The jargon,
translated.
Every term you’ll actually meet, in plain English. No mathematics, no mysticism.
36 terms
A
Agent
An AI system that doesn’t just answer but acts: it works towards a goal in a loop of thinking, using tools and checking results.
Deeper: /explainers/agents
API
The route by which software (rather than a person in a chat window) talks to an AI model. Enterprise legal tools use models through APIs.
Attention
The mechanism by which each word looks at the other words and decides which matter to it — how “it” finds its “cat”.
Deeper: /explainers/transformers
C
Chatbot
A conversational interface to a model: you write, it answers, one turn at a time. Contrast with agent.
Compaction
When a long conversation is summarised to fit the context window, earlier detail is compressed — one reason long chats drift.
Deeper: /explainers/context
Context window
The fixed amount of text a model can see at once: your instructions, documents, the conversation and its own replies all share it.
Deeper: /explainers/context
Cutoff (knowledge cutoff)
The date the training corpus was frozen. The model’s world stops there unless you ground it with newer material.
Deeper: /explainers/training
E
Effort
A setting on modern models controlling how hard the model thinks before answering — low for triage, high for analysis.
Deeper: /explainers/models
Embedding
A word or document turned into a list of numbers so that similar meanings sit close together; how AI search finds “relevant” passages.
Deeper: /explainers/transformers
Enterprise endpoint
A contracted route to a model with commercial terms: data residency, retention limits, no training on your inputs. The difference between a consumer toy and a professional tool.
Deeper: /explainers/confidentiality
F
Fine-tuning
Additional training that teaches a model a format or behaviour using examples, after its general training.
Deeper: /explainers/training
Foundation model
A large general-purpose model trained on broad data, on which products are built.
G
Grounding
Giving the model your actual documents and instructing it to answer from them rather than from memory. The single biggest reliability upgrade.
Deeper: /explainers/grounding
H
Hallucination
A confident, fluent, invented output — a case that doesn’t exist, a quote that was never said. Structural, not a bug.
Deeper: /explainers/hallucination
I
Inference
Running a trained model to get an answer (as opposed to training it). What you pay for per use.
K
Knowledge base
A curated set of documents a tool searches and answers from, rather than relying on the model’s memory.
Deeper: /explainers/grounding
L
Large language model (LLM)
A model trained on vast amounts of text to predict the next word — the technology behind every current AI assistant.
Deeper: /explainers/transformers
Local model
A model that runs entirely on your own machine: nothing leaves the building, at the cost of size and speed.
Deeper: /explainers/confidentiality
M
Multimodal
A model that handles more than text: images, audio, documents-as-scans.
N
Neural network
The underlying structure of a model: layers of simple calculations whose combined behaviour is learned, not programmed.
P
Parameter
One of the learned numbers inside a model; counted in billions. More parameters ≈ more capacity, not automatically more accuracy.
Pretraining
The first, longest stage of training: reading a vast corpus predicting the next word. Where knowledge and the cutoff date come from.
Deeper: /explainers/training
Prompt
Everything you give the model: instructions, context, documents, question. The quality of the answer is set here.
Deeper: /explainers/prompting
Prompt injection
An attack where text inside a document the model reads contains instructions the model then follows. A reason to be careful what tools may act on what content.
R
RAG (retrieval-augmented generation)
Grounding at scale: a search step finds relevant passages in your documents and inserts them into the context automatically.
Deeper: /explainers/grounding
Red-teaming
Deliberately attacking your own work — asking the model to find the weaknesses in a draft, or testing a system for failure modes before someone else does.
RLHF (reinforcement learning from human feedback)
The final stage of training, where humans rank answers and the model is tuned towards the preferred ones. Source of helpfulness — and agreeableness.
Deeper: /explainers/training
S
Skill
A packaged set of instructions and reference material a model loads for a specific task — reusable expertise for the tools you use.
System prompt
Standing instructions the model receives before your message: role, rules, tone. In products, mostly invisible to you.
T
Temperature
A randomness dial on some models: low for consistency, higher for variety. Many current models manage this themselves.
Token
The unit models actually read — a short chunk of characters, roughly three-quarters of a word in English. Context windows and prices are measured in them.
Deeper: /explainers/transformers
Training data
The text a model learned from. Sets what it knows, when its knowledge stops, and what biases it inherits.
Deeper: /explainers/training
Transformer
The architecture behind modern AI models: attention plus feed-forward layers, stacked.
Deeper: /explainers/transformers
V
Vibe coding
Building small software tools by describing what you want and letting the model write the code. Powerful for utilities; risky for anything that matters without review.
Z
Zero data retention
A contractual term: the provider keeps your inputs and outputs for zero days. What to demand for client material.
Deeper: /explainers/confidentiality
Zero-shot
Asking a model to do a task with no examples — relying on instructions alone. Contrast with showing it worked examples, which usually performs better.
Deeper: /explainers/prompting
A term we’ve missed?
Tell us and we’ll add it — the glossary grows with the technology.
Suggest a term