The 12 basics · read this before diving into the glossary
If you've never touched a terminal, the glossary can feel like a mountain. So I pulled out the 12 words to understand first. Each one comes with a concrete mental image. Zero jargon. Read this first, then you can open any entry in the glossary without feeling out of your depth.
Nail these 12 words and the rest becomes readable
I put them in the order I understood them myself. First, what a model is (the brain). Then how you talk to it (prompt, token, context). Then how you build up its memory (embedding, vector, RAG, chunk). Finally, how you give it arms (agent, MCP, workflow). Plus the big fear of the whole field: hallucination.
Read the cards in order. If a mental image doesn't land for you, no worries — the explanation right after puts things back in place. And every card has a link to the detailed glossary entry if you want to dig deeper.
The model, the prompt, the token, the context
The first 4 words. You can do 80% of what you see on LinkedIn with just these.
Model (LLM)
A brain that has read a ton of books and mostly remembers the links between ideas.
When you use ChatGPT, Claude, or Gemini, you're talking to a model. By default it knows nothing about you, but it has digested a huge amount of text and can answer like a well-read assistant.
See the LLM entry →Prompt
The request you make to an assistant — be precise and you get something good, be vague and you get something mushy.
A prompt is just the message you write to the model. The quality of the answer depends 70% on the quality of your request. No magic needed, just being clear about the context, the format, and the result you expect.
See the Prompt entry →Token
Little Lego bricks — the model doesn't read whole words, it reads fragments.
A token is about 3-4 characters. When you see "1,000 tokens," count roughly 750 English words. It matters because everything is billed by the token and every model has a max limit per conversation.
See the Token entry →Context
An assistant's working memory — it only has in mind what you've told it in the current conversation.
Context is everything the model "sees" while it answers you: your request, the history, the files you gave it. Too short and it forgets the beginning. Too long and it gets lost, or it gets expensive.
See the Context entry →The hallucination, the embedding, the vector, the RAG
The model can make things up with total confidence. Luckily, we know how to teach it to check your documents before answering. Here's how.
Hallucination
A coworker who invents an answer with total confidence rather than admitting they don't know.
The model is designed to produce the most likely continuation of a text. If it doesn't know, it can put out a wrong answer that sounds right. This is risk number one. The fix: make it cite its sources and check them.
See the Hallucination entry →Embedding
The fingerprint of a text's meaning — two sentences close in meaning have close fingerprints.
An embedding is a long string of numbers that captures the meaning of a text. You don't need to read it yourself, but it's what lets a machine say "these two paragraphs are about the same thing."
See the Embedding entry →Vector / vector database
A smart filing cabinet that sorts your documents by theme — not alphabetically, by meaning.
A vector database stores all the embeddings of your documents. When you ask a question, it finds the passages closest in meaning. It's the tool behind every "AI that knows my documents."
See the Vector database entry →RAG
An assistant that checks your archives before answering — instead of making things up, it goes and looks.
RAG (Retrieval Augmented Generation) = first we find the right passages in your documents, then we ask the model to answer with those passages in front of it. It's the most effective fix against hallucinations.
See the RAG entry →The agent, the MCP, the workflow, the chunk
A model that answers is nice. A model that acts is useful. Here's how you give it tools, a plan, and how you break up the big documents it has to read.
Agent
A self-directed intern — you give it a mission and it runs through the steps without you approving every click.
An agent is a model that can use tools (send an email, read a file, run a web search) and loop until the task is done. It's what turns a chatbot into a coworker.
See the Agent entry →MCP
A universal jack plug — connect any tool to the model without recoding the connection every time.
MCP (Model Context Protocol) is an open standard that lets a model talk to your tools (Gmail, Notion, your database) cleanly. You install an MCP server, the model gets access to the tool. It's the USB of AI.
See the MCP entry →Workflow
A fixed recipe, the opposite of a cook who improvises — every step is defined ahead of time.
An AI workflow is a series of predefined steps (step A → then B → then C). More reliable and cheaper than an autonomous agent, but less flexible. For 80% of business cases, the workflow is the right call.
See the Workflow entry →Chunk
Tearing a book into bite-sized pieces so the model can swallow them one by one.
A model can't read a 300-page PDF all at once (the context isn't big enough, or it costs too much). So we cut it into pieces (chunks) of a few paragraphs each. It's the step right before creating the embeddings.
See the Chunk entry →Now you can explore the full glossary
You've got the map. You can open any entry in the A-Z glossary without feeling lost. And if a word trips you up, come back here — it's probably one of the 12 basics.
Open the A-Z glossary