Agent
An AI system that can plan, call tools, observe results, and continue through a task instead of returning one immediate answer.
Chapter 6Reference
Glossary & Concept Index
Short definitions for the terms that repeat across the guide, with chapter links for deeper reading.
Alignment
Training and product work that steers a model toward useful, honest, safe, and controllable behavior.
Chapter 2Attention
The Transformer mechanism that lets each token weigh which other tokens matter most for interpreting it.
Chapter 1Context Window
The immediate working memory available to the model: the prompt, retrieved documents, tool outputs, and prior messages it can inspect now.
Chapter 3Diffusion
A generative media approach that starts from noise and repeatedly removes predicted noise until an image, video, or audio sample emerges.
Chapter 5Embedding
A vector representation that places text, images, or other data into a mathematical space where similar meanings sit near each other.
Chapter 3Eval
A repeatable test that checks whether an AI system behaves correctly on cases that matter for the product.
Chapter 8Groundedness
The degree to which an answer is supported by the evidence the system retrieved or was given.
Chapter 8Guardrail
A product-level check or constraint around the model, such as input filtering, output validation, citation checks, or tool permission rules.
Chapter 8LLM-as-Judge
Using a model to grade another model's output with a rubric. It is useful for scale, but it needs calibration because judges can be biased.
Chapter 8Mixture of Experts
A sparse architecture where a router activates only a few specialized expert networks for each token.
Chapter 4Prompt Injection
An attack where untrusted text tries to override the system's instructions, often through retrieved documents, web pages, or user input.
Chapter 8Query, Key, Value
The three vectors used by attention: what a token seeks, what other tokens offer, and the content that gets blended into the result.
Chapter 1Retrieval-Augmented Generation
A pattern where the system retrieves relevant external evidence, places it in context, and asks the model to answer from that evidence.
Chapter 3Test-Time Compute
Extra inference-time work spent on planning, searching, checking, or tool use before the final answer is returned.
Chapter 7