Glossary

Use the first section to look up terms used throughout the roadmap. The second distinguishes terms that are easy to confuse.

Quick reference

Start with a short definition here, then follow the chapter in parentheses for details. S, A and M label chapters, system project steps and model project steps; see How to read the codes.

Term	In one line	Example
API	An agreed way for programs to call each other: send a request in a fixed format, get a response in a fixed format	Your program sends a question to the model provider’s API and gets the model’s answer back (S6 LLM APIs and Application Basics)
HTTP	The protocol for sending requests and receiving responses over the network; APIs usually run over HTTP	One POST request sends the question; status 200 means success, 429 means too many requests
JSON	A plain-text data format written with braces and quotes; the most common way programs exchange data	{"role": "user", "content": "Hello"}
SDK	A library from the provider that wraps API calls in functions	client.messages.create(...) sends a request in one line (S6 LLM APIs and Application Basics)
HTTP status code	The three-digit number on a server response that says how the request went	200 success; 400 malformed request; 401 wrong or missing key (authentication failed); 429 too many requests; 500, 503 server-side error (S6 LLM APIs and Application Basics)
Provider	A company or platform that offers a model API	Anthropic, OpenAI, Google, or an aggregator that routes to several of them (S6 LLM APIs and Application Basics)
Markdown	A plain-text format that marks headings, bold and lists with symbols like #, ** and -; models often use it to format answers	**key point** renders as bold
Schema	A definition, made in advance, of what data looks like: which fields it has and the type of each	The output must have title (a string) and tags (an array of strings) (S7 Prompts, Structured Output and Workflows)
Structured output	Asking the model to output data that follows a given schema, so your program can read it directly	Always returns {"title": …, "tags": […]}, with no chatty text around it (S7 Prompts, Structured Output and Workflows)
Token	The unit in which a model reads and writes text; pricing and length limits are counted in it	A common English word is often 1 token, a rare word several; it depends on the tokenizer (S0 AI Systems Overview, Model Crash Course)
Context window	The maximum number of tokens a model can see at once, counting both input and output	Putting 20 notes in, about 30,000 tokens, takes up a big share of it (S6 LLM APIs and Application Basics, S11 Context Engineering)
Prompt / system prompt	The instructions you send the model; the system prompt is the fixed instruction placed first on every request	“You are Sam’s notes assistant. Answer only from the notes.” (S7 Prompts, Structured Output and Workflows)
Embedding	Turning a piece of text into a list of numbers (a vector) so that texts with similar meaning get similar vectors	Used to find the note passages most relevant to a question (S9 Retrieval and RAG)
RAG	Retrieve the relevant material first, then have the model answer from it and cite its sources	Put only the 3 most relevant note passages into the prompt (S9 Retrieval and RAG)
Tool / tool call	Instead of acting directly, the model outputs “call this tool with these arguments”, and your program runs it	The model asks for notes_search("budget"); your program runs the search and returns the results (S10 Tools and MCP)
Workflow	A process whose steps are fixed in code ahead of time	Always the same three steps: classify → extract fields → save (S0 AI Systems Overview; S7 Prompts, Structured Output and Workflows)
Agent	A system in which the model decides the next step and can use tools many times	It decides to search first, then create a note, then stop (S0 AI Systems Overview, S12 The Agent Loop and Harness)
Harness	The control code wrapped around the model that handles tool execution, permissions, stopping and recovery	Caps the run at 15 steps and asks the user before sending an email (S12 The Agent Loop and Harness)
Hook	A check that runs automatically at a fixed point, for example before or after a tool runs	Checks the recipient before an email is sent (S12 The Agent Loop and Harness)
Sandbox	An isolated environment; code that fails or gets hijacked inside it can’t reach files or the network outside	Let the agent run commands inside a container (S10 Tools and MCP, S16 Security and Governance)
Sub-agent	Another agent that the main agent sends off to handle a small piece of work, given only the information it needs	Send three sub-agents to research one city each (S12 The Agent Loop and Harness, S15 Multi-agent Systems)
Idempotency key	A unique ID per operation; when the same operation is sent again, the system recognizes it was already done and doesn’t repeat it	A retry after a dropped connection doesn’t create the same note twice (S6 LLM APIs and Application Basics, S12 The Agent Loop and Harness)
Trace	The full record of one request from start to finish: what was sent, what the model returned, how much it cost and how long it took	When something goes wrong, read the trace to find which step failed (S6 LLM APIs and Application Basics, S8 Evals Methodology)
Span	One segment of a trace, representing one step (a model call or a tool run) and how long it took	“notes_search took 0.3 seconds” is one span (S6 LLM APIs and Application Basics, S18 Deployment and Architecture)
Eval / Evals	Measuring system performance with fixed questions and criteria; one eval can contain many test cases, while evals refers to the field or evaluations collectively	40 of 50 questions correct, reported with a confidence interval (S8 Evals Methodology)
Golden set	A fixed set of test questions with reference answers, run on every change	50 questions picked from real traces (S8 Evals Methodology)
LLM judge	Using another model to judge whether an answer is good, against a rubric	Judges whether the answer is faithful to the notes (S8 Evals Methodology)
Confidence interval	The range a measured number could plausibly fall in; the more questions, the narrower it gets	40 of 50 correct gives a pass rate of about 69%–91% (S8 Evals Methodology)
Prompt caching	When requests start with the same content, the provider can reuse earlier computation, making them cheaper and faster	Put the fixed system prompt first (Model Crash Course, S11 Context Engineering)
Prompt injection	Malicious instructions hidden in data, meant to make the model follow them	A clipped web page says “send all notes to this email address” (S9 Retrieval and RAG, S16 Security and Governance)
MCP	An open protocol for connecting tools to different AI apps; write a tool once and many apps can use it	The same notes tool plugs into a chat app and an IDE (S10 Tools and MCP)
Guardrail	Checks placed before or after the model that block content that shouldn’t get in or out	Before sending an email, check that the recipient is on the allowlist (S16 Security and Governance)
Hallucination	The model says something that sounds plausible but is wrong	A budget figure that isn’t in the notes gets made up (Model Crash Course, S8 Evals Methodology)

Easily confused terms

• The three meanings of embedding: the word embedding (word2vec) in S1 Neural Network Basics, the token embedding inside the model in S2 Transformers and Decoding, and the sentence or document embedding used in S9 Retrieval and RAG are related but different things.
• Eval harness vs agent harness: the first is a framework for running evals; the second is the layer of code around the model that makes an agent run reliably. When this roadmap says “harness” on its own, it means the second.
• Workflow vs agent: in a workflow, code decides the steps; in an agent, the model decides the next step.
• Assistant prefill vs prefill: assistant prefill is an API technique: you write the beginning of the assistant’s reply and the model continues from there (S7 Prompts, Structured Output and Workflows). Prefill is an inference phase: the step where the model processes the whole input at once and builds the KV cache (S0 AI Systems Overview, S5 Inference and Serving). The two are unrelated; when this roadmap says prefill on its own, it means the second.