Glossary

Agent — An AI system equipped with tools (like web browsers or code interpreters) that can autonomously plan and execute multi-step workflows toward a goal.

Alignment — The engineering discipline of ensuring a model's outputs consistently match human intentions and safety constraints.

Context Window — The maximum amount of data (measured in tokens) a model can hold and process in a single interaction.

Direct Prompt Injection — A user typing malicious commands straight into a chatbot's interface to override its instructions.

Embeddings — Vectors representing the semantic meaning of data, letting a model mathematically compare concepts for similarity.

Few-Shot Prompting — Giving a model a small number of concrete examples in the prompt to demonstrate the exact format or logic wanted.

Fine-Tuning — Further training a pre-trained model on a smaller, specific dataset to specialize it for a task.

Guardrails — Rules, classifiers, or filters placed around a model to block harmful or off-topic outputs.

Hallucination — A model generating false or fabricated information while presenting it with high confidence.

Indirect Prompt Injection (IPI) — A cyberattack where malicious instructions are hidden inside external content (a webpage, PDF, email) an AI is asked to process, hijacking it without the user's knowledge.

Inference — The operational phase where a trained model receives a prompt and produces a response, as opposed to training.

Large Language Model (LLM) — A transformer-based AI trained on massive text data to understand, generate, and predict language.

Multimodal — An AI system that natively understands and generates multiple data types — text, image, audio, video — at once.

Parameters — A neural network's internal weights, adjusted during training; more parameters generally allow more complex pattern capture.

Prompt Injection — The overarching term for tricking an AI into ignoring its original instructions in favor of attacker-supplied ones.

Retrieval-Augmented Generation (RAG) — Connecting a model to an external database so it can search for and use real facts before answering, reducing hallucination.

RLHF (Reinforcement Learning from Human Feedback) — A training phase where human raters score model responses to teach it which answers are helpful and safe.

System Prompt — The hidden instructions a developer gives a model defining its persona, rules, and boundaries before the user ever interacts with it.

Tokens — The chunks of data — pieces of words, characters, or image patches — that a model actually reads and generates.

Vector Database — Storage built to hold and rapidly search embeddings, the infrastructure behind RAG.

World Model — An AI architecture that simulates physical reality — object permanence, gravity, causality — letting users interact with a predicted environment over time.