Learn Full Stack
with tapaScript
By clicking Sign Up you re confirming that you agree with our Terms and Conditions.
Clear, beginner-friendly definitions of AI terminology — with real-world analogies and why each concept matters.
Showing 185 of 185 terms
Also known as: AI
Making computers perform tasks that normally require human intelligence.
💡 A robot that understands instructions and helps with everyday tasks.
It is the umbrella technology behind systems that understand language, recognise patterns, make decisions and generate content.
Also known as: ML
Teaching computers to learn patterns from examples instead of explicitly programming every rule.
💡 Teaching a child to recognise cats by showing many cat pictures instead of listing every feature.
It explains how AI systems learn from data and make predictions rather than merely follow fixed rules.
A type of machine learning that uses multiple layers of artificial neural networks to learn complex patterns.
💡 A student gradually learning simple concepts and combining them to solve complex problems.
It powers many modern AI capabilities, including language understanding, image recognition, speech processing and generative AI.
A computing system loosely inspired by how neurons connect in the human brain.
💡 A group of connected people, each processing information and passing it to the next person.
Neural networks are foundational building blocks of deep learning and help explain how many modern AI models process information.
Also known as: GenAI, Gen AI
AI that can create something new, such as text, images, music, video or code.
💡 An artist creating a new painting from a description.
It enables applications to generate content rather than only retrieve existing information.
A large, pre-trained AI model that can be adapted for many different tasks.
💡 A student who has studied many subjects and can be prepared for different professional roles.
Most modern AI applications build on existing foundation models rather than training models from scratch.
Also known as: LLM
An AI model trained on large amounts of data to understand and generate human language.
💡 A person who has read a huge library and can answer questions or write based on what they learned.
LLMs power chatbots, coding assistants, document analysis, AI agents and many other applications.
AI that can understand or generate multiple types of information, such as text, images, audio and video.
💡 A person who can read, look at photographs and listen to conversations to understand a situation.
It allows applications to understand more than text, such as analysing an image and answering questions about it.
The process of using a trained AI model to generate a prediction or response.
💡 Asking a student a question after they have studied and receiving an answer.
Every model request in an application involves inference, which affects latency and cost.
The process of teaching an AI model using data so it can learn patterns and relationships.
💡 Teaching a student with books, examples, exercises and feedback.
It distinguishes building or adapting a model from using an existing one. Most application developers use trained models.
An AI model that has already learned from a large dataset before you start using it.
💡 Hiring an experienced employee instead of teaching someone everything from the beginning.
Developers can build applications without investing in expensive training infrastructure.
Also known as: Open Weights
A model whose trained parameters, called weights, are publicly available under a particular licence.
💡 Getting a recipe so you can prepare the dish in your own kitchen.
It lets developers download, run or customise models on their own infrastructure, subject to the licence.
A small piece of text that an AI model processes; it may be a word, part of a word or punctuation.
💡 Cutting a sentence into small Lego blocks before assembling and understanding it.
Token usage affects API costs, context limits and processing time.
The process of breaking text into smaller pieces called tokens.
💡 Breaking a sentence into individual word cards before studying it.
Different models may tokenize the same text differently, affecting cost estimates and context usage.
Also known as: Context Length, Context Limit
The maximum amount of information a model can consider in a particular interaction or request.
💡 A desk with limited space for keeping books and documents open at the same time.
It determines how much conversation history, code or documentation can be supplied at once.
Numerical values inside a model that are learned during training and influence its behaviour.
💡 Tiny adjustments made to a musical instrument to produce a desired sound.
Parameter count is often used to describe model size, although more parameters do not automatically mean a better model.
The learned numerical values that determine how strongly different connections influence a neural network's output.
💡 The importance assigned to different subjects when calculating a student's final grade.
Weights represent learned patterns and matter when working with open-weight models, quantization and model hosting.
A generation setting that influences how predictable or varied a model's responses are.
💡 Asking someone for a standard answer versus encouraging a creative answer.
It helps control response variation; lower values generally produce more predictable outputs and higher values encourage more variation.
A sampling setting that limits next-token choices to likely candidates whose combined probability reaches a specified threshold.
💡 Asking a restaurant to choose from the most popular dishes rather than the entire menu.
It provides another way to control response generation and output diversity.
A model designed to use additional computation to work through complex problems before answering.
💡 Giving a student extra time to solve a difficult mathematics problem.
These models are useful for coding, mathematics, planning and multi-step problem-solving.
A reasoning approach involving intermediate steps when solving a problem.
💡 Solving a mathematics problem step by step instead of jumping to the answer.
It helps explain stepwise problem-solving in AI; not every model exposes its internal reasoning.
Also known as: Embeddings, Vector Embedding
A numerical representation of information, such as text or images, that captures aspects of its meaning.
💡 Giving each book a numerical address based on its subject so similar books can be found together.
Embeddings enable semantic search, similarity matching, recommendations and RAG.
An ordered collection of numbers representing information in mathematical space.
💡 Describing a location using coordinates such as latitude and longitude.
Vectors are the basic building blocks of embedding-based search systems and vector databases.
A measurement of how close two vectors are to each other in mathematical space.
💡 Finding people with similar interests even when they describe those interests differently.
It helps identify semantically similar information even when the exact words differ.
Also known as: Confabulation
When an AI produces information that sounds convincing but is incorrect, fabricated or unsupported.
💡 A student confidently giving an answer despite not knowing whether it is correct.
Hallucinations can damage trust and cause business problems; evaluation, grounding and validation can reduce their impact.
Reducing the numerical precision used to represent model weights, making models smaller and often faster.
💡 Compressing a large photograph so it uses less storage while retaining acceptable quality.
It can reduce memory requirements and infrastructure costs, allowing larger models to run on less powerful hardware.
Training a smaller model to imitate the capabilities or behaviour of a larger model.
💡 An experienced teacher creating a simplified study guide with the most important lessons.
It can create smaller, cheaper and faster models for specific tasks.
An instruction, question or information provided to an AI model.
💡 Asking a chef to prepare a particular dish with specific ingredients.
Prompts define the task and provide the immediate input that influences the model's response.
Also known as: Prompting
Designing and refining instructions to get useful and consistent outputs from AI models.
💡 Giving a new employee clear instructions instead of simply saying, “Do something useful.”
Good prompts reduce ambiguity and help developers control how AI applications behave.
The main set of instructions defining how an AI model should behave within an application.
💡 A company's employee handbook explaining roles, responsibilities and rules.
It establishes the model's role, rules, boundaries, response style and task-specific behaviour.
The actual question or instruction provided by the end user.
💡 A customer asking a support representative a specific question.
It represents the immediate task the AI needs to address and is generally combined with system instructions and relevant context.
Giving an AI model a few examples before asking it to perform a task.
💡 Showing a child a few solved puzzles before asking them to solve another.
Examples establish expected behaviour, output structure and patterns without changing the model's training.
Asking an AI model to perform a task without providing examples.
💡 Asking someone to write a letter without showing a sample.
It is useful for straightforward tasks and tests what a model can do from instructions alone.
Selecting and providing the right instructions, information, tools and conversation history to an AI model at the right time.
💡 Preparing an employee with relevant documents, customer history, policies and tools before assigning a task.
The quality of available context often determines the quality of results in sophisticated AI applications and agents.
Deciding what information should be included, retained, summarised or removed from the model's context.
💡 Keeping important documents on your desk and putting unrelated files away.
It prevents context overload, reduces unnecessary token usage and keeps information relevant.
A reusable prompt containing placeholders that can be filled with dynamic information.
💡 A standard email template where the customer's name and order number are replaced each time.
It generates consistent prompts for different users, requests and application scenarios.
Asking an AI model to return information in a predefined format, such as JSON.
💡 Asking someone to fill out a standard form instead of writing an unstructured paragraph.
It makes responses easier to consume programmatically and integrate into frontends, APIs and databases.
A specification describing the expected structure, data types and validation rules of a JSON object.
💡 A form specifying required fields and what kind of information each field accepts.
It helps ensure generated data follows a predictable structure and reduces integration errors.
Restricting an AI model's output to a predefined format, structure or set of allowed values.
💡 Asking a student to answer only Yes or No rather than writing an unrestricted essay.
It improves reliability when outputs must meet strict requirements, such as valid JSON or a fixed list of categories.
A malicious instruction designed to manipulate an AI model into ignoring intended instructions or performing unintended actions.
💡 A stranger slipping fake instructions into an employee's paperwork to trick them into breaking company rules.
It is a major security risk when applications process untrusted documents, websites, emails or user inputs.
Reusing previously processed portions of a prompt when supported by the model provider.
💡 Keeping frequently used documents ready instead of retrieving and reading them from scratch each time.
It can reduce repeated processing costs and improve latency when large instructions or context are sent repeatedly.
Also known as: RAG, Retrieval Augmented Generation
An approach where AI searches relevant external information before generating an answer.
💡 An open-book examination where a student looks up relevant information before answering.
It lets applications answer using company documents, private databases and up-to-date knowledge without retraining the model.
A collection of information that an AI application can search and use to answer questions.
💡 A company library containing policies, product manuals and operational documents.
It provides domain-specific knowledge that general-purpose models may not already possess.
The process of collecting and bringing information into an AI system for processing and retrieval.
💡 Collecting books from different locations and bringing them into a central library.
It is the starting point for knowledge-based applications; data must be prepared before it can be searched effectively.
Breaking large documents into smaller pieces of information.
💡 Dividing a textbook into chapters and sections so a topic can be found quickly.
Smaller chunks make retrieval more focused and provide relevant context without sending entire documents to the model.
A model designed to convert information into numerical vectors representing its semantic characteristics.
💡 A translator that converts sentences into numerical coordinates so a computer can compare their meanings.
It enables semantic search and is an important component of most vector-based RAG architectures.
Also known as: Vector DB, Vector Store
A database designed to store and search vector representations of information.
💡 A specialised library that groups books by similarity of subject and meaning.
It lets AI applications efficiently retrieve semantically similar information and is common in RAG systems.
Searching for information based on meaning and intent rather than exact keyword matches.
💡 Asking a librarian for books about happiness and receiving books about well-being.
It finds relevant information even when users use different words from those in the original documents.
Finding items whose numerical representations are closest to a given vector.
💡 Finding houses most similar to a house you like based on location, size and features.
It is used in vector databases, recommendation systems and semantic search applications.
Additional information attached to a document or data record, such as author, date, department or category.
💡 A library catalogue showing a book's author, publication date, genre and shelf number.
It helps filter search results and retrieve information based on business rules, permissions and document attributes.
Combining keyword-based search with semantic search.
💡 Searching a library by exact title and by a description of what the book is about.
It captures both exact matches and meaning-based relationships, useful for technical and business documents.
Reordering retrieved search results according to how relevant they are to a particular query.
💡 A recruiter shortlisting candidates and then ranking the best ones against job requirements.
It puts the most useful information first before the results are provided to the LLM.
Finding and collecting relevant information from an external knowledge source.
💡 Looking up the right pages in a reference book before answering a question.
It is the information-gathering stage of RAG, and its quality directly influences the generated response.
Connecting an AI-generated response to trusted information, evidence or supplied context.
💡 Asking a student to support an answer with a reference from the textbook.
It helps reduce unsupported claims and makes responses more reliable in business applications.
Showing the source from which information in an AI-generated answer was obtained.
💡 A research paper identifying the book and page from which a fact was taken.
It improves transparency, lets users verify information and helps establish trust.
An approach to RAG that uses knowledge graphs and relationships between entities to improve retrieval.
💡 A librarian finding books about a person and books connecting that person to colleagues, projects and organisations.
It is useful for questions involving relationships across documents, people, events or business entities.
An approach where an AI agent decides how, when and where to retrieve information while solving a task.
💡 A researcher deciding which library to visit, which books to consult and what to investigate next.
It enables flexible research when a question requires multiple searches or information sources.
An AI-powered system that can understand a goal, make decisions and take actions using available tools.
💡 A personal assistant who understands a request, plans the work and uses resources to complete it.
Agents move applications beyond text generation into tasks such as retrieving data, updating records and interacting with services.
An approach to building AI systems that work towards goals by making decisions, taking actions and adapting based on results.
💡 An employee who can work out how to complete an assignment rather than following only a fixed checklist.
It enables software to handle dynamic, multi-step tasks rather than relying entirely on predefined instructions.
Also known as: Agentic Loop
The repeating process in which an AI agent reasons about a task, takes an action, observes the result and decides what to do next.
💡 A driver checking a route, taking a road, observing traffic and choosing the next route.
It is fundamental to how agents execute multi-step tasks and respond to changing situations.
Allowing an AI model to request the execution of an external tool or capability.
💡 An assistant using a calculator or telephone rather than merely explaining how to use one.
It enables AI applications to interact with databases, APIs, search engines and software instead of only generating text.
A structured mechanism through which an AI model requests a specific function with the required arguments.
💡 Asking an assistant to call a particular department and giving the exact extension and information.
It connects AI models to application logic in a controlled and predictable way.
Breaking a complex goal into smaller steps that can be executed in sequence or in parallel.
💡 Planning a holiday by choosing a destination, booking tickets, finding accommodation and preparing an itinerary.
It helps agents organise complex tasks, manage dependencies and decide which actions are necessary.
Analysing information, evaluating possibilities and arriving at a conclusion or decision.
💡 A detective examining clues and connecting evidence before reaching a conclusion.
It helps AI systems handle tasks involving multiple conditions, decisions and problem-solving steps.
A prompting and agent design pattern that combines reasoning and actions in an iterative cycle.
💡 A detective thinking about the next clue, investigating it and deciding what to investigate next.
It helps agents solve tasks by alternating between deciding what to do and interacting with external tools.
A process where AI agents perform multiple connected steps and make decisions while completing a task.
💡 A travel assistant searching flights, comparing prices, checking hotels and preparing an itinerary.
It automates business processes that require flexibility, retrieval and interaction with different systems.
Coordinating multiple models, agents, tools and workflows to complete a larger task.
💡 A project manager coordinating designers, developers, testers and business analysts.
It matters when several specialised components must work together reliably.
A system where multiple AI agents work together to accomplish a task.
💡 A company where finance, HR, sales and operations teams collaborate on a common goal.
It divides complex responsibilities among specialised agents rather than assigning everything to one agent.
A specialised agent that handles a smaller task as part of a larger agent's workflow.
💡 A project manager assigning research to a specialist team member.
It divides complex problems into manageable responsibilities and allows specialised work to happen independently.
Transferring control or responsibility from one agent to another.
💡 A hospital receptionist transferring a patient to the appropriate doctor.
It enables specialised agents to collaborate and pass tasks to the component best suited for the next step.
Also known as: HITL
Involving a human to review, approve or intervene in an AI-driven process.
💡 A bank requiring a manager's approval before a large transaction.
It is essential when actions involve money, sensitive information, legal decisions or irreversible operations.
An AI agent capable of performing tasks with limited continuous human intervention.
💡 A self-driving vehicle navigating a route without someone controlling every movement.
It can automate long-running and repetitive processes, but still needs permissions, monitoring and safety controls.
Information an agent retains and retrieves to maintain continuity across interactions or tasks.
💡 A personal assistant's notebook of preferences, previous requests and ongoing commitments.
It helps agents remember relevant facts, previous interactions and task information instead of starting from zero.
The current working information of an agent, including progress, variables, decisions and intermediate results.
💡 A project board showing tasks that are completed, pending or in progress.
It supports multi-step tasks, recovery from failures and resumption of long-running workflows.
Also known as: Application Programming Interface
A defined interface that allows two software applications to communicate with each other.
💡 A restaurant menu showing customers what they can order and how to request it.
APIs are the foundation of software integration and let AI applications access external services and business systems.
An external capability that an AI model or agent can use to perform a specific action.
💡 Giving an employee a calculator, telephone, computer and database to complete work.
Tools let AI search the internet, query databases, send emails, calculate values and interact with software.
Also known as: MCP
A standard protocol that helps AI applications connect to external tools, resources and prompt templates.
💡 A universal adapter that lets different devices connect to compatible accessories.
It reduces the need for separate custom integrations for every AI client and makes capabilities reusable across compatible clients.
A program that exposes tools, resources and prompts through the Model Context Protocol.
💡 A service counter providing different facilities to authorised visitors.
It lets AI clients discover and use capabilities provided by external applications, services and data sources.
The component that connects to an MCP server and communicates with it using the protocol.
💡 A customer visiting a service counter to request available services.
It lets AI applications discover available tools and resources and interact with external systems.
The AI application that manages MCP client connections and provides the environment for MCP interactions.
💡 A hotel managing its reception desk and coordinating with different service providers.
Understanding the host-client-server relationship helps developers design and debug MCP integrations.
Functions exposed by an MCP server that an AI application can discover and invoke.
💡 Services available at a service centre, such as printing, verification and parcel delivery.
They let AI applications perform actions such as searching records, creating tickets or updating business information.
Data and information exposed by an MCP server for an AI application to read.
💡 A library giving visitors access to books and reference materials.
Resources provide external context such as documents and database records without requiring every interaction to be an action.
Reusable prompt templates made available through an MCP server.
💡 Standard forms and instruction sheets available at an office reception.
They let compatible AI applications reuse predefined instructions and workflows for common tasks.
Also known as: A2A
A protocol that enables different AI agents to communicate, collaborate and delegate tasks.
💡 Different companies collaborating through a common business language.
It helps agents built with different frameworks or platforms work together through a common communication mechanism.
A structured description of an agent's identity, capabilities and connection information.
💡 A business card containing a person's name, role, skills and contact information.
It helps other agents discover what an agent can do and how to communicate with it.
Also known as: Open Authorization
An authorisation standard that lets an application access resources on behalf of a user without directly handling their password.
💡 Giving a hotel permission to collect your luggage without handing over the key to your entire house.
It matters when agents connect to services such as email, calendars, cloud storage and business applications.
A mechanism through which one application automatically sends information to another when an event occurs.
💡 A delivery company automatically messaging you when your parcel reaches a location.
It enables event-driven AI workflows, such as triggering an agent when an invoice arrives or a customer submits a form.
A central entry point that receives and manages API requests before forwarding them to the appropriate services.
💡 A hotel reception desk directing visitors to the correct department.
It helps manage authentication, routing, rate limits, monitoring and access control across AI backend services.
An AI-powered tool that helps developers write, understand, explain and improve code.
💡 A colleague suggesting improvements while you write code.
It accelerates everyday development activities such as code completion, debugging, documentation and implementation.
An AI system that can inspect a codebase, modify files, execute commands and test changes to complete development tasks.
💡 A junior developer who inspects a project, implements tasks, runs tests and reports results.
It moves beyond suggestions and lets AI participate in actual software engineering workflows.
Building software primarily by describing requirements in natural language and iteratively refining AI-generated implementations.
💡 Describing your dream house to an architect and asking for changes until the design looks right.
It makes software creation accessible to more people, but generated code still needs review, testing and maintenance.
Working with AI as a collaborative programming partner during software development.
💡 Working with another developer who discusses approaches and suggests solutions while you code.
It helps developers brainstorm, understand unfamiliar code, debug issues and explore implementation alternatives.
Automatically producing source code based on instructions, specifications or existing code.
💡 Giving an architect a building specification and asking for an initial design.
It reduces repetitive coding effort and helps developers move from requirements to implementation faster.
The ability of an AI system to inspect and understand an existing software repository.
💡 A new employee studying existing processes and documentation before changing them.
It helps agents understand dependencies, architecture and existing implementation before making changes.
Building an understanding of how files, modules, components and dependencies connect within a codebase.
💡 A building blueprint showing rooms, corridors, electrical connections and plumbing.
It helps coding agents navigate large repositories, identify relevant files and understand the impact of changes.
Defining requirements, specifications and expected behaviour before asking AI or developers to implement a solution.
💡 Giving an architect a detailed building plan before construction starts.
It reduces ambiguity, improves implementation quality and makes generated code easier to review against requirements.
Using AI to inspect source code and identify potential bugs, security issues and improvement opportunities.
💡 An experienced colleague reviewing your work before it is submitted to a client.
It provides an additional review layer and can help teams catch problems earlier in development.
Automatically generating test cases or test code using AI.
💡 Asking an assistant to prepare practice questions covering every chapter of a textbook.
It accelerates test creation, improves coverage and helps identify scenarios developers might overlook.
Using AI agents to perform larger software development tasks with reduced continuous human intervention.
💡 A development team that plans, executes and validates a project with limited supervision.
It represents AI participation across planning, implementation, testing and debugging; human review and accountability remain essential.
Allowing an AI system to interact with a computer through its graphical interface by clicking, typing and navigating.
💡 A human assistant opening applications, clicking buttons and entering information.
It lets agents operate applications that do not provide convenient APIs or direct integrations.
An AI agent that interacts with a command-line interface to execute commands and manage development tasks.
💡 A technician operating a machine through a command console rather than physical buttons.
It automates activities such as running tests, installing dependencies, inspecting files and executing scripts.
A software development kit providing reusable libraries and interfaces for integrating AI capabilities into applications.
💡 A ready-made toolkit containing the equipment needed for a job.
It simplifies model communication, streaming, tool calling and structured outputs, reducing integration code.
A collection of abstractions and utilities that helps developers build AI-powered applications.
💡 A construction framework providing standard building blocks instead of starting every building from scratch.
It simplifies recurring problems such as prompt management, model integration, retrieval and workflow execution.
A framework designed to help developers create, execute and manage AI agents.
💡 A management system for organising employees, assigning responsibilities and tracking work.
It provides abstractions for tool use, state management, planning and orchestration.
A framework and ecosystem for developing applications powered by LLMs and external integrations.
💡 Compatible building blocks that can be assembled into a working system.
It provides components for connecting models, prompts, tools and data sources.
A framework for building stateful, graph-based AI workflows and agents.
💡 A flowchart showing different routes a project can take depending on decisions and outcomes.
It supports complex workflows requiring controlled execution, state persistence, branching, human approval and multi-agent coordination.
Making a trained AI model available so applications can send requests and receive responses.
💡 Opening a restaurant kitchen so customers can order and receive meals.
It is necessary when deploying your own models and making them accessible to users or other services.
Software responsible for executing an AI model and generating predictions or responses.
💡 The engine inside a machine that performs the work after receiving instructions.
It influences execution speed, memory usage and hardware efficiency, especially when hosting models yourself.
Also known as: Graphics Processing Unit
A Graphics Processing Unit, a specialised processor capable of performing many calculations simultaneously.
💡 A factory with thousands of workers performing smaller calculations simultaneously.
GPUs accelerate the mathematical operations required for training and running many large AI models.
Running an AI model on your own or a cloud provider's infrastructure and making it available for use.
💡 Owning or renting a kitchen where you prepare food instead of ordering from another restaurant.
It gives developers control over deployment, performance, privacy, availability and infrastructure choices.
A language model that runs on your own computer or infrastructure rather than relying entirely on a remote provider.
💡 Cooking in your own kitchen rather than ordering from an external restaurant.
It can provide greater control over data, privacy and offline use, subject to hardware and model limitations.
A language model hosted by an external cloud provider and accessed through an API or service.
💡 Ordering food from a restaurant instead of maintaining your own kitchen.
It lets developers use powerful models without managing hardware, deployment and scaling infrastructure.
A model where users provide their own AI provider API key to access AI capabilities through an application.
💡 Bringing your own ingredients to a shared kitchen instead of paying for everything the kitchen supplies.
It lets developer tools support multiple providers while users manage their own AI usage and billing.
A central layer that manages and routes requests to different AI models and providers.
💡 An airport control tower directing flights to different runways based on availability and destination.
It simplifies model switching, provider fallback, usage tracking, cost management and access control.
Coordinating multiple AI agents so they collaborate on different parts of a larger task.
💡 A project manager coordinating specialist teams to deliver one large project.
It manages complex applications where specialised agents communicate, delegate work and combine results.
A group of AI agents that collaborate, often in parallel, to solve a larger problem.
💡 Search-and-rescue teams looking in different areas simultaneously for a missing person.
It distributes research, analysis and execution tasks, although coordination overhead and consistency need management.
AI agents that execute tasks over extended periods while maintaining their progress.
💡 A construction project manager tracking a building project over several months.
They are useful for research, monitoring and business processes that cannot finish in one interaction.
AI agents that execute tasks without requiring continuous user interaction.
💡 A washing machine continuing to wash clothes after you leave the room.
They enable asynchronous automation such as document processing, monitoring and scheduled work.
An approach where AI decides what to search, which sources to investigate and what additional searches are needed.
💡 A journalist following leads and discovering new questions while investigating a story.
It supports complex research beyond a single search query, especially when several sources and investigative steps are required.
An AI capability that performs multi-step research across sources and produces a structured analysis.
💡 A research analyst examining reports and preparing a comprehensive report.
It helps with complex research involving source discovery, comparison, synthesis and evidence gathering.
The design of how an AI application stores, organises and retrieves information across interactions.
💡 A company's filing system for finding previous records and decisions.
It determines what a system can remember and how it maintains continuity over time.
A type of memory that stores specific past events or interactions.
💡 A diary containing specific events organised by date.
It helps agents recall previous conversations, completed tasks and decisions.
A type of memory that stores general facts, knowledge and concepts rather than individual events.
💡 An encyclopaedia containing general knowledge rather than a personal diary.
It lets AI systems retain knowledge that can be applied across different conversations and tasks.
Information an AI system currently needs to complete an active task.
💡 A student's rough notes and calculations while solving a problem.
It helps agents maintain relevant context during execution without permanently storing every intermediate detail.
Selecting which AI model should handle a particular request or task.
💡 A hospital directing patients to different specialists based on their needs.
Models differ in reasoning ability, cost and speed; routing balances quality, performance and expenses.
Switching to another model or provider when the primary model fails or is unavailable.
💡 A backup electricity generator when the main power supply fails.
It improves reliability during provider outages or request failures.
Running multiple independent tasks simultaneously instead of waiting for each to finish sequentially.
💡 Five employees completing five separate assignments at the same time.
It reduces execution time when an AI workflow contains independent tasks.
The amount of computation or reasoning effort allocated to solving a problem, where supported by a model.
💡 Giving a student ten minutes for a simple question and an hour for a difficult problem.
It can influence the balance between response quality, latency and computational cost.
An AI agent capable of understanding and acting on text, images, audio, video and other modalities.
💡 An assistant who can read documents, inspect photographs, listen to messages and act on them.
It enables applications to handle real-world tasks involving visual information, voice and other non-text inputs.
A machine learning architecture involving two neural networks competing to generate realistic outputs.
💡 A counterfeiter trying to make realistic money while a detective tries to identify fake notes.
GANs are important in generative AI history and are used in some image generation and data synthesis applications.
A neural network architecture that uses attention mechanisms to process relationships between different parts of information.
💡 A student reading a paragraph and noticing how different words relate to one another.
It is the foundation of most modern LLMs and transformed language processing and generative AI.
A technique that allows a model to focus more on relevant parts of the input while processing information.
💡 Listening to a conversation and paying extra attention to important details.
It helps models understand relationships between words, sentences and other elements in complex inputs.
A mechanism that allows a model to examine relationships between different tokens within the same input sequence.
💡 Reading a paragraph and understanding how its beginning, middle and end relate.
It helps transformer models understand context and relationships between words, even when they are far apart.
Also known as: Fine Tuning
Further training an already pre-trained model using additional task-specific or domain-specific data.
💡 Sending an experienced doctor for specialised training in cardiology.
It adapts a model to specialised tasks or behaviour when prompting and retrieval are insufficient.
A technique that adapts a large model by training a relatively small number of additional parameters.
💡 Training an experienced employee in a new skill without repeating their entire education.
It reduces resources needed to customise models compared with updating all original weights.
A family of techniques that customise models by updating only a small portion of parameters or adding lightweight trainable components.
💡 Giving an employee a short specialist course instead of sending them back to university for a full degree.
It makes model adaptation more affordable without the resources required for full fine-tuning.
Artificially generated data used for training, testing or evaluating AI systems.
💡 Creating imaginary customer records to test a banking application without exposing real customer information.
It creates additional examples, simulates rare scenarios and can help protect sensitive information when real datasets are limited.
Also known as: MoE
A model architecture containing multiple specialised components, with selected experts handling different inputs.
💡 A company with specialists where only relevant experts are assigned to a particular project.
It can increase model capacity while activating only parts of the model for a request, potentially improving computational efficiency.
A structured representation of entities and the relationships between them.
💡 A family tree showing people and their relationships rather than simply listing names.
It helps AI applications understand connections between people, products, organisations and events.
A generative model that learns to create data by gradually reversing a process of adding noise.
💡 Creating a detailed painting by starting with an unclear canvas and gradually refining it.
Diffusion models are widely used in image generation and also applied to audio, video and other generative tasks.
Also known as: ASR, Speech Recognition, STT
Converting spoken language into written text.
💡 A secretary listening to a meeting and typing what participants say.
It enables voice-based applications, meeting transcription, voice commands and conversational AI.
Also known as: TTS
Converting written text into spoken audio.
💡 A person reading a written letter aloud.
It enables voice communication, accessibility and conversational interfaces.
AI systems that process and respond to information continuously or with very low latency.
💡 Having a live conversation rather than exchanging letters and waiting for replies.
It enables interactive voice assistants, live translation, real-time monitoring and responsive experiences.
Running AI models directly on devices such as smartphones, laptops, cameras or embedded systems instead of relying entirely on cloud infrastructure.
💡 Having a small personal calculator in your pocket instead of sending every calculation to a central office.
It can reduce latency, improve offline capabilities and provide greater control over data processing.
Also known as: Evals
Testing and measuring whether an AI system produces useful, accurate and reliable results.
💡 Conducting regular examinations to check whether a student is learning and improving.
AI outputs can vary, so traditional software tests are not enough. Evals measure quality and detect regressions.
A collection of test inputs, expected outcomes or evaluation criteria used to assess an AI system.
💡 A question paper and answer key used to evaluate students consistently.
It provides a consistent basis for comparing models, testing prompts and measuring improvements over time.
Also known as: LLM Judge
Using an AI model to evaluate the quality of responses generated by another AI system.
💡 An experienced teacher evaluating another student's answer using a grading rubric.
It automates evaluation of qualities such as relevance, completeness and clarity, although human validation and calibration remain important.
A measurement of how often a system produces correct results.
💡 Checking how many questions a student answers correctly in an examination.
It is useful for evaluating classification and prediction, although it can hide problems in imbalanced datasets.
The proportion of results identified as positive or relevant that are actually correct or relevant.
💡 A security guard flagging visitors and measuring how many flagged people were actually suspicious.
It measures how often positive predictions or retrieved results are trustworthy.
The proportion of all actual positive or relevant items that the system successfully identifies.
💡 Checking whether a security guard identified all suspicious visitors entering a building.
It measures whether a system misses important information, especially in search, retrieval and classification.
The extent to which an AI-generated response is supported by the information and evidence provided to the model.
💡 Checking whether a student's answer can be justified using the provided textbook.
It helps detect unsupported statements and assess whether responses are based on retrieved documents or supplied facts.
The ability to understand what happens inside an AI application by examining execution, inputs, outputs, tool calls, errors and performance.
💡 CCTV cameras and activity records throughout a factory to understand where problems occur.
It is essential for debugging systems involving LLMs, agents, retrieval and multiple external services.
Recording the sequence of operations performed during a request or workflow.
💡 Following a parcel through every delivery centre to find where it was delayed.
Traces help identify which model call, tool execution or retrieval step caused a failure or delay.
Recording important events, errors, requests and system activities.
💡 A security register recording who entered a building and when.
Logs help investigate failures, understand system behaviour and maintain operational visibility.
The time taken for a system to respond to a request.
💡 Waiting for a restaurant to serve food after placing an order.
High latency makes applications feel slow; monitoring and optimisation are important for user experience.
The amount of work a system can process within a given period.
💡 Measuring how many customers a restaurant can serve in one hour.
It helps determine how many simultaneous users or requests an AI application can support.
The number of input and output tokens consumed during AI interactions.
💡 Checking how much fuel a vehicle consumes during a journey.
It helps monitor model consumption, estimate API expenses and identify inefficient prompts or workflows.
Rechecking previously working functionality after changes to ensure it still works correctly.
💡 Checking that existing electrical connections still work after renovating a room.
AI models, prompts and retrieval pipelines can change behaviour after updates; regression testing detects quality degradation.
Comparing AI models or systems using standardised tests and measurements.
💡 Comparing students using the same examination paper and marking criteria.
It helps evaluate differences in quality, speed, cost and capabilities before selecting a model.
A change in data patterns or system behaviour over time that can reduce model performance.
💡 A navigation system using an old map after roads and traffic patterns have changed.
Monitoring drift helps identify when a system needs updated data, revised prompts, new evaluations or other adjustments.
Rules and controls that restrict AI behaviour and prevent unwanted or unsafe operations.
💡 Safety barriers and rules in a factory that prevent workers entering dangerous areas.
They enforce business policies, limit inappropriate outputs and reduce risks when AI interacts with users or external systems.
Checking and validating information before accepting or processing it.
💡 A security guard checking visitors' identity documents before allowing entry.
It helps prevent malformed, unexpected or malicious input from entering an AI application and downstream systems.
Checking AI-generated responses before using them in an application or executing an action.
💡 A quality inspector checking a manufactured product before it reaches customers.
It prevents invalid JSON, unexpected values and unsafe outputs from entering databases or triggering unintended operations.
The unintended exposure of confidential, sensitive or private information.
💡 Accidentally leaving confidential company documents in a public place.
AI applications may process employee records, customer data and business documents, so preventing leakage is essential.
Information that can identify a particular individual, such as a name, email address, phone number or identification number.
💡 An identity card containing a person's name, photograph and identification number.
AI engineers must recognise and protect PII when collecting, processing, storing or sharing personal information.
Running potentially risky operations in an isolated environment separated from critical systems.
💡 Allowing a child to play in a fenced playground instead of an unrestricted construction site.
It helps prevent unsafe code execution or agent actions from damaging production systems or accessing unauthorised resources.
Giving a user, application or AI agent only the permissions necessary to perform its assigned task.
💡 Giving a delivery person access to the building entrance but not every room.
It reduces potential damage caused by mistakes, malicious instructions or compromised tools.
Verifying the identity of a user, application or service requesting access.
💡 Checking someone's identity card before allowing entry to an office.
It ensures only recognised identities can access protected AI applications and connected resources.
Restricting the number of requests an application or user can make within a specified period.
💡 A shop limiting how many items each customer can buy during a promotion.
It prevents abuse, protects infrastructure and controls unexpected API consumption and costs.
Requiring a human to review and approve an AI-proposed action before it is executed.
💡 A bank manager approving a large transaction before the bank processes it.
It is important for financial transactions, deleting data, external communications and other sensitive operations.
A chronological record of actions, decisions and changes performed by users or systems.
💡 A bank transaction history showing who transferred money, when and to whom.
It provides accountability, supports investigations and helps organisations meet regulatory and compliance requirements.
The policies, processes and controls that determine how AI is developed, deployed and used responsibly.
💡 A company's rulebook defining how employees must use company resources responsibly.
It helps organisations manage legal, ethical, security, privacy and operational risks associated with AI adoption.
The expense associated with running an AI model to process requests and generate responses.
💡 Paying for electricity every time a machine operates in a factory.
It influences the operating cost and profitability of AI products; developers need to understand model pricing, tokens and infrastructure.
Charging based on the number of input and output tokens processed by an AI provider.
💡 Paying for water based on litres consumed rather than a fixed monthly amount.
It helps estimate API expenses and understand how prompt length, response length and model selection affect costs.
The average expense incurred to process one request in an AI application.
💡 Calculating the average cost of preparing one meal in a restaurant.
It helps estimate customer-serving costs, compare model choices and establish sustainable product pricing.
Reducing AI operating expenses while maintaining the required quality and performance.
💡 Reducing a factory's production expenses without compromising product quality.
It matters as usage grows; techniques include model routing, prompt optimisation, caching and model selection.
Reducing the time an AI application takes to produce a response.
💡 Improving restaurant operations so customers receive food faster.
It improves user experience and helps applications support interactive and time-sensitive use cases.
Also known as: SLM
A relatively small language model designed to perform language-related tasks with fewer computational resources.
💡 Hiring a specialist for a small, well-defined job instead of an entire department.
SLMs can provide lower latency, reduced infrastructure costs and specialised capabilities for tasks that do not need a large model.
Improving the amount of work an AI system can process within a given time.
💡 Increasing the number of customers a restaurant serves each hour by improving kitchen operations.
It helps applications handle growing request volumes and simultaneous users without excessive infrastructure costs.
Also known as: TTFT
The time between sending a request to an AI model and receiving the first generated token.
💡 The time between placing a food order and receiving the first item.
It is important for streaming applications because users judge responsiveness by how quickly the first part of an answer appears.
Delivering an AI response gradually as it is generated instead of waiting for the entire response to finish.
💡 Watching a live cricket match rather than waiting for the complete recording.
It improves perceived responsiveness and lets users see answers appear progressively.
Reusing a previous response when a new request is sufficiently similar in meaning to an earlier request.
💡 A restaurant reusing a standard recipe for customers ordering the same popular dish.
It can reduce repeated model calls, lower API costs and improve response times; context, permissions and freshness need care.
Processing multiple requests or records together rather than handling each one independently.
💡 Washing a full load of clothes together instead of washing each item separately.
It can improve efficiency and reduce costs for large-scale, non-interactive tasks such as document analysis and classification.
Choosing the most suitable AI model for a particular task.
💡 Choosing the right vehicle for a journey instead of using a truck for every trip.
Models differ in reasoning ability, speed, cost, context capacity and modalities; selecting appropriately is an important architecture decision.
with tapaScript
By clicking Sign Up you re confirming that you agree with our Terms and Conditions.
