Reference

The AI glossary, in plain English.

32 terms that actually come up when you are building with AI — what each one means, why it matters, and the mistake almost everyone makes with it. No prior background assumed.

Last reviewed 3 September 2026

Every term in this glossary

32 terms, grouped below by what they are about. Jump straight to one, or read the section that covers the part of AI you are working on.

Foundations

Large language model (LLM): A large language model is a neural network trained on enormous amounts of text to predict what comes next. Doing that well enough, at enough scale, produces something that can write, summarise, translate, answer questions and follow instructions.
Token: A token is the unit a language model actually reads and writes — usually a common word, a word fragment, or a piece of punctuation. As a rough rule for English, one token is about four characters, and 100 tokens is about 75 words.
Context window: The context window is the maximum amount of text, measured in tokens, that a model can take into account for a single response. Everything the model knows about your specific problem has to fit inside it — the system prompt, the conversation so far, any documents you pasted, and the answer it is about to write.
Hallucination: A hallucination is output that is fluent, confident and wrong — an invented citation, a function that does not exist, a statistic with no source. It is not a malfunction bolted onto an otherwise reliable system; it is the same next-token prediction that produces correct answers, applied where the model has no grounding.
Multimodal AI: A multimodal model handles more than one kind of input or output — text plus images, audio, video or documents. In practice it means you can hand a model a screenshot, a scanned invoice, a whiteboard photo or a recording and ask questions about it in ordinary language.
Transformer: A transformer is the neural network architecture behind almost every modern language model. Its defining trick is attention: when processing any one word, it can look at every other word in the input at once and weigh which ones matter.
Neural network: A neural network is a stack of simple mathematical functions with adjustable numbers between them, arranged in layers. Training repeatedly adjusts those numbers until the whole stack turns inputs into the outputs you wanted.
Temperature: Temperature is a setting that controls how much randomness goes into choosing each token. Low values make a model pick the most likely option almost every time; high values let it choose less likely ones more often.
RLHF: RLHF is the training step that turns a raw next-token predictor into something that follows instructions. People rank competing answers, those rankings train a reward model, and the language model is then tuned to score well against it.

Building with AI

Prompt engineering: Prompt engineering is the practice of designing what a model sees — the instruction, the context, the examples and the required output format — so that it produces the result you want reliably rather than occasionally.
AI agent: An AI agent is a language model given a goal and a set of tools, left to decide its own next step. Instead of answering once, it works in a loop: choose an action, call a tool, read the result, decide what to do next, and stop when the goal is met.
Fine-tuning: Fine-tuning continues training an existing model on your own examples so that its default behaviour shifts towards them. It is how you teach a model a house style, a rigid output format or the conventions of a specialist domain.
Model Context Protocol (MCP): The Model Context Protocol is an open standard for connecting AI applications to external tools and data sources. It defines one way for an assistant to discover what a system can do and to call it, so an integration written once works with any client that speaks the protocol.
System prompt: A system prompt is a standing instruction placed before the conversation that tells a model who it is, what it should do and what it must not do. It applies to every turn rather than to one message.
Chain of thought: Chain of thought is prompting a model to work through a problem in steps before giving its answer. Producing the intermediate reasoning as text measurably improves accuracy on problems that need more than one step.
Function calling: Function calling lets a model request that your code run a specific function with specific arguments. The model never executes anything itself — it returns a structured request, your code decides whether to run it, and the result goes back into the conversation.
Few-shot prompting: Few-shot prompting means including a handful of worked examples in your prompt so the model can infer the pattern. Zero-shot is the same request with no examples, relying on instructions alone.
Structured output: Structured output constrains a model to return data in a shape you specified — usually JSON matching a schema — instead of prose. It turns "parse the answer and hope" into a contract the model cannot syntactically break.

Data and retrieval

Embedding: An embedding is a list of numbers — typically a few hundred to a few thousand of them — that represents a piece of text, an image or an audio clip as a point in space, positioned so that things with similar meaning end up close together.
Vector database: A vector database stores embeddings and answers one question very quickly: which stored items are nearest to this one? It is the retrieval layer under semantic search and most RAG systems.
Retrieval-augmented generation (RAG): Retrieval-augmented generation means finding the documents relevant to a question and putting them in front of the model, so it answers from your content rather than from memory. It is the standard way to make a model reliable about facts it was never trained on.
Chunking: Chunking is splitting documents into smaller passages before turning them into embeddings and indexing them. The chunk is the unit that gets retrieved, so how you split decides what the model can be given.
Semantic search: Semantic search finds results by meaning rather than by matching words. Query and documents are turned into embeddings, and the closest vectors are returned — so "how do I cancel" can find a passage about ending a subscription.
Hybrid search: Hybrid search runs a keyword search and a semantic search over the same corpus and merges the two result lists into one ranking. It exists because the two methods fail in opposite directions, so the union is more reliable than either half.
Reranking: Reranking is a second pass over search results that reorders them by how well each one actually answers the query. Retrieval fetches a wide set cheaply; the reranker reads each candidate against the query properly and puts the best ones first.

Running AI in production

Inference: Inference is the act of running a trained model to produce an output — the part that happens every time a user asks something, as opposed to training, which happens once. In almost any product that ships, inference is where the money goes.
Evals: Evals are a repeatable set of test cases, with a way of scoring the output, that tell you whether a change to your prompt, model or retrieval made the system better or worse. They are to AI features what a test suite is to ordinary code.
Guardrails: Guardrails are the checks placed around a model — before the request, after the response, and around any tool it can call — that keep a system safe, in scope and predictable when the model does something unexpected.
Prompt injection: Prompt injection is what happens when content a model reads — a web page, an email, a document, a tool result — is treated as instructions rather than as data. The model has no reliable way to tell the two apart, because to it both arrive as text in the same context window.
Quantization: Quantization stores a model at lower numerical precision than it was trained at, so each weight takes fewer bits. The model gets substantially smaller and faster, with some loss of quality.
Jailbreak: A jailbreak is a prompt crafted to make a model ignore its own safety training and produce something it was tuned to refuse. The user is the attacker, which is what separates it from prompt injection, where the attacker is content the model reads.
Distillation: Distillation trains a small model to reproduce the behaviour of a large one. The large model generates the training signal, the small model learns to match it, and the result is a cheaper model that keeps much of the original quality on the tasks it was distilled for.

Terms people mix up

Most confusion about AI is between two terms rather than inside one. These are the six pairs that get swapped most often.

Pairs of AI terms that are commonly confused, what each one actually means, and how the two differ
Term	What it actually is	Not the same as
Embedding	A piece of text turned into a list of numbers that captures what it means.	Vector database — the store those numbers live in, and what you search to find them.
Retrieval-augmented generation (RAG)	Fetching relevant text at answer time and handing it to the model with the question.	Fine-tuning — changing the model itself by training it further on your data.
Semantic search	Matching on meaning, so a search for "car" also finds "automobile".	Hybrid search — running that alongside plain keyword matching and merging both sets of results.
Prompt injection	Instructions hidden inside content the model reads, aimed at hijacking the system.	Jailbreak — the user themselves talking the model out of its own rules.
Evals	Measuring whether the thing actually works, before you ship it and after.	Guardrails — constraining what it is allowed to do while it is running.
Distillation	Training a smaller model to imitate a larger one.	Quantization — shrinking the numbers inside a model you already have.

Which AI terms should I learn first?

Start with model, token and context window, because almost every other term is defined using them. Add hallucination and prompt engineering next, since they explain why outputs go wrong and what you can do about it. Retrieval and embeddings come after that, when you start feeding a model your own material.

What is the difference between an AI model and an AI agent?

A model answers: you give it text, it gives text back, and it does nothing else. An agent uses a model in a loop, deciding what to do next and calling tools to do it — searching, writing files, calling an API. The model is the engine; the agent is what somebody built around it.

How is this glossary different from a dictionary definition?

Every entry answers the same three questions: how it works, why it matters, and what people get wrong about it. That last one is the useful part — most confusion about AI comes from a term that was half-understood rather than never encountered, and a one-line definition will not fix that.

Missing a term you had to look up?

The glossary grows from the questions members actually ask. Tell us what to add and it goes on the list.