AI and LLM Glossary for Professionals: The Terms You Actually Need to Know
If a colleague mentions a token limit, a vector database or an agent in your next planning meeting, this glossary gets you up to speed in minutes. Below are plain-English definitions of the eleven AI and LLM terms that come up most often in UK workplaces right now, with the practical detail you need to ask the right follow-up question.
Why a shared vocabulary matters now
AI terminology has moved from research papers into procurement forms, board decks and ICO guidance in the space of about two years. The trouble is that vendors often use the same words to mean slightly different things: a
"knowledge base" from one supplier might be a folder of PDFs, while from another it means a fully indexed, query-ready system with citations. Getting the definitions straight before you sign anything, or before you brief a project team, saves wasted budget and awkward questions later. This glossary groups the terms into three practical clusters: the building blocks of language models, how AI finds relevant information, and how AI is made to act or connect to other tools.
The building blocks: LLM, token and context window
Large language model (LLM)
A large language model is a statistical system trained on vast amounts of text to predict the most likely next word (or fragment of a word) given everything that came before it. In practice, that simple mechanism, repeated millions of times during training, produces systems that can summarise, draft, translate and answer questions convincingly. An LLM has no built-in access to your company's documents, your contracts or today's news unless you give it that information explicitly, which is exactly the gap that retrieval systems like RAG are built to close.
Token
A token is the basic unit an LLM reads and generates. It is roughly three-quarters of a word in English, so "invoice" might be one token while "unsubscribe" might split into two or three. Tokens matter for two reasons: pricing (most APIs charge per token processed) and limits (every model has a maximum number of tokens it can handle at once, known as the context window). When a vendor quotes you a price per 1,000 tokens, that figure covers both the text you send in and the answer the model sends back.
Context window
The context window is the maximum amount of text, measured in tokens, that a model can consider in a single exchange. Modern models offer windows ranging from roughly 8,000 tokens to well over 100,000, which sounds generous until you try to fit an entire contract bundle, a year of support tickets or a product manual into one prompt. Stuffing everything into the context window is one option; retrieving only the relevant passages with RAG is the other. We compare the two approaches, including cost and accuracy trade-offs, in RAG vs Long-Context Models: When to Retrieve and When to Stuff the Prompt.
How AI finds meaning: embeddings, vector databases and hybrid search
Embedding
An embedding is a numerical representation of a piece of text (a word, a sentence, a paragraph) as a list of numbers, typically several hundred to a few thousand of them. Texts with similar meaning end up with similar numbers, even if they don't share a single word in common, which is how a search for "staff absence policy" can surface a document titled "sickness and leave procedure". If you want the full mechanics without the maths, Text Embeddings Explained: A Plain-English Guide for Non-Specialists walks through it step by step.
Vector database
A vector database stores embeddings and lets you search for the ones that are mathematically closest to a query, a technique called similarity search. It is the storage and retrieval layer underneath most modern search and RAG systems. On its own, though, pure vector search can miss exact matches such as reference numbers, acronyms or product codes, which is why most serious systems pair it with traditional keyword search.
Hybrid search
Hybrid search combines classic full-text search (often based on the BM25 algorithm) with vector similarity search, then merges the two rankings into one result set. This matters because BM25 is excellent at exact terms like "clause 14.3" or "SKU-7781", while vector search is better at conceptual matches. We go into the mechanics, and why one alone usually underperforms, in Hybrid Search: Why BM25 Plus Vectors Beats Either Alone for RAG. Kopik indexes every uploaded document for hybrid search by default, which is part of why answers stay grounded even when a question uses different wording from the source text.
Retrieval-augmented generation (RAG)
RAG is the pattern of retrieving relevant passages from your own documents, then handing those passages to an LLM alongside the user's question so it can generate an answer grounded in that material rather than in its general training data. It is the standard way to get an AI assistant to answer accurately about your policies, your contracts or your product catalogue, without retraining the underlying model. The output typically includes citations back to the source passages, which is what lets a reader verify a claim rather than take it on faith.
Reranker
A reranker is a second-pass model that takes the handful of candidate passages a search system found and re-scores them for relevance to the specific question asked, before any of them reach the LLM. Initial retrieval (hybrid or vector search) is optimised for speed across thousands or millions of chunks; a reranker is slower but more precise, and is usually applied only to the top 20 to 50 candidates. The practical effect is fewer irrelevant passages reaching the final answer, which reduces the chance of a confident but wrong response.
Agents, fine-tuning and MCP: making AI act
Agent
An AI agent is a system that doesn't just answer a question but takes a sequence of actions, deciding which tool to use, calling it, reading the result and deciding what to do next, often in a loop until the task is complete. A support agent might search a knowledge base, check an order status in a separate system, and draft a reply, all without a human choosing each step. The reliability of an agent depends heavily on the quality of the tools and data it can reach, which is why grounding its knowledge retrieval in a well-structured knowledge base matters as much as the reasoning model itself.
Fine-tuning
Fine-tuning means further training an existing model on a smaller, specific dataset so it adopts a particular tone, format or skill, for example drafting replies in your house style or classifying support tickets into your own categories. It changes how the model behaves, not what it knows about your current documents. This is the main reason fine-tuning and RAG solve different problems: fine-tuning teaches a style or a skill once, while RAG keeps answers current because it reads live documents at query time rather than baking facts into model weights that quickly go stale.
Model Context Protocol (MCP)
MCP is an open standard that lets AI assistants such as Claude, Cursor or ChatGPT connect to external tools and data sources, including knowledge bases, in a consistent way, without each tool needing a bespoke integration. In practice, once a knowledge base exposes an MCP server, you can query it directly from inside your coding editor or chat client rather than copying and pasting text back and forth. We cover the mechanics in The Model Context Protocol (MCP) Explained: A Plain-English Guide for UK Teams. Every knowledge base on Kopik can be queried this way, as well as through the website or a REST API, which is documented at /developers.
Hallucination and grounding: why answers go wrong, or right
A hallucination is a response that sounds plausible and confident but is factually wrong, invented, or not actually supported by any real source. LLMs hallucinate because they generate the most statistically likely continuation of text, not because they "know" a fact is true; when a model lacks reliable information on a topic, it fills the gap with something that reads convincingly. Grounding, by contrast, means constraining the model's answer to specific retrieved passages and ideally citing them, so a reader can check the claim against the original source rather than trusting the model's tone of confidence. We cover concrete mitigation techniques in Reducing AI Hallucinations: A Practical Guide to Grounded Answers.
- LLM: predicts the next token; no inherent access to your private documents
- Token: the billing and length unit models use to measure text
- Context window: the maximum text a model can consider in one go
- Embedding: a numeric fingerprint of meaning used for similarity search
- Vector database: stores and searches embeddings
- RAG: retrieves relevant passages, then generates a grounded answer
- Agent: chains tool calls and decisions to complete a multi-step task
- MCP: a standard way to connect AI clients to tools and knowledge bases
- Fine-tuning: retrains a model's behaviour, not its live knowledge
- Hallucination: a confident but unsupported or invented answer
- Reranker: a precision pass that reorders retrieved passages before generation
Quick reference: eleven terms at a glance
| Term | One-line definition | Why it matters |
|---|---|---|
| LLM | A model trained to predict the next token in text | Powers drafting, summarising and answering |
| Token | Roughly three-quarters of a word | Drives pricing and the context window limit |
| Context window | Maximum tokens a model can process at once | Determines whether you need RAG or can stuff the prompt |
| Embedding | Numeric representation of meaning | Lets search match concepts, not just keywords |
| Vector database | Stores and searches embeddings | Underpins semantic search and RAG |
| RAG | Retrieves passages, then generates a grounded answer | Keeps answers current and cited |
| Agent | Chains actions and tool calls to finish a task | Automates multi-step workflows |
| MCP | Standard protocol connecting AI clients to tools | Lets Claude, Cursor or ChatGPT query your data directly |
| Fine-tuning | Retrains behaviour or style on a smaller dataset | Changes how a model writes, not what it knows live |
| Hallucination | A confident but false or unsupported answer | The main risk RAG and grounding aim to reduce |
| Reranker | Re-scores retrieved passages for relevance | Improves precision before generation |
When reading an AI answer, check the citation first
Before you trust a grounded answer, click through to the cited passage and confirm it actually says what the summary claims. This single habit catches most residual errors, even in well-built RAG systems, and takes seconds once it becomes routine.
Putting the glossary to work
Knowing the terms is useful, but most professionals reach for this glossary because they are about to evaluate, build or commission an AI system. If that system needs to answer questions grounded in your own policies, contracts, manuals or research, you are looking for a RAG setup, not a fine-tuned model and not a raw LLM with a huge context window. Kopik lets you create a ready-to-query knowledge base without writing retrieval code: upload PDFs, Word documents, text or Markdown files, and the platform handles chunking, embedding and hybrid indexing automatically. You can keep a base private for internal use, or list it publicly in the catalogue and charge per question if your expertise is worth sharing.
For teams that already have a product or workflow, the practical next step is usually integration rather than a from-scratch build. A knowledge base can be queried through a simple web interface, through a REST API with your own keys, or through MCP so it plugs straight into Claude, Cursor or ChatGPT without custom glue code, as covered in Connect Claude, Cursor or ChatGPT to a Knowledge Base.
Try the terms on a real knowledge base
Create a free knowledge base from your own documents and see embeddings, hybrid search and grounded citations working together in minutes.
A few terms worth knowing next
This glossary covers the terms that come up most in day-to-day work, but the field moves quickly. Chunking strategy, for example, decides how a document gets split before indexing, and getting it wrong is a common cause of poor retrieval quality even when every other part of a RAG system is well built. If you are planning a build in-house, it is worth understanding the main approaches before you commit to a pipeline, and if your organisation operates in the UK, it is also worth checking your data handling against UK GDPR and RAG: What to Check Before You Build a Knowledge Base before any documents containing personal data are uploaded anywhere.
Frequently asked questions
What is the difference between a vector database and a knowledge base?
A vector database is a piece of infrastructure that stores and searches embeddings. A knowledge base is the complete, usable product: your documents, chunked, indexed (often with both vector and keyword search), and wired up to an LLM so you can ask questions and get cited, grounded answers. A vector database is one ingredient; a knowledge base is the finished dish.
Is RAG the same as fine-tuning?
No. Fine-tuning retrains a model's behaviour or style on a fixed dataset, so the knowledge gets baked in and goes stale as things change. RAG retrieves live documents at the moment a question is asked, so answers stay current without retraining, and each answer can be traced back to a specific passage.
What does context window mean in practice?
It is the maximum amount of text, measured in tokens, that a model can read and respond to in one exchange. If your documents exceed that limit, you either summarise aggressively, use a model with a larger window, or use RAG to retrieve only the relevant passages instead of sending everything at once.
Why do LLMs hallucinate?
Because they generate the statistically most likely continuation of text rather than checking facts against a verified source. Without grounding in retrieved documents and citations, a model will often produce a fluent, confident answer even when it has no real basis for the claim.
What is MCP and why does it matter for professionals?
MCP (Model Context Protocol) is an open standard that lets AI assistants like Claude, Cursor or ChatGPT connect to external tools and data sources, including knowledge bases, using one consistent method instead of bespoke integrations for every tool. It means a team can query its own documents directly from the AI client it already uses.
Do I need to understand embeddings to use RAG?
Not to use it day to day, but a basic grasp helps you judge quality. If search results feel off, it is often because of poor chunking or weak embeddings rather than the final model, so understanding the retrieval layer helps you ask the right diagnostic questions.
Get the Kopik newsletter
New knowledge bases, RAG guides and product news. One email every week or two, unsubscribe in one click.
By subscribing you agree to receive our newsletter. We never share your address.