LLM

What an LLM is

LLM stands for large language model. It is a neural network trained on a very large amount of text that predicts the probability of the next piece of text. A sentence is therefore not looked up but calculated piece by piece. That explains the most important property of these systems: they write fluently and convincingly without owning a database of facts.

The building blocks are called tokens, meaning parts of words rather than words. One token is roughly three to four characters depending on the language. Computing power and cost are measured in tokens, which is why the length of input and answer is a direct cost factor.

How it produces usable results

The models do not answer from stored documents but from patterns. For tasks that need accuracy, a second component is added. With RAG, matching passages are retrieved from your own sources and handed to the model so that the answer rests on those passages. Tool calls belong to the same idea: the model may trigger a function, for instance look up an order, and works the result into the answer.

How much text the model can consider at once is described by the context window. It is the upper limit for input and output together and therefore the limiting factor with long product data or contracts.

Why it hallucinates

• Patterns instead of knowledge: when a detail is missing it is not reported as missing but filled in plausibly.
• Outdated state: a model only knows what was public up to its training date.
• A convincing tone: the answer sounds confident even when it is wrong. That is the real danger in operations.
• No sources: without RAG or tools the model cannot show where a statement comes from.

What that means for a shop

Language models are good at wording, summaries, translations and rewording existing material. They are not a source for prices, stock, delivery times or legal statements. The practical rule is therefore to give the model the facts and leave the decision with people. Anyone publishing model output needs a review step, and every paragraph needs a person responsible for it.

Frequent questions

• Does a language model think? It calculates probabilities. The result looks thought through but is statistics.
• Is my input confidential? Only if the provider commits to it contractually, and only on a plan that excludes training on your data.
• Can I feed it my own data? Yes, with RAG without training, with fine tuning with training. For most shops RAG is the faster and cheaper route.