AI & open-source terms, explained

An AI agent is a system that uses a large language model to plan and take actions toward a goal — calling tools, reading results, and iterating with little or no human input.

MCP is an open standard that lets AI assistants connect to external tools, data, and services through one consistent interface, so any MCP client can use any MCP server.

RAG is a technique that retrieves relevant documents from a knowledge base and feeds them to a language model as context, so its answers are grounded in your data instead of only its training.

A vector database stores embeddings — numeric representations of text, images, or code — and finds the most similar ones fast, powering semantic search and RAG.

A large language model is a neural network trained on massive text corpora to predict the next token, giving it the ability to generate and understand human language.

A local LLM is a language model that runs on your own machine or servers instead of a hosted API — for privacy, offline use, and cost control.

An AI coding assistant uses an LLM to help you write, edit, and debug code — from inline autocomplete to autonomous agents that implement whole tasks across a repository.

Browser automation drives a real web browser programmatically — clicking, typing, and navigating — for testing, scraping, and increasingly AI agents that operate the web.

An embedding is a numeric vector that represents the meaning of text, code, or an image, so similar items sit close together in vector space and can be compared mathematically.

Fine-tuning adapts a pre-trained model to a specific task or domain by continuing training on a smaller, targeted dataset, changing its behaviour without training from scratch.

Prompt engineering is the practice of crafting a language model's input — instructions, examples, and structure — to reliably get the output you want.

A transformer is the neural-network architecture behind modern language models, using self-attention to weigh how every token in the input relates to every other.

Inference is the process of running a trained model to generate output — for an LLM, turning a prompt into tokens — and the stage where latency, throughput, and cost are decided.

Quantization shrinks a model by storing its weights at lower numeric precision — for example 4-bit instead of 16-bit — cutting memory and speeding inference with a small quality trade-off.

A context window is the maximum amount of text — measured in tokens — a language model can consider at once, covering both your input and its generated output.

A token is the unit of text a language model processes — roughly a word-piece — and the thing models are billed by and limited to.

A multi-agent system coordinates several AI agents — often with specialised roles — that communicate and divide a task, rather than relying on one agent to do everything.

Function calling lets a language model invoke external tools or APIs by emitting a structured request your code runs and feeds back — the mechanism behind tool-using agents.

An agentic workflow is an automation where an AI agent — not a fixed script — decides the steps: planning, calling tools, checking results, and adapting until a goal is met.

Semantic search finds results by meaning rather than exact keywords, comparing embeddings of the query and documents to surface conceptually similar matches.

An AI hallucination is when a language model produces fluent but false or fabricated information, stating it with the same confidence as fact.

AI guardrails are the checks around a language model that constrain its inputs and outputs — blocking unsafe content, enforcing formats, and catching hallucinations.

A small language model is a compact LLM — typically a few billion parameters or fewer — designed to run cheaply on local or edge hardware while staying capable for focused tasks.

A multimodal model understands or generates more than one type of data — such as text plus images, audio, or video — within a single model.

LoRA is a parameter-efficient fine-tuning method that trains small adapter matrices instead of the full model, cutting the cost of customising an LLM dramatically.

Mixture of experts is a model design that routes each token through only a few of many specialised sub-networks, giving large total capacity at a fraction of the compute per token.

Self-hosting means running software on infrastructure you control instead of a vendor's cloud — for privacy, cost control, customisation, and no per-use fees.

Observability is the ability to understand a system's internal state from its outputs — logs, metrics, and traces — so you can debug and monitor it in production.

WebAssembly is a portable binary format that runs near-native-speed code in the browser and beyond, letting languages like Rust, Go, and C target the web and the edge.

Reranking reorders an initial set of retrieval results with a more accurate — and costlier — model, putting the most relevant items on top.

A knowledge graph stores information as entities and the relationships between them, letting you query connected facts rather than isolated rows or documents.

A diffusion model generates images, audio, or video by starting from random noise and iteratively denoising it toward a sample that matches a prompt.