Do I need to know all of these to use ChatGPT or Claude?

No. You can get useful work done knowing only three of these: prompt, token, and context window. The rest matter if you start building something — a chatbot for your company, a search tool over your own files, a custom workflow.

What is the difference between a prompt and a system prompt?

A prompt is what you type in the chat box right now. A system prompt is a standing instruction that sits above your chat, often written by the developer of the app, that shapes how the AI responds to everything you send. When you "customize" a GPT or a Claude Project, you are mostly editing the system prompt.

Is RAG the same as fine-tuning?

No, and the difference matters. RAG keeps the model as-is and lets it look things up at answer time. Fine-tuning changes the model itself by continuing to train it on your data. RAG is cheaper, easier to update, and better for factual Q&A. Fine-tuning is for changing style or teaching narrow, repetitive tasks.

How many tokens is a typical document?

Rough rule: one token is about 0.75 English words, so 1,000 tokens is roughly 750 words or a page and a half. A 20-page PDF is around 10,000 tokens of clean text — more if there are tables and footnotes.

Why do people keep telling me to convert my PDFs to markdown?

Because PDFs are built for printing, not for AI. The underlying text is often out of order, broken across columns, or locked inside images. Markdown is clean, ordered, and token-efficient, so the model spends its attention on your content instead of untangling layout.

What is the difference between an embedding and a vector database?

An embedding is one piece of text turned into numbers. A vector database is a system for storing millions of embeddings and finding the closest matches to a query quickly. You need both to do semantic search over your own documents.

Glossary of AI & Document Terms: Plain English

Fifteen terms keep showing up in AI conversations, product docs, and Twitter arguments. Most explanations assume you already know what the other fourteen mean. This page fixes that. Each entry starts with one plain sentence, then adds a real example so the word sticks. If you are new to the whole area, read Markdown basics and choosing a format first — they use most of these terms in context.

Jump to a term using the bar below. Every definition links out to the tools, guides, and blog posts where the term actually matters.

Chunking

The act of cutting a long document into smaller pieces before feeding it to an AI.

AI models have a limit on how much text they can read at once, and searching works better on smaller pieces anyway. So a 500-page employee handbook gets sliced into chunks — often a few paragraphs each — before it goes into a search index. When you ask a question, the system retrieves the few chunks most relevant to your question and sends only those to the model. Good chunking respects natural boundaries: section headings, paragraph breaks, table rows. Bad chunking cuts sentences in half and confuses everything downstream.

Context window

The maximum amount of text an AI model can read and remember in a single conversation.

Think of it as the model's short-term memory. Once the conversation plus the files you've pasted exceed that limit, the oldest stuff starts getting pushed out or summarized. In early 2023, 8,000 tokens (about 6,000 words, or 12 pages) was generous. In 2026, models routinely handle 200,000 tokens (a 300-page book) and Gemini and Claude can go to 1 million and beyond. Bigger is not always better — long contexts get expensive and models often pay less attention to the middle.

Embedding

A list of numbers that represents the meaning of a piece of text, so computers can find things with similar meaning.

Here is the trick. Run "the cat sat on the mat" and "a feline rested on the rug" through the same embedding model and you get two long lists of numbers, maybe 1,536 numbers each. Those two lists end up very close to each other in math space even though they share almost no actual words. That closeness is what lets semantic search work. You turn the user's question into an embedding and find the document chunks whose embeddings sit nearest. No keyword matching required.

How embeddings line up

"the cat sat on the mat" → [0.12, -0.33, 0.87, ...]

"a feline rested on the rug" → [0.14, -0.31, 0.85, ...] close

"quarterly revenue report" → [-0.41, 0.62, -0.19, ...] far

Fine-tuning

Continuing to train an existing AI model on your own data so it learns your style, terminology, or tasks.

A law firm might fine-tune a model on 5,000 of their own contracts so it writes clauses the way they actually write them. Fine-tuning does not stuff new facts into the model reliably — it shapes behavior. If you want the model to answer "what is our 2025 expense policy?" accurately, use RAG, not fine-tuning. If you want it to always reply in your brand voice, or always return JSON in a specific shape, fine-tuning earns its keep.

Grounding

Tying an AI answer to a specific source document so it stops making things up.

A grounded answer comes with a citation: "According to section 4.2 of your employee handbook, bereavement leave is five days." An ungrounded answer is what the model guessed based on training data, which may or may not match your actual handbook. NotebookLM is built entirely around grounding — every sentence in an answer points back to the line in your source where it came from. For enterprise work, grounding is non-negotiable.

Hallucination

When an AI confidently states something that is not actually true.

The canonical case is Mata v. Avianca (S.D.N.Y. 2023). A lawyer asked ChatGPT for case law, and the model invented half a dozen cases with plausible names and citations that did not exist; the attorneys got sanctioned by Judge Castel. Hallucinations are not bugs in the ordinary sense. They come from how these models work: they predict plausible next words, and plausible is not the same as true. Defenses are grounding, citations, retrieval, and not trusting confident prose without checking the source.

Markdown

A lightweight way of writing formatted text — headings, lists, links, tables — using plain characters like # and *.

# Title is a heading. **bold** is bold.- item is a bullet point. That's 90% of it. Markdown is the format this entire site nudges you toward because it is the cleanest, cheapest, most AI-legible way to hand documents to a model. Unlike DOCX or PDF, it is pure text — no hidden layout, no binary blobs, no XML soup. GitHub, ChatGPT, Claude, Obsidian, Notion exports, and almost every AI RAG system treat markdown as the native tongue.

MCP (Model Context Protocol)

An open standard that lets AI assistants pull in data and run tools from external systems without custom glue code.

Before MCP, connecting Claude (or any assistant) to your internal wiki, your file storage, or your database meant writing bespoke integrations for each combination. MCP defines a common shape: a server exposes tools and resources, the assistant discovers and calls them. The same server works for Claude Desktop, for coding agents, for any MCP-aware client. We ship an MCP server that lets Claude convert documents to markdown on the fly — install once, works everywhere.

OCR (Optical Character Recognition)

Technology that looks at an image of text and turns it into actual, searchable characters.

A scanned PDF of a 1987 commercial lease is just pictures of pages. Your computer cannot search it. Copy-paste returns nothing. AI cannot read it. OCR fixes that. Traditional OCR (Tesseract, Abbyy FineReader) works by pattern-matching letter shapes. Modern AI vision models go further and handle handwriting, merged-cell tables, and grainy microfilm scans that break traditional tools. The state of the art in 2026 is no longer about digitizing perfect laser-printed PDFs; it is about making a 1950 handwritten birth certificate useful.

Prompt

Whatever you type into an AI chat. The instruction plus any context you give it.

"Summarize this contract in three bullets" is a prompt. "You are a senior paralegal. Summarize this contract in three bullets, flag any unusual indemnification, return nothing else" is a better prompt. Prompts are the entire user-facing API of a generative model. A good prompt tells the model who it is playing, what the task is, what the output should look like, and what to ignore. Most "the AI gave bad answers" complaints are really prompt problems.

RAG (Retrieval-Augmented Generation)

A pattern where the AI looks up relevant documents first, then writes an answer grounded in what it found.

The acronym is ugly but the idea is plain. You ask "what is our bereavement policy?" The RAG system searches your internal docs, pulls the three chunks most relevant to bereavement, and hands those chunks plus your question to the model. The model writes an answer using only what it was given. This is how ChatGPT's custom GPTs, Claude Projects, NotebookLM, and every enterprise AI chatbot actually work under the hood.

RAG in four steps

User asks: "What's our bereavement policy?"
System embeds the question, finds closest doc chunks
Retrieved chunks + question → sent to the model
Model writes answer, cites chunks

System prompt

The behind-the-scenes instructions that shape how an AI responds, set by whoever built the app.

When you open ChatGPT, you only see the chat box. Invisible above your messages is a system prompt that might say: "You are ChatGPT, a large language model trained by OpenAI. Today's date is X. Never do Y." When you build a Custom GPT, you are writing a system prompt. When you set up a Claude Project with instructions, same thing. Your regular "user" messages come later; the system prompt frames everything.

Temperature

A dial between 0 and 1 (or higher) that controls how creative vs predictable an AI response is.

At temperature 0, the model picks its single most likely next word every time. Same prompt, nearly identical answers. At temperature 1, it samples more broadly — fine for brainstorming names for a coffee shop, terrible for extracting structured data from invoices. If you're using an API for anything resembling extraction, classification, or summarization, start at 0.0–0.3. For creative writing, 0.7–1.0. The ChatGPT web app hides this control; the API exposes it.

Token

A chunk of text — usually a word or piece of a word — that an AI counts for pricing and length limits.

A real example. "The cat sat on the mat" is about 7 tokens, roughly one per word. "Tokenization" is three tokens ("token", "iz", "ation") because the model splits long or unusual words. English averages ~0.75 words per token. Tokens matter because they are the unit of pricing. If Claude costs $3 per million input tokens and you send it a 300,000-token book, that is 90 cents on one query. Cleaner documents (say, markdown instead of the messy HTML of a scraped web page) use fewer tokens for the same content.

Vector database

A database built to store and search embeddings, so you can find text by meaning rather than exact keywords.

Regular databases are good at "find me the row where email = alice@example.com". They're bad at "find me the three paragraphs in my 10,000-document library that are most similar in meaning to this question". Vector databases — Pinecone, Weaviate, Qdrant, pgvector, Chroma — are built for that second query. They store embeddings, they index them for fast nearest-neighbor search, and they return the closest matches in milliseconds. Every serious RAG system has one at its core.

Still confused?

Most of these terms only matter once you start trying to do something specific, such as summarizing a long report, building a custom chatbot, or cutting API costs. If you are just getting started, skip the jargon and head to AI file uploads, which walks through what to do with ChatGPT, Claude, and Gemini without assuming you know any of this. Prefer to start with the format itself? Try the Markdown editor or convert an existing document with the PDF to Markdown converter.