A living glossary of the AI vocabulary that actually appears in IT job descriptions, vendor documentation and meetings. Fifteen core terms are defined here in plain English with an admin-context example each — LLM, token, embedding, prompt, RAG, agent, vector database, fine-tuning, inference, API, transformer, context window, hallucination, multimodal and system prompt — plus a repeatable method for absorbing new terms as the field evolves.
This is a working glossary, not a dictionary. Each term is defined in one or two sentences of plain English and immediately anchored to something an IT professional touches — a log file, an Intune policy, a PowerShell script. The fifteen core terms above cover roughly 90% of what you will read in vendor docs and job posts in 2025: LLM (the model itself), token (its unit of text), embedding (text as vectors), prompt (your instruction), RAG (grounding answers in your documents), agent (AI that acts through tools), vector database (where embeddings live), fine-tuning (specialising a model), inference (the model producing output), API (programmatic access), transformer (the architecture), context window (its working memory), hallucination (confident wrongness), multimodal (text+image+audio) and system prompt (the hidden instructions steering behaviour).
Vocabulary is the entry fee to the AI conversation. In job interviews you will be asked to explain RAG or the difference between fine-tuning and prompting; in vendor meetings you will hear 'embeddings', 'vector store' and 'agentic' within five minutes; in security reviews you will need to know what a system prompt is to discuss prompt injection. Without the words, you cannot even form the questions. With them, documentation, changelogs and AI strategy documents become readable — and you stop nodding politely in meetings you could be leading.
Read the terms in dependency order rather than alphabetically. Start with the model stack: transformer → LLM → parameters → training → fine-tuning → inference. Then the text units: token → context window → system prompt → prompt. Then the memory layer: embedding → vector database → RAG. Finally the behaviour layer: hallucination, multimodal, agent, API. Each layer only makes sense on top of the previous one, which is exactly how the earlier topics in this level built up. New terms that appear later (MCP, guardrails, evals) slot into these same layers.
Start: everyday AI vocabulary
↓
Core: Tokens, Prompts, Context
↓
Systems: RAG, Vector DB, Agents
↓
Ops: API, Inference, Fine-tuningWhere these terms actually meet you: your security team asks whether 'Copilot can leak data via embeddings stored in a vector database' — you can now explain that Microsoft 365 Copilot uses your tenant's Graph data at inference time under existing permissions. A vendor pitches 'we fine-tune an LLM for your tickets' — you can ask whether retrieval (RAG) over current tickets wouldn't be fresher and cheaper. A job post wants 'experience with AI APIs and prompt engineering' — you know it means calling Azure OpenAI or Graph-connected models programmatically and writing structured prompts.
Create glossary.md with the 15 terms above as headings. For each, paste the definition and write your own one-line example from your environment, e.g. Token: "my 300-line IME log excerpt is about 4,000 tokens." Personal examples stick.
Ask: "Quiz me on Tokens, RAG, embeddings, context window and system prompts — 5 questions, one at a time, mark my answers." Getting tested is far more effective than re-reading, and the AI adapts difficulty to your gaps.
Ask the AI: "Show a minimal PowerShell example that calls an LLM API with my system prompt and a user prompt." Seeing prompt, system prompt, API and inference in 15 lines of real code converts definitions into skills.
Add a recurring task: every Friday, ask the AI "What new AI terms became relevant for IT admins recently?" (mind the knowledge cutoff), verify the genuinely new ones against vendor docs, and append three entries to glossary.md. A glossary you maintain beats one you downloaded.
Deliberately use three learned terms in real work: write a ticket comment citing "context window limits" for a log you could not fully paste, mention "RAG" when discussing the internal wiki bot, and note "inference cost" in a Copilot licensing discussion. Usage cements knowledge.
Invoke-RestMethod -Method Post -Uri "$env:AZURE_OPENAI_ENDPOINT/openai/deployments/gpt-4o-mini/chat/completions?api-version=2024-10-21" -Headers @{ Authorization = "Bearer $env:AZURE_OPENAI_KEY" } -ContentType "application/json" -Body (@{ messages = @(@{ role = "system"; content = "You are an IT glossary assistant. Answer in one line each." }, @{ role = "user"; content = "Define Token, RAG and Context Window." }); temperature = 0; max_tokens = 200 } | ConvertTo-Json -Depth 5)Three glossary terms are security-critical. System prompt: the hidden instruction set that defines what an AI may do — prompt injection attacks try to override it through untrusted input. Embeddings and vector databases: they encode your documents; where they are stored, who can query them and whether they contain personal data must be reviewed like any database. API keys: LLM APIs are paid resources — keys belong in secret management (Azure Key Vault), never in scripts committed to repos, and should be rate-limited and rotated.
Retrieval-augmented generation means the AI answers from your documents instead of only its training memory. When a question arrives, we search a knowledge base for relevant passages, hand them to the model together with the question, and the model answers citing that material. Benefits: answers use current, company-specific facts, and we can control what it reads. It is how an internal AI helpdesk knows our actual Intune standards without retraining the model.
An embedding is a list of numbers (a vector) that represents a piece of text so that texts with similar meanings get similar vectors. A vector database stores millions of these and can find the nearest ones in milliseconds. That combination is the engine inside RAG: your documents are embedded once, and at question time the few most relevant passages are retrieved in milliseconds for the model to read.
The system prompt is the persistent instruction set that defines role, rules and boundaries for every conversation — written by the application, hidden from end users. The user prompt is what a person types. Injection attacks target the system prompt by smuggling instructions inside user content or retrieved documents, which is why untrusted input must be treated as data, never as instructions.
Turn glossary.md into a flashcard deck: front = term, back = your one-line definition + your personal example. Run 15 cards daily for a week using any spaced-repetition app, then have the AI quiz you in random order. Finish by writing a 200-word mini-article for your team: 'The 10 AI terms our helpdesk should know in 2025' — teaching is the final test of vocabulary.
The core terms here — LLM, token, embedding, RAG, transformer, context window — have been stable for years and underpin everything else. What churns fast is the product and buzzword layer (agent frameworks, MCP, evals). That is why the topic teaches a living-document method: a stable core plus a weekly habit of adding verified new terms.
For endpoint and cloud roles: LLM, token, context window, prompt, RAG, fine-tuning vs prompting, hallucination and API. Expect the classic questions — explain RAG, why models hallucinate, prompt vs system prompt — plus scenario ones like how you would safely let an AI read your logs. Levels 2 and 3 of this curriculum cover the scenario answers.
The word is overused, but the concept is real: an agent is an LLM given tools and a loop — it can plan, call tools (APIs, scripts, search), observe results and continue. The difference matters operationally: a chatbot answers; an agent acts, which changes the security review completely. Treat agent claims as architecture questions: which tools, which permissions, which logging.