# Glossary

> **How to use this**
> Not a chapter to read front to back but a **reference page**. Terms you have
> not met yet are expected — look one up whenever another chapter stalls you.

## IT basics

Start here if you have no technical background. The detail is in the
[IT basics](/guide/it-what-is-computer) part.

| Term | Meaning |
|---|---|
| **Memory vs storage** | what is open now vs what survives power off. Why "restart it" works → [detail](/guide/it-what-is-computer) |
| **Path** | the one-line address saying which folder a file is in → [detail](/guide/it-files) |
| **Terminal · CLI** | the window where you instruct the computer in text → [detail](/guide/it-terminal) |
| **PATH · command not found** | the folder list commands are looked up in. Usually fixed by a new terminal window |
| **Runtime · Node.js** | the program that executes code. Node.js is the one for JavaScript → [detail](/guide/it-programs) |
| **IP · port** | which machine, and which program on it → [detail](/guide/it-network) |
| **localhost** | this very machine. An address that never leaves it |
| **Domain · DNS** | the name, and the desk that turns a name into an address |
| **Server · client** | the side that answers and the side that asks. A role, not a class of machine → [detail](/guide/it-server) |
| **Cloud · on-premises** | computers you rent and computers inside your walls |
| **HTTP request · response** | the single round trip the web is made of → [detail](/guide/it-http) |
| **Status codes** | 401 key · 403 permission · 404 address · 429 rate · 500 server |
| **API · endpoint** | the counter built for programs, and one address on it → [detail](/guide/it-api) |
| **SDK · library** | that counter pre-wrapped per language |
| **JSON · CSV** | the notation programs exchange, and a table held as text → [detail](/guide/it-data-formats) |
| **Encoding (UTF-8)** | the agreement turning characters into numbers. Mismatch it and text turns to garbage |
| **Table · row · column · key** | the shape a database holds data in → [detail](/guide/it-database) |
| **SQL** | how you ask a database. `SELECT` reads; `UPDATE`/`DELETE` are hard to undo |
| **Authentication · authorisation** | who you are, and what you may do → [detail](/guide/it-accounts) |
| **API key · token** | the key for programs. One per purpose; retire it if it leaks |

## Basics

| Term | Meaning |
|---|---|
| **Language model** | a program that predicts the next word. Summarising, translating, and drafting all come from this → [more](/guide/ai-language-model) |
| **Prompt** | everything you give the model: instruction + context + material |
| **In-context learning** | changing behaviour by putting examples in the prompt. One kind of learning, but **the model does not change** and it is gone when the conversation ends → [more](/guide/ai-rag-vs-finetune) |
| **Token** | the unit of billing and length. Roughly one to two per English word |
| **Hallucination** | a plausible wrong answer. Structural, not a defect → [more](/guide/ai-hallucination) |
| **Training cutoff** | the end date of what the model learned. It knows nothing after it |

## Conversation and memory

| Term | Meaning |
|---|---|
| **Context window** | everything the model sees at once. What "remembering" actually is → [more](/guide/ai-context-window) |
| **Context overflow** | the window fills and the earliest part falls off → [more](/guide/ai-context-overflow) |
| **Context rot** | quality sliding quietly as clutter accumulates, before anything overflows → [more](/guide/ai-memory) |
| **Compaction** | keeping recent turns verbatim and replacing earlier ones with a summary |
| **Memory** | a record written outside the conversation and re-read in later sessions → [more](/guide/ai-memory) |

## Working with material

| Term | Meaning |
|---|---|
| **RAG · grounding** | finding relevant material and passing it in with the question. **Non-parametric memory** — kept outside the model, so the store learns rather than the model → [more](/guide/ai-grounding) |
| **Embedding** | text turned into a list of numbers. Close in meaning, close in numbers |
| **Semantic search** | retrieval by closeness of meaning, using embeddings |
| **Chunk** | a piece a document is split into for retrieval |
| **Rerank** | a second pass that filters and orders retrieved chunks |
| **Pre-training** | the stage where a model learns language itself from a vast corpus. **Most of its ability is settled here, and you do not do this part** → [more](/guide/ai-rag-vs-finetune) |
| **Fine-tuning** | **further** training, laying your data on top of an already pre-trained model. One part of training, and for voice and format rather than facts → [more](/guide/ai-rag-vs-finetune) |

## Tools and agents

| Term | Meaning |
|---|---|
| **Tool use** | the model calling an external capability. The system executes it → [more](/guide/ai-tool-use) |
| **Agent** | a structure that takes a goal and loops, choosing tools as it goes → [more](/guide/ai-agent-basics) |
| **MCP** | the standard for attaching tools → [more](/guide/mcp-why-standard) |
| **MCP server** | the side that exposes capabilities through that standard |
| **Subagent · delegation** | splitting large work and handing pieces to other agents |
| **HITL (human in the loop)** | placing human approval in front of irreversible actions → [more](/guide/ha-guardrails) |
| **Guardrail** | any mechanism that stops speed from turning into errors |

## Models and cost

| Term | Meaning |
|---|---|
| **Model tier** | top, mid, light. Capability and price move together → [more](/guide/ai-choosing-models) |
| **Input/output tokens** | what goes in and what comes out. Output costs roughly 5x |
| **Caching** | storing a repeated prefix and reusing it to cut cost |
| **Thinking depth** | how hard it reasons before answering. Higher is slower and pricier |
| **Credit** | the balance usage is drawn from in Connect → [more](/guide/cn-usage) |

## Evaluation and operations

| Term | Meaning |
|---|---|
| **Evals** | a system for checking answer quality in numbers → [more](/guide/ai-evaluation) |
| **Golden dataset** | test cases with confirmed correct answers |
| **LLM-as-judge** | having a model do the scoring |
| **Baseline** | the measurement taken **before** deployment. Without it you cannot prove effect → [more](/guide/ax-measurement) |
| **Cycle time** | time to process one item. The metric executives read most easily |
| **Rework rate** | share of AI output a human had to fix. A hidden cost |
| **Trace** | a tree-shaped record of the model and tool calls one request branched into → [more](/guide/ax-audit-trail) |

## Security and regulation

| Term | Meaning |
|---|---|
| **Prompt injection** | an attack where text in material the model reads acts as an instruction → [more](/guide/mcp-prompt-injection) |
| **Indirect injection** | the form arriving via web, document, or email with no user action |
| **Least privilege** | grant only what is needed now, widen later → [more](/guide/mcp-security) |
| **Irreversible action** | sending email, payment, deletion, external publishing. The criterion for approval points |
| **High-impact AI** | AI in fields materially affecting rights — hiring, lending, healthcare → [more](/guide/ax-compliance) |
| **AI user-business** | a business supplying products or services using someone else's AI. Most companies |
| **Watermark / labelling duty** | the obligation to mark output as AI-generated |

## Transformation and method

| Term | Meaning |
|---|---|
| **Adoption vs transformation** | inserting a tool versus moving a metric → [more](/guide/ax-adoption-vs-transformation) |
| **POC** | the stage where you build it and confirm the effect → [more](/guide/ax-phase3-poc) |
| **Gate** | the criterion for whether you may proceed to the next stage |
| **Hyper-agile** | a short loop of intent → orchestration → validation → delivery → [more](/guide/ha-what-changes) |
| **Micro sprint** | that loop taken at its shortest → [more](/guide/ha-micro-sprint) |
| **IntentOps** | treating intent itself as something to be managed → [more](/guide/ha-intentops) |

## Product terms — HyperTeams

| Term | Meaning |
|---|---|
| **HyperTeams** | the control room for driving and watching your own machine from the web → [more](/guide/ht-what-is-it) |
| **Work folder** | the folder scope an agent is allowed to touch → [more](/guide/ht-working-directory) |
| **Autonomy level** | how far it may go on its own → [more](/guide/ht-autonomy) |
| **Skill** | a repeated procedure captured as a document for reuse → [more](/guide/ht-skills) |
| **Tunnel** | the add-on that lets you reach your machine from outside |

## Product terms — Connect

| Term | Meaning |
|---|---|
| **Workspace** | the isolation unit where data does not mix → [more](/guide/cn-workspace) |
| **Thread** | a single conversation → [more](/guide/cn-thread) |
| **Department** | people grouped along the org chart → [more](/guide/cn-groups) |
| **Conversation group** | conversations grouped by topic, with a visibility setting |
| **Drive** | the file store that feeds answers → [more](/guide/cn-drive) |
| **System** | one company machine attached to the workspace → [more](/guide/cn-systems) |
| **Channel** | a route by which conversations arrive from an outside service → [more](/guide/cn-channels) |
| **Queue** | the list of scheduled and waiting work → [more](/guide/cn-queue) |
| **Playground** | where you test an agent before putting it into service → [more](/guide/cn-playground) |

---

This page is here to come back to. On to the main track →
[Adoption is not transformation](/guide/ax-adoption-vs-transformation)