The language around private AI is full of terms that sound harder than they are. This page explains the ones you are likely to meet — in a meeting with us, in a vendor's brochure, or in a conversation with your IT team.
Nothing here is specific to our product. It is the vocabulary of the field.
What the technology actually is, underneath the marketing. Five terms that make everything below make sense.
Software that works out its own rules from examples, rather than being given the rules by a programmer.
Show it ten thousand invoices and it learns what an invoice looks like. Nobody sits down and writes a description of one.
The structure most machine learning uses: a very large number of simple units, arranged in layers, each passing numbers to the next.
The name comes from a loose analogy with brain cells. The resemblance is thin — this is arithmetic, not biology, and no part of it thinks.
The particular design of neural network behind every modern language model. Published in 2017, and the reason the field moved as quickly as it did afterwards.
Its central idea is attention: when the model handles one word, it works out how much every other word in the passage bears on it. Earlier designs read strictly left to right and lost the thread over long passages.
Generative Pre-trained Transformer. Three plain claims: generative, it produces text rather than choosing from options; pre-trained, the expensive learning was done in advance; transformer, the design above.
How a model is wired: how many layers, how wide they are, how they connect. The blueprint, as distinct from the weights that fill it.
The handful of words that everything else is built on.
The AI itself — the part that reads and writes. ChatGPT, Claude, Llama and Gemma are all large language models.
A model is a very large file of numbers, produced by reading an enormous amount of text. It has no memory of you and no connection to your files unless it is given them.
Finding the relevant pages in your own documents first, then asking the model to answer using only those pages.
Instead of asking someone to answer from memory, you hand them the right three pages and say: answer from these, and tell me which page you used.
An AI service running on someone else's computers, reached over the internet. Your question, and any document you attach, leaves your building to be answered.
The hardware physically lives in your building, on your network, rather than in a data centre belonging to someone else.
Models whose files are published for anyone to download and run on their own hardware. Llama, Gemma, Qwen and Nemotron are examples.
"Open-weight" and "open source" are often used interchangeably, but they are not the same. Open-weight means the finished numbers are published. Genuinely open source would also mean publishing the training data and the code that produced them, which is rarer. The distinction matters less for using a model than for scrutinising how it was built.
Using a finished model to answer a question. The opposite of training, which is the process of building the model in the first place.
Inference is asking the expert a question. Training is putting them through university.
You will meet the word constantly, because it is the only thing most AI systems actually do. An inference server or inference engine is simply the software that runs the model and takes questions.
Taking a finished model and nudging it to be better at one narrow thing, using additional examples. Much cheaper than training from scratch, much more involved than retrieval.
The instruction given to the model — the question, plus whatever context is supplied alongside it.
A standing instruction given to the model before any conversation starts — its role, its tone, what it must do and must not do. Users never see it.
The unit a model reads and writes in. Not a letter and not quite a word — a chunk of text somewhere between the two.
Common words are usually a single token each. "The cat sat on the mat" is six words and six tokens. Longer or less common words get broken into pieces: "unbelievable" becomes something like un · believ · able, three tokens for one word. Unusual names, long numbers and technical terms split the most.
Average that across ordinary English and you land at roughly four tokens for every three words. That is where "three-quarters of a word" comes from — it is an average across a lot of text, not a rule about any particular word.
The material a model was built from — very large quantities of text, gathered before the model was released.
The model itself: one very large file of numbers. Running a model means loading that file into memory.
The name is literal. It is the same "weight" as in a weighted average — when you weigh up a decision, you decide how much each factor counts. Inside a model are simple units, each receiving several inputs, and each input arrives with a number attached saying how much it counts. That number is its weight. The term comes from the earliest neural networks of the 1940s and 50s, where the weights were physical dials controlling how strongly a signal passed through.
Training is nothing more than adjusting those numbers. You start with random values, so the model produces nonsense; you show it an example, measure how wrong it was, and nudge every weight slightly in the direction that would have been less wrong. Then repeat, an astronomical number of times. Nobody writes the weights, and nobody can read an individual one and say what it does.
A model has two parts: the architecture — how the units are wired, short enough to write on a page — and the weights. The architecture is the blueprint. The weights are the building.
The date after which a model knows nothing, because its training finished. Ask about last month's events and it will have no idea.
The vocabulary that decides whether a tool is usable at all in regulated work.
The protection that keeps communications between a lawyer and their client confidential and outside the reach of disclosure.
Under data protection law, the controller decides why and how personal data is used. A processor acts on the controller's instructions. The two carry different obligations.
The United Kingdom's data protection regime, retained and amended from the EU General Data Protection Regulation, sitting alongside the Data Protection Act 2018.
An assessment organisations carry out before deploying technology likely to present a high risk to people's rights — identifying what is processed, why, what could go wrong, and what reduces that risk.
Two rights individuals hold over their personal data: to obtain a copy of what an organisation holds, and in defined circumstances to have it deleted.
The documented history of a piece of evidence — who held it, when, and what was done to it — sufficient to show it has not been altered.
The stage of legal proceedings in which parties exchange the documents relevant to a case.
European Union legislation regulating AI by risk category, with heavier obligations on higher-risk uses. It applies to organisations operating in the EU; the UK has taken a different, sector-led approach.
These entries explain terms you may encounter. They are descriptions, not legal advice — your own advisers and data protection officer should determine what applies to your organisation.
What the system is actually pointed at, and the state it tends to arrive in.
The body of documents a system searches. Your matter files, case files, archives, or whichever subset you choose to include.
Reading text off a scanned image so it becomes searchable. A scanned page is a photograph until something converts it.
The system a firm already keeps its documents in — iManage, NetDocuments, SharePoint, or a structured set of network folders.
Information about a document rather than in it — author, date, matter number, document type, who last edited it.
Permanently removing or obscuring specific material in a document before it is shared — names, addresses, privileged passages.
Turning recorded speech into text — interviews, recorded calls, meetings — so it can be searched and quoted like any other document.
A system that handles more than text — images, audio, and video alongside documents.
What actually happens between your question and the reply.
The one-off process of reading your documents and organising them so they can be searched by meaning. Done once at the start, then topped up as new documents arrive.
Splitting documents into passages small enough to be retrieved on their own. A five-hundred-page bundle is stored as thousands of pieces, each tagged with where it came from.
Turning a passage of text into a list of numbers that captures its meaning, so the system can find passages that mean something similar rather than passages that happen to use the same words.
Searching by meaning rather than by matching words. The opposite of keyword search, which finds only the exact terms you typed.
The reference attached to an answer, pointing at the exact source — document and page — that the statement came from.
Restricting the model to answering from supplied documents rather than from whatever it absorbed during training.
When a model states something untrue with complete confidence. It is not lying — it is producing text that fits the pattern of a plausible answer.
Rules placed around a model to constrain what it will do — topics it declines, actions it cannot take, information it must not repeat.
How much text the model can hold in mind at once, measured in tokens.
How many pages you can keep open on the desk before you have to close one.
An AI that works in several steps on its own — look something up, decide what to do next, check another source, then answer — rather than replying in one pass.
A model that works a problem through in steps before answering, instead of replying immediately. The working is generated internally and usually kept out of sight.
Standard tests used to compare models against one another, usually reported as a score out of a hundred.
The hardware vocabulary — most useful if you have an IT team who will ask.
A graphics card. Built originally to draw video games, but the same maths — thousands of small sums at once — turned out to be exactly what AI needs.
An ordinary processor is one very capable person working through problems in order. A graphics card is thousands of simple calculators working at the same time. AI is thousands of small sums, so the calculators win.
The graphics card's own memory, separate from the computer's ordinary memory and far faster.
Think of it as the desk the system works at. The whole model has to fit on the desk at once. If it does not fit, it does not run slowly — it does not run.
How quickly the graphics card can move data in and out of its own memory, measured in gigabytes per second.
If memory capacity is the size of the desk, bandwidth is how fast your hands move across it. A large slow desk is worse than a smaller fast one.
The number of adjustable values inside a model. "70B" means seventy billion of them. More generally means more capable, and always means a larger file.
In conversation, parameters and weights mean the same thing. Strictly, parameters include a few other values as well, but the two words are used interchangeably and both point at the same idea: the numbers that make up the model.
Storing a model's numbers less precisely so the whole thing takes less space.
Saving a photograph at a smaller file size. A little quality is lost, a lot of space is saved, and in normal use you cannot tell.
Memory that checks itself for errors. All memory occasionally suffers a flipped bit — from heat, electrical interference, even background radiation. ECC memory detects it and corrects it. Ordinary memory does not notice.
Parts of the graphics chip built for one job: multiplying grids of numbers together. That happens to be nearly everything an AI model does.
The software layer that lets programs use NVIDIA graphics cards, and the name for the thousands of small processing units inside them.
The model file is not the only thing occupying memory. Every question in progress needs scratch space, and the system itself takes a share.
Measures of raw calculating speed — trillions of operations per second.
The practical questions a facilities or IT manager will raise.
A computer with no monitor, keyboard or mouse. It is administered over the network and used through a web browser on people's ordinary machines.
A machine with no connection to the internet at all.
Where your data physically sits, and which country's law therefore applies to it.
How much electricity the machine uses, measured in watts — and therefore how much heat it puts into the room.
A battery that keeps the machine running through a power cut for long enough to shut down cleanly.
Rules determining which staff can see which documents, and therefore which documents can inform their answers.
A record of who asked what, and when.
Backup is a copy kept elsewhere in case something is lost. Redundancy is duplication inside the machine — typically mirrored drives — so a single failure does not stop it working.
Administering, updating and diagnosing the machine over a network connection rather than standing in front of it.
Two physical formats. A rackmount machine is built to bolt into a standard server rack; a tower stands on the floor like a desktop computer.
Security certifications. ISO 27001 is an international standard for managing information security; Cyber Essentials is a UK government-backed scheme covering baseline controls.
Terms that come up when comparing options, whoever you buy from.
The position of being unable to leave a supplier without losing something you depend on — your data, your workflow, or the ability to operate at all.
Charging for each named user, every month, for as long as they have access.
A limited trial on real material, run before committing, to establish whether the thing works on your documents rather than in a demonstration.
The written commitment on support — how quickly a supplier responds, during which hours, and what happens when something fails.
No terms match that search. Try a shorter word.
That is the normal starting position, and it is what a discovery visit is for. We come to you, look at the documents and the work they support, and tell you honestly whether private AI is worth it for your organisation — including when it is not.
Book a discovery visit