The AI terms that matter.

Retrieval-Augmented Generation (RAG) #

Also known as RAG

Retrieval-Augmented Generation (RAG) retrieves relevant passages at query time and feeds them into an LLM's context.

Embeddings #

Also known as vector embeddings

Embeddings are numeric vectors that place similar content close together in vector space.

Context Window #

The context window is the maximum text, measured in tokens, a model can consider at once.

Token #

A token is the basic unit of text an LLM reads and writes.

Model-Agnostic Architecture #

A model-agnostic architecture lets an application swap or route between LLM providers without a rewrite.

Fine-Tuning (vs RAG) #

Fine-tuning trains a model's weights on your own examples, changing its behavior directly.

Vector Search (Semantic Search) #

Also known as semantic search, similarity search

Vector search finds the nearest matches to a query by comparing embeddings, not exact keywords.

Chunking #

Chunking splits source documents into smaller passages before embedding, so retrieval returns focused text.

Agent (Agentic AI) #

An agent is an LLM given a loop, memory, and tools it can call.

Tool Calling #

Also known as function calling

Tool calling lets a model request a structured action instead of only generating text.

Model Context Protocol (MCP) #

Also known as MCP

MCP is an open standard for connecting LLM applications to tools and data sources.

Context Engineering #

Context engineering decides what occupies a model's context window at every step, and in what order.

Tool Contract #

A tool contract is the schema a tool exposes to a model.

Idempotency #

An idempotent operation produces the same result no matter how many times it runs.

Structured Output #

Structured output constrains a model's response to a defined schema instead of free-form prose.

Guardrails #

Guardrails are the checks that keep an LLM or agent inside acceptable bounds in production.

Golden Path #

A golden path is the single sanctioned way to do a common piece of work.

MCP Gateway #

An MCP gateway is a shared layer in front of several MCP servers, centralizing controls.

Context Rot #

Context rot is a model's declining ability to use context well as it fills with stale material.

Compaction #

Compaction summarizes or drops older context to make room in a filling window.

Context Isolation #

Context isolation gives each delegated unit of work its own scoped context window.

Eval Suite (Evals) #

Also known as evals, evaluation suite

An eval suite is a versioned set of test cases that measures whether outputs are good.

LLM-as-Judge #

LLM-as-judge uses another LLM to score a model's outputs against a rubric, at scale.

Observability #

Observability captures traces of every prompt, retrieval, tool call, and response for debugging.

Hallucination #

A hallucination is a confident, fluent output that is factually wrong or unsupported.

Prompt Injection #

Prompt injection is untrusted input crafted to override a model's system prompt or task.

Shadow Deployment #

Also known as shadow testing, dark launch

A shadow deployment runs a candidate release against real production traffic, offline, before release.

Inference Cost #

Inference cost is what it costs to run a trained model on a request.

Notebook-to-Production #

Notebook-to-production is the gap between a working prototype and a system that runs reliably.

Canary Release #

Also known as canary deployment, canary rollout

A canary release exposes a change to a small, deliberately chosen slice of traffic first.

Output Drift #

Also known as model drift, quality drift

Output drift is a gradual change in a live system's output, with no release behind it.

Model Deprecation #

Model deprecation is a provider retiring a model version on its own schedule.

Acceptance Rate #

Acceptance rate is the share of an agent's proposed changes a human reviewer approves.

Data Processing Agreement (DPA) #

Also known as DPA

A DPA is the contract naming a vendor as a processor of personal data.

Data Residency #

Data residency is where data is physically processed and stored, not where users are located.

Red Teaming #

Red teaming is deliberately attacking your own AI system to find what breaks first.

AI Register #

An AI register is a current list of every system that touches a model.

High-Risk AI System #

A high-risk AI system is one classified as carrying enough decision impact to trigger heavier compliance.

Human Oversight #

Human oversight is a checkpoint where a person can meaningfully approve, reject, or intervene.

System Card #

A system card is a short document stating an AI system's purpose, limits, and owner.

Shadow AI #

Also known as shadow use

Shadow AI is a model or tool used without going through the sanctioned path.

Tool Poisoning #

Tool poisoning hides an instruction in a tool's description that the model reads as legitimate.

Confused Deputy #

A confused deputy is a component tricked into using more authority than a request deserves.