality metric before it is considered acceptable.

Quantitative Evaluation

An evaluation approach based on numerical measurements, metrics, statistical analysis, or predefined scoring methods.

Qualitative Evaluation

An evaluation approach based on human judgment, expert review, user feedback, or structured analysis of AI outputs rather than numerical metrics alone.

Queue

A data structure or system mechanism that temporarily stores tasks, requests, or messages waiting to be processed.

Queue-Based Processing

An architecture in which requests or tasks are placed into a queue and processed asynchronously by available workers or services.

RAG — Retrieval-Augmented Generation

An AI architecture that retrieves relevant information from external data sources and provides it as context to a generative model to improve the relevance, accuracy, or grounding of its responses.

RAG Pipeline

The sequence of processes involved in a RAG system, typically including document ingestion, chunking, embedding, indexing, retrieval, context construction, generation, and evaluation.

RAG Evaluation

The process of evaluating the retrieval and generation components of a RAG system, including retrieval relevance, factuality, groundedness, completeness, and answer quality.

RAGAS — Retrieval Augmented Generation Assessment

An evaluation framework designed to assess RAG systems using metrics related to retrieval quality, faithfulness, context relevance, and answer relevance.

Reranking

A retrieval technique in which an additional model evaluates an initial set of retrieved results and reorders them according to their relevance to the query.

Re-ranker

A model or component that scores and reorders retrieved documents or passages to improve the relevance of the final context provided to an AI model.

Retrieval

The process of finding and returning relevant information from a dataset, database, document collection, search index, or knowledge base.

Retrieval-Augmented Agent

An AI agent that dynamically retrieves external information as part of its reasoning or task execution process.

Retrieval-Augmented Generation Evaluation

The systematic evaluation of how effectively a RAG system retrieves relevant information and uses it to produce accurate, grounded responses.

Retrieval Accuracy

A measure of how effectively a retrieval system identifies information that is relevant to a user’s query.

Retrieval-Augmented Search

A search approach that combines information retrieval with generative AI to retrieve relevant information and produce a natural-language response.

Recall

A classification or retrieval metric representing the proportion of relevant or positive items that were successfully identified.

Recall@K

A retrieval metric measuring the proportion of relevant items that appear within the top K retrieved results.

Reinforcement Learning (RL)

A machine learning approach in which an agent learns to make decisions by interacting with an environment and receiving rewards or penalties.

Reinforcement Learning from Human Feedback (RLHF)

A training approach in which human preferences or feedback are used to optimize an AI model toward desired behaviors or outputs.

Reasoning

The process by which an AI system derives conclusions, selects actions, or solves problems based on information, rules, learned patterns, or intermediate steps.

Reasoning Model

An AI model designed or optimized to perform more complex multi-step problem solving, planning, or inference.

Reasoning Trace

A sequence of intermediate information or steps associated with how a system arrives at an answer or decision.

ReAct — Reasoning and Acting

An agentic approach that combines reasoning steps with actions such as tool calls, searches, or interactions with external systems.

Recommendation System

An AI system that predicts or ranks items, content, products, or actions that may be relevant or useful to a user.

Regression

A supervised machine learning task in which a model predicts a continuous numerical value.

Regularization

A technique used during model training to reduce overfitting by discouraging overly complex models or parameter configurations.

Reproducibility

The ability to obtain consistent or comparable results by repeating an experiment or process under the same documented conditions.

Response Generation

The process through which a generative AI model produces an output in response to an input, prompt, or retrieved context.

Response Latency

The time between an AI system receiving a request and returning a response.

Response Validation

The process of checking an AI-generated response against predefined rules, schemas, sources, or quality requirements before it is used.

Response Streaming

A technique in which an AI system sends generated output progressively to the user instead of waiting for the complete response to be generated.

Retrieval-Augmented Fine-Tuning

A training approach that combines model adaptation with retrieved or externally provided information to improve performance on a specific domain or task.

Retrieval Index

A data structure optimized to allow efficient searching and retrieval of relevant documents, passages, vectors, or other information.

Retrieval Quality

The degree to which a retrieval system returns information that is relevant, complete, accurate, and useful for the downstream task.

Retrieval Failure

A situation in which a retrieval system fails to return the information needed to answer a query correctly, even when relevant information exists in the underlying data.

Risk Assessment

The systematic identification, analysis, and evaluation of potential risks associated with an AI system, including technical, operational, security, legal, ethical, and business risks.

Responsible AI

An approach to designing, developing, deploying, and governing AI systems with attention to safety, fairness, transparency, privacy, accountability, and compliance.

Robustness

The ability of an AI system to maintain acceptable performance when inputs, conditions, data, or environments vary from expected conditions.

Role-Based Access Control (RBAC)

An access-control mechanism that grants permissions based on a user’s or service’s assigned role.

Root Cause Analysis

A systematic process used to identify the underlying cause of a problem rather than only addressing its visible symptoms.

Safety

The property of an AI system operating without causing unacceptable harm or producing outputs that violate defined safety requirements.

Safety Evaluation

The process of testing an AI system for unsafe, harmful, or policy-violating behaviors before and after deployment.

Safety Guardrail

A technical or procedural control designed to prevent, detect, or mitigate unsafe or unwanted AI behavior.

Sampling

The process of selecting data points or generating outputs from a probability distribution or dataset.

Sampling Temperature

A parameter that controls the randomness of token selection during generative AI inference; higher values generally produce more varied outputs.

Semantic Search

A search approach that retrieves information based on meaning and semantic similarity rather than relying only on exact keyword matches.

Semantic Similarity

A measure of how similar two pieces of content are in meaning, often calculated using vector representations such as embeddings.

Semantic Retrieval

A retrieval approach that identifies information based on semantic meaning rather than exact lexical matches.

Semantic Chunking

A document segmentation technique that creates chunks based on meaningful semantic boundaries rather than only fixed character or token lengths.

Sequence-to-Sequence (Seq2Seq)

A model architecture designed to transform one sequence of data into another sequence, commonly used for tasks such as translation and summarization.

Sequence Length

The number of tokens or elements contained in an input or output sequence processed by a model.

Self-Attention

An attention mechanism that allows a model to determine which parts of an input sequence are most relevant to each other when processing the sequence.

Self-Supervised Learning

A machine learning approach in which training signals are generated from the data itself rather than requiring manually labeled examples for every training instance.

Sentiment Analysis

An NLP task that identifies the sentiment or emotional polarity expressed in text, such as positive, negative, or neutral.

Serving

The process of making an AI model available to applications or users so that it can receive inputs and return predictions or generated outputs.

Single-Shot Evaluation

An evaluation approach in which a model is tested using a single example, request, or interaction rather than a broader evaluation set.

Single Source of Truth (SSOT)

A designated authoritative source that an AI system or workflow should rely on for a particular piece of information.

Small Language Model (SLM)

A language model with substantially fewer parameters and lower computational requirements than large language models, often optimized for specific tasks or deployment constraints.

Softmax

A mathematical function that converts a set of numerical values into a probability distribution whose values sum to one.

Source Attribution

The practice of identifying the sources from which information used in an AI response was obtained.

Source Grounding

The process of ensuring that an AI response is supported by specified external sources or retrieved information.

Speech Recognition

The process of converting spoken language into machine-readable text or structured information.

Speech-to-Text (STT)

A technology that converts spoken audio into written text using speech recognition models.

Speech Synthesis

The process of generating human-like speech from text or other structured inputs.

System Prompt

A high-priority set of instructions that defines an AI model’s role, behavior, constraints, or operating context.

Structured Data

Data organized according to a predefined schema or format, such as tables, databases, JSON objects, or relational records.

Structured Output

An AI-generated response that follows a predefined format or schema, such as JSON, XML, or a specified set of fields.

Structured Generation

The process of constraining an AI model to generate outputs that conform to a predefined structure, schema, or grammar.

Supervised Learning

A machine learning approach in which a model learns from labeled examples containing inputs and their expected outputs.

Support Vector Machine (SVM)

A supervised machine learning algorithm that finds a decision boundary designed to separate data points belonging to different classes.

Summarization

An AI task that produces a shorter representation of a larger body of information while attempting to preserve its key meaning and relevant details.

System Evaluation

The systematic assessment of an AI application as a whole, including models, retrieval, prompts, tools, interfaces, and business workflows.

System Prompt Injection

An attack or manipulation attempt intended to override, interfere with, or exploit instructions provided through an AI system’s system-level configuration.

Safety Alignment

The process of designing or adapting an AI system so that its behavior follows predefined safety principles, policies, and human-defined objectives.

Scalability

The ability of an AI system to handle increasing numbers of users, requests, data volumes, or workloads while maintaining acceptable performance and cost.

Schema Validation

The process of checking whether structured data or an AI-generated output conforms to a predefined schema.

Search-Augmented Generation

An AI architecture that combines search or retrieval capabilities with generative models to produce responses using externally retrieved information.

Shadow Deployment

A deployment approach in which a new AI model or system processes real production inputs without its outputs being directly exposed to users, allowing performance to be evaluated safely.

Temperature

A parameter controlling the randomness of an AI model’s output during text generation; higher values generally allow more diverse outputs, while lower values tend to produce more predictable outputs.

Tensor

A multidimensional numerical structure used to represent and process data in machine learning models.

TensorFlow

An open-source machine learning framework developed by Google for building, training, and deploying machine learning models.

Text Classification

An NLP task in which text is assigned to one or more predefined categories.

Text Embedding

A numerical vector representation of text designed to capture semantic or contextual information so that texts can be compared mathematically.

Text Generation

The process of automatically producing text using an AI model based on an input, prompt, context, or other information.

Text-to-Image Generation

The process of generating images from textual descriptions using a generative AI model.

Text-to-Speech (TTS)

A technology that converts written text into synthesized spoken audio.

Token

A unit of text processed by a language model, which may represent a word, part of a word, punctuation mark, or other text fragment.

Token Budget

The maximum number of input and/or output tokens allocated to a model call or AI workflow.

Token Cost

The cost associated with processing input and output tokens when using a language model.

Token Limit

The maximum amount of tokenized input and output that a model can process within a defined context or request.

Tokenization

The process of converting text into smaller units called tokens so that a language model can process the text numerically.

Token Probability

The probability assigned by a language model to a possible next token during text generation.

Top-K Sampling

A text generation technique that restricts the next-token selection to the K tokens with the highest predicted probabilities.

Top-P Sampling

A text generation technique that restricts token selection to the smallest group of tokens whose cumulative probability reaches a specified threshold.

Tool Calling

The ability of an AI model or agent to request the execution of an external function, API, database query, or other tool.

Tool Use

The capability of an AI system or agent to interact with external tools or services to obtain information or perform actions.

Tool-Augmented AI

An AI system enhanced with external tools such as search engines, APIs, databases, calculators, or software applications.

Transformer

A neural network architecture based primarily on attention mechanisms and designed to process sequential or structured data efficiently.

Transformer Architecture

The architectural design underlying many modern language and multimodal models, based on components such as self-attention, feed-forward networks, and positional information.

Transfer Learning

A machine learning approach in which knowledge learned from one task or dataset is reused or adapted for another related task.

Training

The process of adjusting a model’s parameters using data so that it learns to perform a specific task or represent useful patterns.

Training Dataset

The dataset used to train a machine learning model by providing examples from which the model learns patterns or relationships.

Training Pipeline

The sequence of data preparation, training, validation, evaluation, and other processes used to build or update a machine learning model.

Training Data

The data used by a machine learning algorithm to learn patterns, relationships, or representations.

Training-Time

Referring to processes, computations, or decisions that occur while an AI model is being trained rather than when it is serving predictions or responses.

Test Dataset

A dataset reserved for evaluating a trained model on data that was not used during training.

Test Set

A collection of examples used to measure the final performance of a trained machine learning model.

Test Case

A defined input, expected behavior, and evaluation criterion used to verify whether an AI system behaves correctly under a particular scenario.

Test Harness

A set of software components and procedures used to systematically execute tests and evaluate the behavior of an AI system.

Testing Dataset

A dataset specifically prepared to evaluate an AI model or application against predefined scenarios and expected outcomes.

Threshold

A predefined value used to determine whether a score, probability, metric, or signal is sufficient to trigger a particular decision or action.

Throughput

The amount of data, requests, tokens, or tasks an AI system can process within a given period of time.

Time-to-First-Token (TTFT)

The time between an AI request being submitted and the first generated token being returned to the user.

Time-Series Data

Data collected or recorded sequentially over time, often used for forecasting, anomaly detection, and trend analysis.

Trace

A record of the sequence of operations, model calls, tool calls, retrieval steps, and other events involved in processing an AI request.

Tracing

The process of recording and analyzing the execution path of an AI application to understand its behavior, performance, errors, and dependencies.

Tool-Calling Agent

An AI agent capable of selecting and invoking external tools to gather information or perform actions as part of a task.

Token-Level Evaluation

An evaluation approach that analyzes model behavior or predictions at the individual token level rather than only evaluating the complete output.

Trustworthy AI

An approach to AI development that emphasizes properties such as reliability, safety, transparency, privacy, fairness, accountability, and robustness.

Underfitting

A modeling problem in which a model is too simple to capture the relevant patterns in the training data, resulting in poor performance on both training and unseen data.

Uncertainty

A measure or representation of how unsure an AI system or model is about a prediction, classification, generated output, or decision.

Uncertainty Estimation

The process of estimating how confident or uncertain a model is about its predictions or outputs.

Uncertainty Quantification (UQ)

The process of measuring and representing uncertainty in model predictions, data, parameters, or system outputs.

Unsupervised Learning

A machine learning approach in which a model learns patterns, structures, or representations from data without predefined labels for each training example.

Unstructured Data

Data that does not follow a predefined tabular or rigid schema, such as free-form text, documents, images, audio, or video.

Unstructured Text

Text that is not organized according to a predefined structured schema, such as articles, emails, reports, or natural-language documents.

Universal Embedding Model

An embedding model designed to generate useful vector representations across multiple domains, languages, or types of data.

Universal Language Model

A language model designed to perform effectively across a broad range of languages, domains, or language-related tasks.

User Intent

The underlying goal or purpose behind a user’s query, request, or interaction with an AI system.

User Intent Classification

The process of identifying and categorizing the purpose of a user’s request so that it can be handled by the appropriate workflow, model, agent, or tool.

User Feedback

Information provided by users about the usefulness, accuracy, relevance, or quality of an AI system’s behavior or outputs.

User Feedback Loop

A process in which user feedback is collected, analyzed, and used to improve an AI system, its prompts, data, models, or workflows.

User Embedding

A vector representation of information about a user, their preferences, behavior, or interactions, used in applications such as personalization or recommendation.

User-Level Evaluation

An evaluation of an AI system based on individual user interactions, behaviors, outcomes, or satisfaction rather than only aggregate technical metrics.

User Acceptance Testing (UAT)

A testing phase in which intended users or business stakeholders verify that an AI system meets practical business and functional requirements before deployment or release.

Utility Function

A function used to represent the value or desirability of different outcomes, actions, or decisions within an AI or optimization system.

Utility Model

A model or representation used to estimate the usefulness, value, or preference associated with different possible outcomes or actions.

Unicode

A universal character encoding standard that allows text from different writing systems and languages to be represented consistently by computer systems.

Unit Test

A software test that verifies the behavior of a small, isolated component of an application or AI workflow.

Unit Testing

The practice of testing individual software components or functions independently to detect errors and regressions.

Usage-Based Pricing

A pricing model in which the cost of an AI service depends on the amount of resources consumed, such as tokens, requests, compute time, or storage.

Usage Monitoring

The process of tracking how an AI system is used, including request volume, users, tokens, model usage, latency, costs, and other operational indicators.

Uncertainty-Aware AI

An AI system designed to explicitly account for uncertainty when making predictions, decisions, or generating outputs.

Universal Approximation

The theoretical property that certain neural network architectures can approximate a broad class of functions given sufficient capacity and appropriate conditions.

Validation

The process of checking whether an AI model, dataset, system, or output meets predefined requirements, constraints, or quality criteria.

Validation Dataset

A dataset used during model development to evaluate performance and tune model or training decisions without using it as the final test dataset.

Validation Loss

A loss value calculated on validation data to measure how well a model generalizes to data that was not used directly for parameter updates.

Validation Set

A subset of data reserved for evaluating and tuning a model during development.

Value Function

In reinforcement learning, a function that estimates the expected future reward associated with a state or state-action pair.

Variational Autoencoder (VAE)

A generative neural network architecture that learns a probabilistic latent representation of data and can generate new samples from that representation.

Vector

A numerical representation consisting of an ordered set of values, commonly used in AI to represent data such as text, images, or other objects.

Vector Database

A database designed to store, index, and efficiently search high-dimensional vector representations such as embeddings.

Vector Embedding

A numerical vector representation of an object, such as text, an image, or audio, designed to capture meaningful relationships or similarities.

Vector Index

A data structure optimized for efficiently searching and retrieving vectors based on similarity or distance.

Vector Search

A search technique that retrieves information by comparing vector representations and finding items that are mathematically similar to a query vector.

Vector Similarity

A measure of how closely two vectors are related according to a mathematical similarity or distance function.

Vector Store

A storage system used to store embeddings and associated metadata so that they can be efficiently retrieved using vector search.

Vector Retrieval

The process of retrieving information by searching for vectors that are similar to a query vector.

Vector Quantization

A technique that represents vectors using a smaller or more compact set of numerical values to reduce storage or computational requirements.

Verification

The process of checking whether an AI system or output conforms to specified requirements, rules, formats, or expected behavior.

Verification Dataset

A dataset used to verify that an AI system behaves correctly for predefined scenarios, requirements, or constraints.

Versioning

The practice of tracking and managing different versions of AI models, prompts, datasets, configurations, code, or system components.

Model Versioning

The practice of tracking different versions of an AI model and their associated configurations, datasets, metrics, and deployment states.

Prompt Versioning

The practice of tracking changes to prompts so that different prompt configurations can be tested, compared, reproduced, and managed.

Vector Database RAG

A RAG architecture that uses a vector database to store and retrieve embeddings representing chunks of external information.

Vision Model

An AI model designed to process, interpret, classify, or generate information from visual data such as images or video.

Vision-Language Model (VLM)

A multimodal AI model capable of processing both visual and textual information and relating information across these modalities.

Visual Question Answering (VQA)

An AI task in which a system answers natural-language questions about the content of an image or other visual input.

Video Understanding

The use of AI to analyze and interpret the content, events, actions, objects, or relationships within video data.

Voice AI

AI technology that enables systems to understand, process, generate, or interact through spoken language.

Voice Activity Detection (VAD)

A technology that identifies whether speech is present within an audio signal, often used in voice-based AI systems.

Vocabulary

The set of tokens, words, symbols, or subword units that a language model or tokenizer can represent.

Vocabulary Size

The number of unique tokens or symbols contained in a model’s vocabulary.

Vulnerability

A weakness in an AI system, model, application, infrastructure, or workflow that could be exploited or could lead to unintended behavior.

Vulnerability Assessment

The systematic process of identifying, analyzing, and prioritizing weaknesses that could affect the security or reliability of an AI system.

Weight

A learned numerical parameter in a machine learning model that influences how strongly an input contributes to the model’s output.

Weights

The collection of numerical parameters learned by a machine learning model during training.

Weight Initialization

The process of assigning initial values to a model’s weights before training begins.

Weight Decay

A regularization technique that penalizes large model weights during training to help reduce overfitting.

Weight Pruning

A model compression technique that removes or reduces the importance of selected model weights to decrease computational or memory requirements.

Weight Quantization

A technique that represents model weights using lower numerical precision to reduce memory usage and potentially improve inference efficiency.

Weight Sharing

A technique in which the same model parameters are reused across different parts of a neural network to reduce the number of parameters and computational requirements.

Web Scraping

The automated extraction of information from websites, often used to collect data for analysis, search, or machine learning applications.

Web Search

The process of retrieving information from web-based sources in response to a query, which can be integrated into AI systems to provide access to external or current information.

Web-Augmented Generation

An AI generation approach in which a model uses information retrieved from the web as additional context when producing a response.

Web Retrieval

The process of retrieving relevant information from web sources in response to a query.

Webhook

A mechanism that allows one system to automatically send information to another system when a specified event occurs.

Workflow

A defined sequence of tasks, actions, decisions, or system operations used to accomplish a particular objective.

AI Workflow

A workflow that incorporates one or more AI models, agents, retrieval systems, tools, or automated decision steps.

Workflow Automation

The use of software, AI, and predefined rules to execute repetitive or multi-step business processes with limited manual intervention.

Workflow Orchestration

The coordination, sequencing, and monitoring of multiple tasks, services, models, tools, or agents within a workflow.

Weak Supervision

A machine learning approach that uses incomplete, noisy, approximate, or automatically generated labels instead of relying entirely on manually labeled training data.

Word Embedding

A numerical representation of a word in a vector space designed to capture relationships between words based on their usage or context.

Word-Level Tokenization

A tokenization method that divides text into individual words or word-like units.

World Model

An internal representation learned or constructed by an AI system to model aspects of an environment, its states, and how they may change.

Worker

A process, service, or computing component responsible for executing a task, job, inference request, or workflow step.

Worker Queue

A queue containing tasks waiting to be processed by available workers, commonly used for asynchronous or distributed AI workloads.

Worst-Case Performance

The performance of an AI system under particularly difficult, adverse, or unfavorable conditions.

XAI — Explainable Artificial Intelligence

A set of methods and techniques designed to make the behavior, predictions, or decisions of AI systems more understandable to humans.

Explainability

The degree to which the behavior or output of an AI system can be understood and explained to humans.

Explanation

Information provided to help a user or stakeholder understand why an AI system produced a particular prediction, decision, or output.

Explainable Model

A model whose predictions or decisions can be interpreted or explained using understandable features, rules, representations, or other techniques.

Feature Attribution

A technique that estimates how much individual input features contributed to a model’s prediction or output.

SHAP — SHapley Additive exPlanations

An explainability method based on Shapley values that estimates the contribution of individual features to a model’s prediction.

LIME — Local Interpretable Model-Agnostic Explanations

An explainability technique that approximates the behavior of a complex model around a specific prediction using a simpler interpretable model.

XGBoost — Extreme Gradient Boosting

An optimized gradient boosting machine learning algorithm commonly used for structured or tabular data, classification, regression, and ranking tasks.

X-Vector

A learned vector representation commonly used in speaker recognition and related speech-processing applications to represent characteristics of a speaker.

X-Entropy

A term sometimes used informally to refer to cross-entropy, a loss function commonly used for classification and language modeling.

Cross-Entropy

A loss function that measures the difference between a target probability distribution and the probability distribution predicted by a model.

XAI Evaluation

The process of assessing whether explanations generated by an AI system are understandable, faithful to the model’s behavior, useful, and appropriate for their intended users.

XAI for LLMs

The application of explainability techniques to large language models to help users understand aspects of their outputs, behavior, or decision processes.

XML — Extensible Markup Language

A structured markup language used to represent, store, and exchange structured information, sometimes used for AI prompts, data pipelines, and system integrations.

YAML — YAML Ain’t Markup Language

A human-readable data serialization format commonly used for configuration files, AI pipelines, model deployment, workflow definitions, and infrastructure configuration.

Y-Label

The target or expected output variable in a supervised machine learning problem, commonly represented as y.

YAML Schema

A set of rules defining the expected structure, data types, and constraints of a YAML document.

YAML Configuration

A YAML-formatted configuration file used to define parameters, environments, workflows, services, or other settings in an AI system.

Y-Variable

The dependent or target variable that a machine learning model attempts to predict from input features.

YAML-Based Pipeline

An AI or machine learning pipeline whose stages, parameters, dependencies, or deployment configuration are defined using YAML.

YAML Validation

The process of checking whether a YAML document is syntactically correct and conforms to a required schema or structure.

Y-Axis

The vertical axis of a chart or visualization, commonly used in AI evaluation dashboards to represent metrics, values, probabilities, or model performance.

Yield

The proportion of useful or successful outputs produced by an AI or data-processing system relative to the total number of inputs or processing attempts.

YOLO — You Only Look Once

A family of real-time object detection models that identify and locate objects in images or video using a single neural network inference process.

YOLOv8

A version of the YOLO object detection family designed for tasks such as object detection, image classification, and image segmentation.

Yottabyte

A unit of digital information equal to approximately one septillion bytes, relevant when discussing extremely large-scale data infrastructure.

Zero-Shot Learning

A machine learning approach in which a model performs a task or recognizes categories that it was not explicitly trained on using previously learned knowledge or semantic relationships.

Zero-Shot

A model’s ability to perform a task without being given task-specific examples in the prompt or without having been specifically trained for that task.

Zero-Shot Classification

A classification task in which a model assigns inputs to categories without being explicitly trained on examples of those specific categories.

Zero-Shot Prompting

A prompting technique in which an AI model is given an instruction or task without providing example inputs and outputs.

Zero-Shot Retrieval

A retrieval approach in which a model performs retrieval or ranking without being specifically trained on labeled examples for the target retrieval task.

Zero-Shot Evaluation

The evaluation of a model on a task without providing task-specific examples or fine-tuning the model for that task.

Zero-Shot Transfer

The ability of a model to apply knowledge learned from one set of tasks, domains, or categories to a new task or category without task-specific training examples.

Z-Score

A statistical measure indicating how many standard deviations a value is above or below the mean of a distribution.

Zero-Shot Intent Classification

The classification of user intents using an AI model without providing labeled examples for the specific intents during inference.

Zero-Temperature

A model sampling configuration using a temperature value of zero or near zero to reduce randomness in generated outputs, although it does not necessarily guarantee deterministic behavior.

Z-Loss

A loss function or training technique used in some machine learning architectures to improve numerical stability or training behavior.

Z-Order

An ordering concept used in computer graphics and spatial or multidimensional data systems to determine the relative position or visibility of objects or data elements.

ZSL — Zero-Shot Learning

An abbreviation for Zero-Shot Learning, referring to the ability of a model to generalize to previously unseen classes or tasks without task-specific training examples.