Glossary

The terms, defined before they are sold to you.

How we explain each term to clients in a scoping call: what it actually means, where it is used, and what it costs when it is done properly. When a term deserves a full guide, the link is there.

RAG (Retrieval-Augmented Generation): Retrieval-augmented generation (RAG) is a technique where the system fetches the relevant pages from your documents at question time, hands them to the language model along with the question, and has it answer from that material instead of relying on training memory. It is the standard pattern behind every "chat with your documents" product.
Embedding: An embedding is a list of numbers — typically a few hundred to a few thousand values — computed from a piece of text so that texts with similar meaning get similar numbers. It lets software compare meaning instead of keywords: the mechanism behind semantic search, recommendations and retrieval in RAG systems.
Large Language Model (LLM): A large language model (LLM) is a neural network trained on massive text collections for one task: predict the next word. From that training it acquires grammar, facts, reasoning patterns and multiple languages — enough to power chatbots, copilots, search and agents. It predicts plausible text, not verified truth, which is the root of both its strengths and its failures.
Prompt injection: Prompt injection is an attack where hostile instructions hide inside data a language model will read — a user message, an uploaded PDF, a web page — to make it ignore its real instructions and do what the attacker wants: leak data, misuse tools, produce manipulated output. It ranks first in the OWASP LLM Top 10.
AI agent: An AI agent is a system that does not just produce text: it calls tools — APIs, databases, calendars — keeps state across steps, and pursues an outcome, like booking a slot, processing a refund or triaging tickets. A chatbot advises; an agent acts, which is why its failures are measured in consequences, not word quality.
Fine-tuning: Fine-tuning is continuing a model's training on your own examples — hundreds to thousands of input-output pairs — so its weights shift and it reliably produces that style, format or behaviour. It changes how the model acts; it does not add facts. For feeding it your documents, RAG is the right tool; fine-tuning serves style, format and tone.
Vector database: A vector database is a store built for embeddings: lists of numbers that capture the meaning of texts. Given a new text's numbers, it finds the most similar stored ones in milliseconds, using distance metrics rather than keyword matching. It is the retrieval engine behind semantic search, recommendation systems and every RAG deployment that answers from documents.
Temperature (LLM): Temperature is a numeric setting — usually between 0 and 1, sometimes up to 2 — that controls how randomly a language model picks its next word. Low values make output deterministic and consistent; high values allow more surprising, varied phrasing. For factual, repeatable tasks like support answers, temperature near zero is standard; creative drafting tolerates more variety.
Grounding: Grounding is anchoring a language model's answers to retrieved facts — pages from your documents, rows from your database — handed to it at question time, instead of relying on what it memorized in training. The model reasons over your material, cites it, and is instructed to say "I don't know" when the material does not contain the answer.
Hallucination: A hallucination is an answer a language model produces with complete fluency and confidence despite having no basis in fact — invented statistics, fabricated citations, policies that never existed. It happens because the model is trained to produce plausible text, not true text: when it lacks the answer, it completes the sentence anyway, and the completion reads exactly like knowledge.
Strangler fig pattern: The strangler fig pattern is a modernization strategy where new code is placed in front of a legacy system and takes over one slice of functionality at a time, while everything else keeps routing to the old one. Each release is small and reversible, the old system keeps working throughout, and there is no cutover weekend.
MVP (Minimum Viable Product): A minimum viable product is the smallest version that can test a real demand hypothesis in front of real users — one core path, end to end, everything else deliberately cut. Its job is learning, not revenue: to show whether strangers will pay or genuinely use the thing before more money goes into what may not be wanted.
TCO (Total Cost of Ownership): Total cost of ownership is the full cost of a solution over a defined period — purchase price plus operation, maintenance, training and exit. The sticker price is the visible fraction: a cheaper tool that eats hours of your team's time every month often costs more over a year than a pricier flat-fee one. TCO is the honest comparison unit.
AI receptionist: An AI receptionist is software that does the first round of reception work: it answers customer messages on WhatsApp, Telegram or the website, checks real availability, proposes only free slots and registers the booking, with automatic reminders. Unlike a simple chatbot it acts on real systems — and hands over to a human the moment the conversation leaves its perimeter.
NIS2: NIS2 — Directive (EU) 2022/2555 — is the EU's cybersecurity law, replacing the 2016 NIS directive. It obliges medium and large companies in 18 sectors, split between "essential" and "important" entities, to manage cyber risk, report incidents on a strict clock and secure their supply chains. The transposition deadline was October 2024; enforcement is rolling out through 2025 and 2026.