Yuval Avidani
Author
If you're already using ChatGPT or Claude every now and then, but feel like you're just scratching the surface, this glossary is exactly for you. There's a huge gap between "asking a question and getting an answer" and actually squeezing real value out of AI at work, and that gap is made mostly of words: context window, embeddings, RAG, agent, MCP. Let's break them down one by one, and by the end we'll see how they turn into uses that genuinely change your workday.
How a model actually "thinks": guessing the next token
Let's start with the engine. At its core, an LLM does one thing: it takes in text, and guesses the most likely token (a word-piece) to come next, over and over. That's it. There's no "understanding" in the human sense inside there, just remarkably precise probabilistic prediction trained on a massive pile of text. To feel this with your own hands, play with this tiny engine that guesses the next word:
The cool part is that out of this simple thing, guessing token after token, an ability emerges to write code, summarize a legal document, and translate. That's exactly why even a very smart model can sometimes be confidently wrong: it doesn't "know" a fact, it generates the most probable continuation.
Tokens, context window, and parameters
Three numbers that determine a lot. A token is a chunk of text (roughly four characters in English, and in Hebrew usually fewer characters per token, meaning Hebrew is "more expensive" in tokens). Price is measured by tokens, and memory is measured by tokens too.
And here's the concept most worth internalizing. The context window is the amount of tokens the model can hold "in its head" at once, your entire prompt plus the entire response. Think of it like a desk: the bigger it is, the more documents you can spread out on it at once, but it's not infinite. Once you cross the limit, older information in the conversation "falls off the desk" and the model starts forgetting the beginning. That's why in long conversations it pays to summarize and start fresh.
And the third number is parameters. Parameters are the internal "knobs" the model tuned during training, and their count (in billions) is a rough gauge of the model's size and capability. More parameters usually means a smarter model, but also more expensive and slower to run, which is why models come in different sizes for different tasks.
How you let AI access your data: embeddings and RAG
So far we've talked about the knowledge the model learned in advance. But the real power kicks in when you connect it to your own data. This is where two concepts come in. An embedding is a way of turning text into a list of numbers that represents its meaning, so that texts with similar meaning get numbers that are close together. This is what enables "semantic search": searching by intent rather than exact words, so "how do I cancel a subscription" also finds a document that says "terminating an agreement."
And RAG is built on top of that. RAG, short for Retrieval-Augmented Generation, is a technique where before the model answers, it automatically pulls the relevant chunks of your documents and attaches them to the prompt, so the answer is grounded in your data rather than just what the model remembers. Think of it like a brilliant student who was given access to exactly the right pages in the book right before the exam. This is the common solution for a company assistant: it answers accurately, stays up to date, and can point to the source.
RAG vs. Fine-tuning vs. Prompting: when to use what
There are three ways to adapt AI to your task, and people mix them up constantly:
In short, and this matters: a good prompt is the fastest and cheapest, and solves most cases; RAG gives the model access to your sources; and fine-tuning shapes a fixed style and behavior, but is more expensive and slower. Most people think they need fine-tuning, but in practice 90 percent of cases are solved with a good prompt or with RAG.
AI agents, Tool Use, and MCP: when AI doesn't just talk, it acts
And here's the biggest leap of recent years. An AI agent is a model that doesn't just write an answer, but operates in a loop: it thinks, picks a tool, runs it, reads the result, and decides on the next step, until the task is complete. The capability behind this is called Tool Use. Tool Use is what lets a model perform actions in the real world: search the web, read a file, send an email, or pull data from a system. Without it, the model only talks; with it, the model executes.
And the link that connects it all is called MCP. MCP, short for Model Context Protocol, is an open standard (from Anthropic) that defines a unified language through which a model can connect to external tools and data sources, like Gmail, Google Drive, or your publishing system. Think of it like a universal USB port for AI: instead of building a dedicated connection for every service, you connect through the same standard. This is what turns the idea of "let AI look through my emails and find my insurance policy" from a wish into a one-minute action.
Reasoning models, temperature, and multimodal
Three more terms worth knowing. A reasoning model is a model trained to "think before it answers," breaking a problem down into steps internally, which makes it much stronger at math, code, and planning, but slower and more expensive. Alongside it, ==temperature is a dial that controls how "creative" versus "predictable" the model is: low = consistent, careful answers (good for data), high = more variety and surprise (good for ideas). And finally, multimodal means a model that understands and generates more than one type of media, text plus image, voice, or video together.== This is what lets you upload a screenshot and ask about it, or generate an ad with visuals and narration together.
How to choose a model: quarter of the value
With so many models out there, the practical question isn't "who's the smartest" but "who gives me the most smarts per shekel." The way I like to look at this is a graph of intelligence vs. price:
What you're looking for is the top-right corner: lots of smarts, little money (on an RTL graph, cheap sits on the right). Models like Sonnet 5 and GLM-5.2 stand out on value-for-money, while the strongest models at the top cost significantly more per token. And to actually feel the money in practice, here's a little calculator: how much does a typical call cost, by model?
My rule: start with the cheap, fast model, and only upgrade to the strong one for tasks that actually require it. Most day-to-day work doesn't need the most expensive model.
Real value at work: the uses that change your day
Now let's bring it all down to earth, because this is where it actually turns into money. Here are real advanced uses. Inside email: connect AI to your inbox (via MCP) and ask "find my insurance policy, my bank statement, and the last three invoices, and summarize them for me in a table," including pulling data out of PDFs and slide decks. In marketing: connect your Meta ads account and ask "analyze this month's campaigns, where is ROAS lowest, and what should I turn off." In research: an organized competitor review in minutes instead of hours, and here a precise prompt makes all the difference:
And there's a lot more. Build an entire app through conversation ("vibe coding"), even with zero coding background, just by describing what you want and running it right away. Prepare a presentation that follows your brand guidelines (colors, fonts, tone) from a short brief. Write and structure a fair, well-reasoned employee performance review from a set of bullet points you collected. And even produce a whole ad, visual, or podcast, text, voice, and image together, from a single idea. What all of these share: we define a clear goal and give it access to the data and tools, and the AI does the grunt work.
Bottom line, and my take
So let's wrap up the map. The model guesses tokens; the context window is its temporary memory; embeddings and RAG connect it to our data; an agent with Tool Use and MCP lets it act in the world; and choosing the right model is a balance between smarts and price. Once this map is clear, you stop "playing with a chatbot" and start building processes.
In my view, the difference between a regular user and a power user isn't secret knowledge, it's a habit: define a precise goal, give it context and tools, and check the output. The catch, and it's fair to say this: the more access and autonomy you give AI, the greater the need for oversight, security, and verification, because an agent that makes a mistake does it fast and at scale. And all the numbers here (the intelligence index, the prices) are accurate as of July 2026 and change all the time, so it's always worth double-checking against the official pricing page.
So here's the question I'll leave you with: if AI can already not just answer but act, which recurring, annoying process at your job, the kind you do manually every single week, would you most want to turn into an agent that just does it for you?