AI terms in the glossary

Agentic AI is the broad term for AI systems that act autonomously, plan and make decisions in complex multi-step environments.

An AI agent is a software system around a language model that autonomously performs tasks, uses tools, and iterates decisions toward a goal.

Chain of Thought is a prompting technique that has the LLM reason step by step before giving the final answer, dramatically improving accuracy on complex tasks.

ClaudeBot is Anthropic's web crawler collecting public content for training and powering Claude's search and research capabilities.

The context window is the maximum amount of text (tokens) an LLM can process at once — its 'memory' size for one interaction.

An embedding is a numerical vector representation of text, image or other data, letting AI systems compute semantic similarity.

Few-shot learning is the technique of providing an LLM a few examples in the prompt (in-context learning) to teach it the desired style, format or reasoning for new input.

Fine-tuning is adapting a pre-trained LLM with domain-specific data to perform better on niche tasks or match a style.

A foundation model is a large-scale AI model trained on broad, diverse data and then adapted to a wide range of specific tasks — the backbone of modern generative AI.

Function calling is the mechanism by which an LLM decides to invoke an external function or API with structured parameters, rather than generating only text.

Gemini is Google's family of multimodal foundation models, introduced December 2023, natively handling text, images, audio and video.

Generative AI is the umbrella term for AI systems that create new content — text, image, audio, video or code — based on trained models rather than retrieving existing information.

GPTBot is OpenAI's web crawler, used to collect public web content for training future GPT models and for SearchGPT.

Guardrails are the safety measures around an LLM — input filters, output validation and monitoring — that prevent the model from producing harmful, inaccurate or unintended output.

A jailbreak is a prompting technique that persuades an LLM to ignore its safety training and produce content it would normally refuse.

LLMO is the discipline of optimising content and technical configuration so that LLMs (ChatGPT, Claude, Gemini, Perplexity) use and cite your brand and information in their answers.

Microsoft Copilot is Microsoft's AI assistant product line, built on OpenAI models and Microsoft's own models, integrated into Windows, Microsoft 365, Bing and Edge.

MCP is an open standard from Anthropic (2024) letting AI agents uniformly use external tools, databases and services.

Multimodal AI can process and generate text, images, audio and video together, unlike systems mastering only one modality.

PerplexityBot is the Perplexity AI crawler that indexes content for real-time use in Perplexity's answers, including source citations.

Prompt injection is an attack technique where a malicious user or external data source injects instructions into an LLM that override or bypass the original system prompt.

RLHF (Reinforcement Learning from Human Feedback) is the training method that fine-tunes LLMs after pre-training using human raters who compare and rank outputs.

SearchGPT is OpenAI's AI-powered search product, integrated into ChatGPT, that combines real-time web results with LLM answers including source citations.

A system prompt is the initial, hidden instruction that defines the behaviour, persona and constraints of an LLM before the user asks anything.

Temperature is a sampling parameter between 0 and 2 controlling how randomly an LLM selects its next token — low = deterministic, high = creative.

A token is the base unit an LLM processes text in — typically a sub-word fragment, roughly 4 characters or 0.75 words in English.

Training data is the collection of texts, images or other examples an AI model learns patterns from before deployment.

The Transformer is the neural network architecture underlying nearly all modern language models since 2017 — GPT, Claude, Gemini and Llama.

A vector database is a specialised database that efficiently stores vector embeddings and allows similarity search, essential for RAG systems.

Zero-shot learning is an LLM's ability to perform a task correctly without a single example in the prompt — entirely from pre-training.