Skip to content

AI Glossary

Plain-language explanations of key AI concepts, from LLM and RAG to fine-tuning and agents.

LLM (Large Language Model)

A neural network trained on massive text that predicts the next token, powering chat, writing, coding, and more.

Token

The basic unit of text a model processes; roughly 4 English characters or 1–2 Chinese characters per token. Billing and context limits are measured in tokens.

RAG (Retrieval-Augmented Generation)

A technique that retrieves relevant documents from a knowledge base and feeds them to the model so it answers grounded in your data.

Embedding

A numeric vector representation of text (or images) that captures semantic meaning, enabling similarity search.

Fine-Tuning

Further training a pre-trained model on your own data so it masters a specific style, format, or task.

LoRA (Low-Rank Adaptation)

A parameter-efficient fine-tuning method that freezes base weights and trains small low-rank matrices, drastically cutting cost.

Prompt Engineering

The craft of designing inputs (prompts) to reliably get high-quality outputs from a model.

AI Agent

An AI system that autonomously plans, uses tools, and takes multi-step actions to accomplish a goal.

MCP (Model Context Protocol)

An open standard that lets models connect to external tools and data sources in a unified way.

Context Window

The maximum number of tokens a model can consider at once; content beyond it is truncated or forgotten.

Hallucination

When a model confidently generates false or fabricated content. Mitigate with RAG, tool use, and verification.

Multimodal

Models that understand and generate across multiple data types, such as text, images, audio, and video.

Inference

The process of running a trained model to produce outputs, as opposed to training it.

Open Source / Open Weights

Models or tools whose source code or weights are publicly available to use, study, and self-host.

Vector Database

A database that stores embeddings and performs fast similarity search, the backbone of RAG systems.

Transformer

The neural network architecture behind modern LLMs, using self-attention to process all tokens in parallel.

Attention Mechanism

A mechanism that lets a model weigh which parts of the input matter most for each output token.

Tokenizer

The component that splits raw text into tokens the model can process, and converts them back to text.

Diffusion Model

A generative model that creates images by gradually denoising random noise, powering tools like Stable Diffusion.

RLHF (Reinforcement Learning from Human Feedback)

A training technique that aligns models with human preferences using a reward model learned from human ratings.

Chain-of-Thought (CoT)

A prompting technique where the model reasons step by step, improving accuracy on complex tasks.

Few-Shot Learning

Guiding a model by providing a few examples in the prompt, without any additional training.

Temperature

A sampling parameter that controls randomness: lower values are more focused, higher values more creative.

Quantization

Reducing the numeric precision of model weights (e.g. to 4-bit) to shrink size and run models on less hardware.

GGUF

A file format for storing quantized models, widely used to run LLMs locally with tools like llama.cpp and Ollama.

Mixture of Experts (MoE)

An architecture that routes each token to a few specialized sub-networks, boosting capacity without proportional cost.

Function Calling / Tool Use

A capability that lets a model output structured calls to external functions or tools, enabling agents and integrations.

Knowledge Distillation

Training a smaller 'student' model to mimic a larger 'teacher' model, retaining most quality at lower cost.

AI Alignment

The field focused on ensuring AI systems behave in line with human values and intended goals.

Zero-Shot

Asking a model to perform a task with no examples, relying entirely on its pretrained knowledge.

Agentic AI

AI systems that pursue a goal autonomously — planning, using tools, and taking multi-step actions — rather than just responding to a single prompt.

Vibe Coding

A coding style where you describe what you want in natural language and let an AI generate and iterate on the code, focusing on intent and results over writing every line by hand.

Reasoning Model

An LLM trained to 'think' before answering — spending extra compute on internal step-by-step reasoning to solve harder math, coding, and logic problems.

Text-to-Video

AI models that generate video clips directly from a written description, controlling scene, motion, and style through the prompt.

Foundation Model

A large model trained on broad data that serves as a general-purpose base, adaptable to many downstream tasks via prompting or fine-tuning.

Guardrails

Safety controls around an AI system that constrain what it can say or do — filtering harmful content, limiting tool access, and enforcing policies.

System Prompt

The hidden instruction that sets an AI assistant's role, rules, and style before any user message — the foundation of its behavior.