// GLOSSARY · 12 TERMS · NO OBLIGATION

The words, in plain English

Every one of these gets used to sell you something. None of them is complicated once the jargon is stripped out. Read it, use it, take it to a different consultant if you like — that is the point.

LLM (Large Language Model)

The engine behind ChatGPT and its competitors — a program trained on enormous amounts of text that predicts what words come next. Useful because that turns out to cover writing, summarising, classifying and answering questions.

On-premises / on-prem

Running software on hardware you own, inside your building, instead of renting it from a vendor over the internet. For AI it means your documents are never transmitted to anyone else.

RAG (Retrieval-Augmented Generation)

Giving an AI model access to your own documents at the moment it answers, so it responds from your files rather than from general training. This is how you get a system that knows your policies, contracts and part numbers.

Inference

The act of the model actually answering — as opposed to training it. Every question you ask is one inference, and inference is what your hardware has to be sized for.

Fine-tuning

Adjusting a model on examples specific to your business so it adopts your format, terminology and tone. More expensive than RAG and usually not the first thing to reach for.

Hallucination

When a model states something false with complete confidence. It is a property of how these systems work, not a bug that gets patched, which is why anything consequential still needs a human check.

Token

The unit these models actually read and write — roughly three-quarters of a word. Pricing and size limits are quoted in tokens, which is why they appear on invoices.

Context window

How much text a model can consider at once. Exceed it and the earliest material silently drops out — the usual explanation for a system that seems to forget the start of a long document.

Embedding

A numerical fingerprint of a piece of text that lets a computer find related passages by meaning rather than by exact keyword. The mechanism underneath most document search built in the last few years.

Vector database

Where those fingerprints are stored so they can be searched quickly. A necessary component of RAG, and one more thing that can live entirely on your own hardware.

Agent

A model wired up to take actions — send the email, update the record, place the order — rather than only produce text. Powerful and correspondingly worth constraining carefully.

Sovereign AI

AI capability you control end to end: your hardware, your data, no external service that can change terms, raise prices or read your files.

// θ → 2π · NEXT

Still want a translator in the room?

Knowing the vocabulary is not the same as knowing which of it applies to your operation. That is what the sweep answers.