KI-Glossar

Klare Definitionen des KI-Vokabulars aus Nachrichten, Verträgen und Produktdokumentation.

A

AI agent: An AI agent is a language model connected to tools it can call, given a goal rather than a single question, and allowed to loop — calling a tool, reading the result, deciding the next step — until it judges the goal met or hits a limit. The loop is what distinguishes it from a chatbot.
Alignment: Alignment is the process of shaping a trained model's behaviour so it follows instructions, tells the truth as best it can, and declines harmful requests. It happens after pre-training, mainly through supervised fine-tuning on demonstrations and reinforcement learning from human or AI feedback.

Benchmark: A benchmark is a fixed set of tasks with known answers, used to score and compare models. Public benchmarks give a rough capability ranking but are heavily optimised for, vulnerable to training-data contamination, and shaped like short well-specified problems rather than real production work.

Context window: The context window is the maximum number of tokens a model can process in a single request. It holds the system prompt, tool definitions, conversation history, any attached documents and the generated output. Anything outside the window is invisible to the model.

Distillation: Distillation trains a smaller student model on the outputs of a larger teacher model, so the student reproduces much of the teacher's behaviour at far lower cost. It is the main reason small models improved so quickly, and it is restricted or prohibited by many commercial API terms.

Embedding: An embedding is a fixed-length vector of numbers produced by a model to represent a piece of text, image or audio. Items with similar meaning end up near each other in that vector space, which is what makes semantic search, clustering and recommendation possible.

Fine-tuning: Fine-tuning continues training a pre-trained model on a smaller, task-specific dataset, adjusting its weights so it reliably produces a particular format, tone or narrow skill. It is a poor way to teach facts, which change and are better supplied by retrieval at request time.

Hallucination: A hallucination is output that is fluent and confident but factually wrong or unsupported — an invented citation, a non-existent function, a plausible number with no source. It happens because the model generates statistically likely continuations, and a likely-sounding claim is not necessarily a true one.

Inference: Inference is the act of running a trained model on new input to produce output, as opposed to training, which produces the model in the first place. Training is a one-off cost; inference repeats for every request, and over a widely used model's life it dominates both spend and energy consumption.

Large language model (LLM): A large language model is a neural network with billions of parameters, trained on large text corpora to predict the next token in a sequence. That single objective, at sufficient scale, yields the ability to answer questions, write code, translate and summarise — none of which were trained for directly.

Mixture of experts (MoE): Mixture of experts splits parts of a network into many parallel sub-networks and uses a router to send each token through only a few of them. Total parameter count stays large while the computation per token stays small, which is why several very large models are cheaper to serve than their size implies.
Model Context Protocol (MCP): The Model Context Protocol is an open standard that defines how an AI application discovers and calls external tools, resources and prompts. A server implements it once and any compatible client can use it, replacing per-product custom connectors.
Multimodal model: A multimodal model can process more than one type of input in the same request, most commonly text together with images, and increasingly audio and video. Non-text inputs are converted into the same internal representation as tokens, which is why a single image can consume a substantial share of the context window.

Open weights: Open weights means the trained parameter file is published for download, so you can run, inspect and fine-tune the model on your own hardware. The licence attached decides what you may legally do with it, and many open-weight models ship under custom licences that do not meet the open source definition.

Prompt injection: Prompt injection is an attack in which instructions are embedded in content a model processes — a web page, an email, a document, a code comment — and the model follows them as though they came from its operator. There is no reliable way for a model to distinguish trusted instructions from text it was asked to read.

Quantisation: Quantisation reduces the numerical precision of a model's weights — for example from 16-bit floating point to 8-bit or 4-bit integers — so the model occupies less memory and runs faster. Quality loss is small at 8-bit and becomes noticeable, though often acceptable, at 4-bit.

Retrieval-augmented generation (RAG): Retrieval-augmented generation searches a document collection for passages relevant to a question, inserts them into the model's prompt, and asks the model to answer using them. The model's weights are unchanged; the knowledge arrives as context at request time, so updating a document updates the answer immediately.

Token: A token is the smallest unit of text a language model processes: a common word fragment produced by a tokeniser. In English one token averages roughly four characters, or about 0.75 words. Context limits and API pricing are both counted in tokens, not words.