Glossary

Definitions for terms that come up across multiple courses.

Context window: The maximum amount of text a language model can take in — and keep track of — at one time, usually measured in tokens (small chunks of text, roughly parts of words). It covers everything in a single exchange combined: the prompt/instructions, any text pasted in, the conversation history, and the model's own response. If the total goes over that limit, older or excess text is cut off and the model can no longer see it.
Embedding: A way of converting a word, sentence, or document into a list of numbers that captures what it means, not just which letters or words it contains. Two pieces of text with similar meaning — like "How do I get a refund?" and "What's your return policy?" — end up with similar number lists, even though they don't share any of the same words. It's a bit like giving every idea a GPS coordinate: ideas that mean similar things end up as nearby points on the map, so a computer can find "nearby" ideas the same way it finds nearby cities.
Embedding Model: The AI tool that actually creates embeddings — it reads a piece of text and produces the list of numbers representing its meaning. Think of it as a translator, except instead of translating between languages, it translates between human meaning and the language of numbers that a computer can measure and compare.
Large Language Model: An AI system trained on enormous amounts of text — books, websites, articles — until it learns to predict what word is likely to come next in a sentence. That one skill, repeated over and over, turns out to be enough for it to write essays, answer questions, summarize documents, and hold a conversation. Tools like ChatGPT and Claude are large language models — "large" refers to the huge scale of both the training text and the model itself. Think of it like an extremely well-read assistant who has skimmed a giant slice of the internet and, from all that reading, developed a strong intuition for what a good next sentence sounds like.
Prompt Engineering: The skill of carefully wording the instructions or questions you give an AI so it produces a better, more useful response. Asking a librarian "Do you have any books?" gets a very different answer than "Do you have any mystery novels set in Japan?" — the same is true for AI. A vague prompt like "Write about dogs" gets a vague result, while a specific one like "Write a 100-word bedtime story about a lost puppy finding its way home" gets you much closer to what you actually wanted, without changing anything about the AI itself.
RAG (Retrieval-Augmented Generation): A technique that lets an AI look up real, current information before it answers a question, instead of relying only on what it memorized during training. It works like an open-book exam: rather than answering purely from memory, the AI is allowed to quickly search a set of documents, pull out the relevant pages, and write its answer based on what it just read. This is how AI assistants can answer questions about a company's internal documents or yesterday's news, even though that information didn't exist yet when the AI was originally trained.
Semantic Search: A style of search that looks for results with a similar meaning to your question, instead of only matching the exact words you typed. A traditional keyword search for "how do I get my money back" might miss a page titled "Refund Policy" simply because that page never uses the word "money." Semantic search understands that the two phrases mean roughly the same thing, so it can find the right answer even when the wording is completely different — like a helpful store employee who understands what you're asking for, even if you don't use the exact word printed on the receipt.
Token(s): The small pieces of text a language model actually reads and generates, one at a time. A token is usually smaller than a whole word — often a word, a common word-fragment, or a single punctuation mark (for example, "unbelievable" might be split into "un," "believ," and "able"). Models don't see letters or full sentences the way people do; they break everything into tokens first, like chopping a sentence into puzzle pieces before working with it. This matters day to day because AI tools are often priced and limited by the number of tokens processed — it's why a context window is measured in tokens rather than in words or pages.
Vector: The list of numbers that comes out of an embedding model. Each vector acts like a coordinate that pinpoints where a piece of text "lives" in an imaginary space of meaning — similar to how a street address pinpoints where a building sits on a map. Text with similar meaning gets a vector that sits nearby; unrelated text ends up far away.
Weights: The millions or billions of internal numbers inside a language model that quietly encode everything it has learned — facts, grammar, writing style, reasoning habits. Picture a massive control panel covered in tiny dials: during training, the model makes a guess, checks how wrong it was, and nudges some of those dials slightly to do better next time. Repeat that billions of times across huge amounts of text, and the final dial settings — the weights — are what make the model behave the way it does. "Training" or "fine-tuning" a model means adjusting these dials; techniques like prompt engineering and RAG change what you feed the model without ever touching a single one.