RAG, agents, and other buzzwords decoded
A field guide to the jargon. What each term means, when it matters, when someone's just flexing.
AI conversations move fast and borrow heavily from research papers, so the vocabulary piles up quickly. Some terms describe genuinely important ideas; others are used mainly to sound current. This guide defines the terms you will hear most, explains when each one actually matters, and flags the warning signs that a word is being used as decoration.
The foundations
Large language model (LLM)
A neural network, almost always a transformer, trained on very large amounts of text to predict the next token. Fine-tuned versions follow instructions and hold conversations. It matters because it is the engine underneath nearly every term below. It is flexing when every product with a text box is described as "built on cutting-edge LLM technology".
Token
The unit of text a model reads and writes - a word, part of a word or a punctuation mark. In English, one token averages roughly three-quarters of a word. It matters because models are priced, limited and measured in tokens.
Context window
The maximum amount of text, measured in tokens, a model can consider at once - your instructions, the conversation so far, any pasted documents and its own reply. Anything outside the window is invisible to the model. It matters when you work with long documents or long conversations. A bigger window is useful, but a model can still miss details buried in the middle of a very long input.
Inference
Running a trained model to get an output, as opposed to training it. Every chatbot reply is an inference call. It matters for cost and speed: training happens once, inference happens every time someone uses the product.
Making models know things
Embeddings
A way of turning text (or images, audio and so on) into a list of numbers that captures its meaning. Texts with similar meanings get similar numbers, even if they share no words - "car won't start" sits close to "vehicle battery problem". It matters because embeddings power semantic search, recommendations, clustering and RAG.
Vector database
A database optimised for storing embeddings and quickly finding the ones closest to a query. It matters when you need to search large collections by meaning. It is flexing when a few hundred documents - which a simple search or even a single prompt could handle - are presented as needing specialised infrastructure.
RAG (retrieval-augmented generation)
A pattern, not a product. Before the model answers, a retrieval step finds relevant passages from your own documents and inserts them into the prompt. The model then answers using that material, ideally citing it.
Retrieval-augmented generation. The model does not "learn" your documents; relevant excerpts are fetched and placed in its prompt at answer time.
It matters when answers must reflect private, recent or frequently changing information - company policies, product manuals, legal documents. It reduces, but does not eliminate, made-up answers. It is flexing when "RAG" is used to describe simply pasting a document into a chat.
Fine-tuning
Further training an existing model on your own examples to change its behaviour - its style, format or skill at a narrow task. It matters when you need consistent behaviour at scale that prompting cannot achieve. The common misconception: fine-tuning is usually a poor way to teach a model facts. For knowledge, RAG is typically cheaper, easier to update and easier to verify.
Prompting, RAG or fine-tuning? Start at the top and move down only when you have to.
Hallucination and grounding
A hallucination is a confident but false output - an invented statistic, quote or citation. Grounding means tying answers to verifiable sources, for example through RAG or web search with citations. Grounding lowers the hallucination rate; it does not make checking unnecessary.
Making models do things
Tool use (function calling)
The model is given descriptions of tools - search the web, query a database, send an email, run code - and can respond with a structured request to use one. Your software performs the action and returns the result to the model. It matters because it is how language models move from talking to acting.
MCP (Model Context Protocol)
An open standard, introduced by Anthropic in late 2024, for connecting AI applications to tools and data sources in a consistent way - often compared to a universal adapter. A tool exposed once through MCP can be used by any compatible AI app. It matters if you build integrations; most users will simply see more apps that "connect" to their AI assistant.
Agent
The most stretched word in AI. A reasonable working definition: a system where a model decides its own next steps toward a goal, using tools in a loop - plan, act, observe the result, adjust - rather than following a fixed script.
Autonomy is a spectrum. Many products marketed as "agents" sit on the left.
It matters for open-ended, multi-step tasks like research or software changes, where the path cannot be scripted in advance. It is flexing when a single prompt followed by a single API call is described as an "autonomous agent". A useful test: does the system ever choose to do something different based on what it observed?
Guardrails
Checks placed around a model to keep it within bounds: input filters, output validation, permission limits on tools, and human approval for risky actions. It matters more as systems gain the ability to act. Good guardrails are specific and testable; vague claims of "enterprise-grade safety" are not.
Dials and labels
Temperature: a setting controlling randomness. Lower values give more predictable output; higher values give more variety. Useful to know, rarely worth agonising over.
Multimodal: a model that handles more than one type of input or output - text plus images, audio or video.
RLHF: reinforcement learning from human feedback, a training stage where human preference ratings shape the model's answers to be more helpful and safer.
Open-weight model: a model whose trained weights are published so anyone can run or modify it. Not always the same as fully "open source", which would also include training data and code.
AGI: artificial general intelligence, a hypothetical system matching human ability across virtually all intellectual tasks. A research goal and a frequent marketing word; there is no agreed test for it.
How to spot buzzword overuse
The term replaces an explanation. If someone cannot say what the system does in one plain sentence, the jargon is doing the work.
The architecture is bigger than the problem. Vector databases and multi-agent systems for a task a well-written prompt could handle.
There is no evaluation. Claims of "accurate", "autonomous" or "safe" with no description of how they were tested.
Everything is an agent. Ask what the system decides on its own, and what happens when a step fails.
One question that cuts through: "What would go wrong if we used the simplest possible version instead?" A good answer reveals whether the complexity is earning its place.