Home › Glossary

AI glossary

Plain-language definitions of 55 terms you will meet when choosing and using AI tools.

Agent (AI agent): A program that uses a language model to decide on steps, call tools or APIs, and work toward a goal with limited step-by-step instruction from a person.
Agentic workflow: A process in which a model plans, takes actions, checks results and repeats, instead of answering a single prompt once.
API (application programming interface): A defined way for one piece of software to request data or actions from another. Model providers expose APIs so developers can send prompts and receive responses.
Benchmark: A standard test set used to compare models on a specific skill, such as coding, math or reading comprehension. Scores depend on the test and how it is run.
Chain of thought: A prompting style or model behavior in which the model writes out intermediate reasoning steps before giving a final answer.
Chatbot: A program that holds a conversation with a user in text or voice. Modern chatbots are usually built on large language models.
Context window: The maximum amount of text, measured in tokens, that a model can consider at one time, including the prompt and its own reply.
Copilot: A common product name for an assistant built into another tool, such as an editor or office suite, that suggests or completes work alongside the user.
Deep learning: A branch of machine learning that trains multi-layer neural networks on large amounts of data.
Diffusion model: A type of model, widely used for images and video, that learns to turn random noise into a coherent picture step by step, guided by a text prompt.
Embedding: A list of numbers that represents the meaning of a piece of text, image or other data so that similar items sit close together and can be searched by similarity.
Evaluation (evals): Tests that measure how well a model or an AI application performs on tasks that matter to its builder, often run automatically before each release.
Few-shot prompting: Including a small number of worked examples in a prompt so the model can follow the same pattern.
Fine-tuning: Continuing to train an existing model on a smaller, specialized dataset so it performs better on a particular task or style.
Foundation model: A large model trained on broad data that can be adapted to many tasks, such as a general-purpose language or image model.
Function calling (tool use): A feature that lets a model return a structured request to run a named function or tool, which the application then executes and reports back.
Generative AI: Models that create new content, such as text, images, audio, video or code, based on patterns learned from training data.
GPU: A graphics processing unit, a chip that performs many calculations in parallel. GPUs are the main hardware used to train and run large models.
Grounding: Supplying a model with trusted source material, such as documents or search results, so its answer is based on that material.
Guardrails: Rules, filters or checks placed around a model to limit unwanted inputs or outputs, such as unsafe content or off-topic answers.
Hallucination: When a model states something that sounds plausible but is false or unsupported by its sources.
Inference: Running a trained model to produce an output. Inference cost and speed are key factors when choosing a model provider.
Jailbreak: A prompt crafted to make a model ignore its safety rules or intended instructions.
Large language model (LLM): A neural network trained on large amounts of text to predict and generate language. LLMs power most chatbots and writing assistants.
Latency: The delay between sending a request and receiving a response. Voice and real-time applications are especially sensitive to it.
LoRA: Low-Rank Adaptation, a technique that fine-tunes a model by training a small set of added weights instead of the whole model, which lowers cost.
Machine learning: Building software that improves at a task by learning patterns from data instead of following only hand-written rules.
MCP (Model Context Protocol): An open protocol for connecting AI applications to external tools and data sources through a common interface, so one integration can work across different clients.
Mixture of experts (MoE): A model design in which only some specialized sub-networks are used for each input, which can reduce the computing needed per response.
Model: The trained system, made of numerical weights, that takes an input and produces an output such as text, an image or a prediction.
Multimodal: Able to work with more than one kind of input or output, such as text, images, audio and video.
Neural network: A computing system made of layers of connected units whose strengths are adjusted during training.
Observability: Tools and practices for recording and inspecting what an AI application did, including prompts, responses, costs and errors, so problems can be traced.
Open source: Software released under a license that lets others view, use, change and share the source code, within the license terms.
Open-weight model: A model whose trained weights are published so others can download and run it. Licenses vary and may not be open source in the strict sense.
Orchestration: Coordinating the steps of an AI application, such as calling models, tools and databases in order, often with a framework.
Parameters: The adjustable numerical values inside a model that are set during training. Parameter count is one rough indicator of model size.
Prompt: The input given to a model, which can include instructions, questions, examples and background material.
Prompt engineering: Writing and refining prompts to get more useful, reliable output from a model.
Prompt injection: An attack in which text in a document, web page or message tries to override the instructions given to an AI application.
RAG (retrieval-augmented generation): A pattern in which an application first retrieves relevant documents, often from a vector database, and gives them to the model so its answer draws on that material.
Reasoning model: A model trained or configured to spend extra computation working through a problem in steps before it answers.
Reinforcement learning from human feedback (RLHF): A training method that uses people's ratings of model outputs to steer the model toward preferred behavior.
Sandbox: An isolated environment where code or an agent can run without being able to affect the rest of a system.
Semantic search: Search that matches the meaning of a query rather than only its exact words, usually by comparing embeddings.
Speech-to-text (STT): Software that converts spoken audio into written text, also called transcription or automatic speech recognition.
System prompt: Instructions given to a model by the application developer, separate from the user's message, that set its role and rules.
Temperature: A setting that controls how varied a model's output is. Lower values give more predictable text and higher values give more varied text.
Text-to-speech (TTS): Software that turns written text into spoken audio, often with a choice of voices.
Token: A small unit of text, such as a word or part of a word, that a model reads and writes. Many APIs bill by the number of tokens.
Training data: The examples a model learns from. Its content and quality strongly affect what the model can do.
Transformer: A neural network design built around attention, which lets a model weigh the relationship between all parts of its input. Most modern language models use it.
Vector database: A database designed to store embeddings and quickly find the ones most similar to a query.
Voice cloning: Creating a synthetic voice that imitates a particular person's voice from recordings, which raises consent and misuse concerns.
Zero-shot prompting: Asking a model to perform a task with instructions only, without giving examples.