🔹 RAG (search first, then answer)
What it is: An architecture that makes a neural network look for up-to-date information in external databases before it generates an answer (unlike ordinary generation, which relies on the model’s own memory alone).
How it works: It’s like an open-book exam. You connect the AI to a folder of work documents. When you ask “What’s our holiday policy?”, the AI doesn’t make things up: it finds the specific PDF, reads it and gives you an accurate answer based on your own data.
🔹 Tokens
What they are: The basic units of data that an algorithm breaks text into for processing. They aren’t necessarily whole words: more often they’re groups of letters or single characters.
How it works: The short Ukrainian word “кіт” (cat) is 1 token. But the AI will split the long word “заробіток” (earnings) into “за”, “роб” and “іток” (3 tokens). When you use paid versions of AI, these little pieces of processed information are exactly what you pay for. To give you an idea: 1,000 tokens is roughly a 2-page article.
🔹 Context window
What it is: The maximum amount of information (in tokens) that a model can analyse and hold in its “working memory” at the same time during a single session.
How it works: It’s how much the AI can remember before it forgets the start of the conversation. Modern models have windows of 128,000 tokens or more. That means you can upload an entire Harry Potter book into the chat, and the AI will take every page into account when answering your questions about the text.
🔹 Temperature
What it is: A mathematical parameter of the model (from 0 to 2) that controls how much randomness goes into choosing the next word as the text is generated.
How it works: 🧊 Low (0.1): “Accountant” mode. The AI picks the most likely words. The answer is dry and precise. Ideal for code or working with numbers.
🔥 High (0.8+): “Poet” mode. The AI picks less likely options. The text becomes creative and unusual, but there’s a risk the AI will write nonsense.
🔹 Fine-tuning
What it is: The process of giving an already trained base neural network extra training on highly specialised datasets or on your own data.
How it works: It’s like sending a medical student off to specialise as a surgeon. You take an ordinary neural network and “feed” it 500 of your best posts. After that, the AI starts writing in exactly your style, using your signature words.
🔹 Embeddings
What they are: Vector (numerical) representations of text or images. They’re a mathematical way of turning words into coordinates on a multidimensional map.
How it works: AI doesn’t understand letters. But thanks to embeddings, the words “king” and “queen” sit close to each other on its mathematical map. So when you search a shop for “winter clothes”, it will show you “puffer jackets” and “hats”, even though you never typed those words. It understands meaning rather than just matching letters.
🔹 Zero-shot and few-shot (how to set tasks)
What they are: Ways of building prompts (your requests) that determine how much context you give the model before it carries out a task.
How it works: 🎯 Zero-shot (from scratch): You write “Write a post about cucumbers.” The AI generates it with nothing to go on.
📝 Few-shot (with examples): You write “Here’s an example of a post about courgettes. Here’s one about tomatoes. Now write a post about cucumbers in the same style.” The result will be many times more accurate!
🔹 Parameters
What they are: A model’s internal settings (numerical weights), which are formed during training and determine how strong the connections between artificial neurons are. Essentially, they reflect how much “knowledge” and complexity a neural network has.
How it works: When you read about an “8B model” (8 billion parameters), that’s a lightweight program you can run even on a powerful smartphone. A “400B model”, on the other hand, is a giant that needs entire rooms full of servers to run, but it can solve PhD-level problems.
