AI Terms Explained: A Working Glossary

AI vocabulary spreads faster than its definitions. Terms arrive in product announcements, get repeated in headlines, and settle into general use before most people have been told what they mean. The result is a lot of conversations where everyone nods along at the word “agent” while privately picturing something different.

This glossary covers the terms that come up most often in practice. Each entry explains the concept, notes what it is commonly confused with, and says why it matters when you are comparing tools. The definitions are deliberately about mechanisms rather than specific products, because mechanisms stay put while products change every few months.

The foundations

Large language model (LLM)

A system trained on very large amounts of text to predict what comes next, which turns out to be enough to write, summarise, translate, answer questions and produce code. “Large” refers to the number of internal parameters — the adjustable numbers learned during training — which typically runs into the billions.

Commonly confused with: the chat product built on top of it. The model is the engine; the assistant you use is the engine plus a system prompt, tools, memory features, safety layers and an interface. Two products using the same underlying model can behave quite differently.

Token

The unit a model actually processes. Text is split into tokens before anything happens, each one roughly a short word or a fragment of a longer one. Usage limits, pricing and context sizes are all measured in tokens rather than words.

Why it matters: token counts are how you predict cost and whether your input will fit. A rough working figure for English prose is that a token averages somewhat under a word, so a few hundred words of text is a few hundred tokens plus a margin. Code, tables and non-English text run higher.

Context window

The maximum amount of text a model can consider at once, counting the system instructions, your documents, the conversation history and the response it is generating. It is working memory, not storage. Nothing persists between separate conversations unless a product explicitly saves it.

Commonly confused with: knowledge. Putting a document in context lets the model read it now; it does not teach the model anything lasting.

Parameters and weights

The numbers inside the model, adjusted during training, that encode everything it has learned. Parameter count gets quoted as a proxy for capability, but it is a weak one — training data quality, training method and post-training work matter at least as much. A well-trained smaller model routinely outperforms a poorly-trained larger one.

Ways of adapting a model

Prompting

Instructing the model in plain language. The cheapest and most reversible form of adaptation, and usually the first thing to try properly before reaching for anything heavier. Providing examples of the output you want inside the prompt is often called few-shot prompting.

RAG (retrieval-augmented generation)

A pattern where the system first searches a body of documents for passages relevant to your question, then places those passages into the model's context alongside the question. The model answers from the retrieved text rather than from memory.

This is the standard approach for building an assistant over your own material — internal documentation, a knowledge base, a policy library. Its advantages are concrete:

• Answers can cite the source passage, so they can be checked.
• Updating the knowledge means updating the documents, not retraining anything.
• Access controls can be applied at retrieval time, so users only see what they are entitled to.

Its main weakness is that a RAG system is only as good as its retrieval step. If the search fails to surface the right passage, the model answers from memory and the whole benefit evaporates. When evaluating a RAG-based tool, ask how retrieval quality is measured, not just which model it uses.

Fine-tuning

Further training of an existing model on a curated set of examples, adjusting its weights so it internalises a pattern. Good for teaching consistent format, tone or a specialised task style. Poor for teaching facts, which change and which retrieval handles better.

The practical rule: fine-tune for behaviour, retrieve for knowledge. Fine-tuning also carries real overhead — you need a decent dataset, an evaluation method, and a plan for redoing the work when you move to a newer base model.

What models can do and how they are wired up

Multimodal

A model that handles more than one type of input or output — typically text plus images, and increasingly audio and video. A multimodal model can read a screenshot, describe a photograph, or interpret a chart, because the non-text input is converted into the same internal representation as text.

Worth checking: which modalities are supported for input versus output. Many systems accept images but only produce text, which is a different capability from image generation.

Tool use / function calling

The mechanism that lets a model do something other than write. The application describes available functions; the model responds with a structured request to call one; the application executes it and returns the result into context. Search, calculation, database queries and sending messages all work this way.

This is the bridge between a model that talks about things and software that does things, and it is worth knowing because it determines what a tool can genuinely act on.

Agent

The loosest term in circulation. In its useful sense, an agent is a system where the model works toward a goal over multiple steps, deciding which tools to invoke, observing the results, and continuing until it is done or stuck. The key distinction from a chatbot is the loop: the model's own output determines the next action.

Because the word is applied to everything from genuine multi-step systems to a single scripted API call, treat it as a prompt for questions rather than a description:

• What actions can it actually take, and against which systems?
• Where does a human approve consequential steps?
• What happens when a step fails — does it retry, stop, or continue on bad data?
• Is there a log of what it did, and can the work be reversed?

Inference, latency and streaming

Inference is the act of running a model to produce output. Latency is the delay before you see results, usually split into time-to-first-token and generation speed thereafter. Streaming means tokens appear as they are produced rather than all at once — the same total time, but a much better experience for long responses.

Terms that describe limitations

Hallucination is fluent, confident output that is simply untrue, arising because the model optimises for plausible continuation rather than accuracy. Knowledge cutoff is the point where training data ends, after which the model knows nothing unless told. Temperature controls randomness in output selection; lower is more consistent, higher more varied. Grounding means tying a response to verifiable source material, usually via retrieval.

The practical takeaway: when you are assessing an AI tool, translate the marketing into these terms and the picture usually clarifies. Ask what model family sits underneath, whether it retrieves from your data or answers from memory, whether it can take actions or only produce text, and where a human sits in the loop. Those four answers tell you more about whether a tool will work for you than any benchmark score.