AI Terms, Defined
The vocabulary you need to read a vendor deck without being misled, grouped by what each term is actually about. Definitions only — no hype.
BeginnerVerdeshell Team · 8 min read · Last reviewed
Most confusion about AI is vocabulary confusion. These are the terms that carry real distinctions, defined so the distinctions survive.
Key takeaways
- Model-mechanics terms — tokens, context window, temperature — explain cost, limits and variability.
- Prompting terms — system prompt, few-shot, chain-of-thought — explain how to steer a model.
- Retrieval terms — RAG, embeddings, chunking — explain how models use your own documents.
- Risk terms — hallucination, prompt injection, guardrails — explain what can go wrong.
How to use this
Grouped by what the term is about rather than alphabetically, because the groupings carry information: a term about model mechanics answers a different kind of question than a term about risk.
Where a definition has a caveat that matters commercially, the caveat is in the definition rather than a footnote.
Foundations
- Artificial intelligence (AI)
- Systems performing tasks that normally require human cognition — reasoning, perception, decision-making. Includes rule-based systems that do not learn.
- Machine learning (ML)
- Systems that learn patterns from data rather than following rules written by a person.
- Deep learning
- Machine learning using neural networks with many layers, which learn hierarchical representations of the input.
- Generative AI
- Models that produce new content — text, images, audio, code — rather than classifying or scoring existing content. Today almost always built on deep learning.
- Large language model (LLM)
- A model pre-trained on very large text corpora to predict the next token, then further trained to follow instructions. Usable for summarising, drafting, extraction and code.
Model mechanics
- Token
- The unit a model reads and writes — roughly a word fragment. Cost and context limits are counted in tokens, not words.
- Context window
- The maximum amount of text a model can consider in a single request, including both the input and its own output.
- Embedding
- A numeric vector representing the meaning of a piece of text, so that similar meanings sit close together and can be searched by proximity.
- Temperature
- A parameter controlling output randomness. Higher is more varied; lower is more repeatable. It does not control accuracy.
- Inference
- Running a trained model to get an output. Distinct from training, and the part you pay for per request.
- Transformer
- The neural network architecture behind nearly all current language models, introduced in 2017. Built on attention rather than reading text in order, which lets it train in parallel at scale.
- Attention
- The mechanism by which each token weighs every other token in the input to decide what is relevant to it — how a model links a pronoun to the noun it refers to.
- Pre-training
- The first and largest training stage: learning to predict the next token across a very large body of text. Produces a base model that continues text but does not reliably follow instructions.
- Instruction tuning
- Further training on examples of instructions paired with good responses, so that following instructions becomes the model’s default behaviour.
- RLHF
- Reinforcement learning from human feedback: tuning a model using human judgements of which of two answers is better. The step that makes assistants helpful — and shapes what they refuse.
- Reasoning model
- A model trained to work through a problem in intermediate steps before answering. Better at maths, code and multi-step logic; slower and more expensive per answer.
- Open-weight model
- A model whose trained parameters are published, so it can be run and fine-tuned on your own infrastructure. Not necessarily open source in the full sense — check the licence.
- Fine-tuning
- Further training of an existing model on task-specific data to adapt its behaviour.
Prompting
- System prompt
- Instructions set for the whole conversation — the model’s role, rules and output conventions — separate from the user’s individual messages.
- Few-shot prompting
- Including a small number of worked examples in the prompt so the model infers the pattern. Zero-shot means no examples.
- Chain-of-thought
- Prompting a model to work through intermediate steps before its final answer. Helps standard models on multi-step tasks; reasoning models do it on their own.
- Context engineering
- Curating everything that goes into a model’s context window at each step — instructions, documents, tool results, history — and what is left out. The successor framing to prompt engineering for agents.
Retrieval and tools
- RAG (retrieval-augmented generation)
- Retrieving relevant documents at query time and supplying them to the model, so it answers from your material rather than training data alone.
- Vector database
- Storage optimised for finding embeddings nearest to a query embedding — the retrieval half of RAG.
- Chunking
- Splitting source documents into passages small enough to retrieve and fit in context. Chunk boundaries strongly affect retrieval quality.
- Tool use / function calling
- Letting a model invoke defined functions or APIs, so it can act on systems rather than only describe actions.
- MCP (Model Context Protocol)
- An open standard for connecting models to external tools and data through one consistent interface instead of bespoke per-pair integrations.
- Grounding
- Whether an answer is actually supported by the supplied source material. Retrieval makes grounding possible; it does not guarantee it.
Agents
- AI agent
- A system given a goal rather than an instruction, which selects and takes its own actions toward that goal and adjusts based on results.
- Agentic
- An adjective for how much a system decides for itself. A spectrum, not a category.
- ReAct
- A pattern where a model alternates between reasoning about a problem and acting in an environment, using each result to inform the next step.
- Agent loop
- The cycle of act, observe, reason and retry that an agent repeats until a testable stop condition is met.
- Stop condition
- The testable state that ends a loop — for example, the test suite passing. Without one, an agent runs forever or stops too early.
- Multi-agent system
- Several specialized agents coordinating, each owning one role, with defined handoffs and paths for failure.
Behaviour and risk
- Hallucination
- Fluent, confident output that is false or unsupported by the source material. A property of how generative models work, not a bug that gets patched.
- Prompt injection
- Content the model reads — a web page, a document, a tool result — containing instructions that hijack its behaviour. The main security concern for tool-using systems.
- Guardrails
- Constraints applied around a model — input filters, output validation, permission limits — rather than inside it.
- Evaluation (evals)
- Systematically measuring output quality against defined criteria. The substitute for the accuracy metric a classifier gives you for free.
- Human in the loop
- A design where a person reviews or approves before an action takes effect. The usual first bound on delegated authority.
Want this built properly?
We design and build AI systems for clients. Tell us the problem and we will tell you honestly whether AI — and which kind — is the right fit for it.