Generative AI glossary

A large language model is a system trained on vast amounts of text to predict the next stretch of text, and which derives answers, summaries and translations from that single capability.

The model's knowledge is fixed when training ends and lives in its parameters: it does not include your organisation's documents and does not update itself.

A token is the smallest unit a language model breaks text into: a short word, a piece of a longer one, or a punctuation mark.

It is also the unit of cost: commercial model pricing is quoted per input and output token, not per question.

The context window is the maximum amount of text a model can consider in a single interaction, counting the question, the documents supplied and the answer together.

It is a design constraint, not a detail: it determines how many documents can be placed in front of the model at once, and therefore how they must be selected.

An embedding is the numerical representation of a piece of text, built so that texts with similar meaning end up close to one another.

It is what lets a search for "early termination" retrieve a paragraph about "withdrawal": the closeness is of meaning, not of words.

A hallucination is an answer that is plausible in form and false in content, produced by a model that generates the most probable text even when it has no source.

It is not an unpredictable defect to be tolerated: it follows from how the model works, and it is reduced architecturally by grounding answers in documents and making it explicit when a source is missing.

Prompt engineering is the design of the instructions that govern a model's behaviour: role, constraints, answer format, and conduct in uncertain cases.

In a production system it is not an individual skill but a design layer: instructions are versioned, tested and reviewed like the rest of the code.

RAG is the architecture that makes a language model answer using an organisation's own documents: on each question the system retrieves the relevant passages and places them in front of the model, which answers citing the source.

The operational difference from a model left to its own memory is traceability: every statement traces back to a document, a version and an access permission.

Fine tuning is the further training of a model on your own examples, to adapt its style, format or vocabulary to a specific domain.

It adapts behaviour, it does not refresh knowledge: answering over documents that change requires RAG. The two techniques answer different questions and often coexist.

Chunking is the stage where documents are split into coherent units before indexing, so that retrieval returns passages that make sense rather than fragments cut mid-thought.

It is the least visible and most decisive stage of a RAG system: much of what gets blamed on the model originates here.

A vector database is a store that holds text in numerical form and retrieves it by similarity of meaning rather than by word match.

It finds what resembles. Finding what is connected calls for a knowledge graph.

A knowledge graph represents knowledge as explicit entities and relations — a rule that repeals another, a supplier belonging to a group — and lets those connections be traversed.

It answers questions that similarity alone does not cover: which circulars depend on an amended article, which contracts involve the same counterparty.

An AI agent is a system that does not merely answer but acts: it queries a business system, fills in a case file, opens a ticket, and chooses its next step from the outcome of the last one.

The leap from an assistant is not answer quality but accountability: a system that acts needs permissions, explicit limits and a record of what it did.

An AI system is the combination of model, data, integrations, controls and interface that makes a model usable inside a process: the model is one component, not the whole.

It is the distinction that separates a demo from something that can go into production, and it explains why two projects using the same model differ in cost and return.

Document Intelligence is the automated reading of unstructured documents — PDFs, scans, certified email — to classify them, extract the relevant data and tie every piece of information back to the point in the document it came from.

Enterprise Search is search that answers employees' questions from the organisation's internal archives, respecting each person's permissions and citing the document the answer comes from.

What separates it from traditional search is what it returns: not a list of documents to read, but an answer with a verifiable reference.

OCR is the conversion of an image of text — a scan, a photograph of a document — into text a machine can read.

It is the precondition for any work on paper documents, and the quality of this stage caps the quality of everything downstream.

Speech AI covers the technologies that turn speech into text and text into speech, applied to call centre recordings, meeting minutes and voice assistance.

Vision AI covers the technologies that interpret images and video: quality control on a production line, reading plates and meters, conformity checks on photographed documents.

A multimodal system processes different data types together — text, images, audio — within the same request, for instance reading a technical drawing alongside the manual that describes it.

A guardrail is a control that bounds the behaviour of an AI system: what it may answer, from which sources, with which data, and what it must refuse or hand to a person.

It lives outside the model, as a verifiable control rather than an instruction inside the prompt, precisely because it has to hold when the model gets things wrong.

Human in the loop is the point at which an AI system stops and requires a human decision, typically above a risk threshold or below a confidence threshold.

It is not a stopgap for immature systems: where accountability stays with a person, it is a design requirement.

An audit trail is the record that keeps, for every answer an AI system gives, who asked what, which sources were consulted and which version of the system replied.

It is what makes an answer verifiable months later, when someone needs to reconstruct how a conclusion was reached.

An evaluation set is a collection of questions with their correct answers established in advance, used to measure an AI system before and after every change.

Without one, the only available measure is the impression of whoever tries the system, which cannot tell you whether a change made it better or worse.

Drift is the gradual degradation of an AI system's performance caused by change around it: updated regulations, new documents, different ways of writing.

It is why an AI system carries a running cost and not only a build cost.

On-premise describes a system installed on the organisation's own infrastructure, where data never leaves the corporate perimeter.

The choice between on-premise, cloud and hybrid setups is settled by the nature of the data and by regulatory constraints, not by technology preference.

Data residency is the constraint establishing which jurisdiction data must be stored and processed in.

For public bodies and financial institutions it is often non-negotiable, and as such it precedes the choice of model.

The AI Act is the European regulation that classifies artificial intelligence systems by risk level and imposes obligations that grow with that risk.

The obligations fall on those who adopt a system, not only on those who build it: this is why it concerns every organisation introducing AI into its processes.

A high-risk system is, in the AI Act's classification, an AI system whose use affects rights, access to services or personal safety, and which is therefore subject to the regulation's most demanding obligations.

Technical documentation, risk management, data quality, human oversight and event logging are requirements, not optional good practice.

AI governance is the set of roles, criteria and controls that establishes who decides what an AI system may do and who answers when it gets things wrong.

An AI assessment is the initial analysis that establishes, on a single process, whether an AI intervention is feasible, what it is worth and on what conditions, before any development commitment.

It produces a documented decision, including the case where the decision is not to proceed.