This list is sorted by thematic order, not strictly alphabetically — this is so the explanations build upon one another. You can find any term quickly using your browser’s table of contents (Ctrl+F).
Models and Architecture
- Model
- The actual “AI file.” GPT-4, Llama 3.1, Claude 3 are models. Determines what the AI knows and can do.
- LLM (Large Language Model)
- Large language model. A model that was trained on text and generates, understands, and translates text.
- Parameter
- Internal adjustment knobs of the model. More parameters usually mean more knowledge, but also more hardware requirements. Common sizes today: 7B, 70B, 405B (billions).
- Open Weight
- A model whose weights (i.e., the model file) are freely available. Examples: Llama, Mistral, Qwen. Can run locally — in contrast to GPT-4 or Claude.
- Quantization
- Shrinking the model so that it runs on less hardware. Saves RAM and storage, costs little quality. Common: 4-bit quantization halves the storage requirement with minimal loss of accuracy.
- Fine-Tuning
- Further training a pre-trained model with your own data. Labor-intensive, only worthwhile in rare cases — RAG is usually the better choice for SMEs.
Queries and Answers
- Token
- Building blocks into which text is decomposed. A German word like “Werkstattauftrag” is decomposed into 3-4 tokens. A page of text has approx. 500-700 tokens.
- Context / Context Window
- The amount of tokens that the model can read at once. Modern models: 8,000 to 128,000 tokens.
- Prompt
- The instruction to the AI. Example: “Write a short summary of this order.”
- System Prompt
- The invisible fundamental rule for the model. Example: “You are a polite assistant in an automotive workshop. Answer concisely and in German.”
- Inference
- The process in which the AI answers a question (in contrast to training). Inference requires GPU power and electricity — training requires both x 1000.
Reliability and Security
- Hallucination
- The AI invents plausible-sounding but false content. See dedicated article.
- Bias (Distortion)
- If the training data was one-sided, the answers are one-sided. Classic example: AI prefers job titles in one gender.
- Jailbreak
- An attempt to move the AI to provide answers through tricky prompts that it should actually refuse. Relevance in business is usually low, but it exists.
Integrating Own Knowledge
- Embedding
- Mathematical “fingerprint” of a text. Two similar texts have similar embeddings — therefore RAG finds the matching passages for a question.
- Vector Database
- Stores embeddings and searches for similar vectors in milliseconds. Examples: Qdrant, Chroma, pgvector.
- RAG (Retrieval-Augmented Generation)
- Connecting AI with your own documents. The most important technique for useful business AI. See dedicated article.
Tools and Automation
- Tool Use
- The AI uses external tools — calculator, database, web search, email API. This makes it more reliable for tasks that require accuracy.
- Agent
- An AI that processes several steps independently. Example: “Read the email, create an offer, send it for approval.” Connects tool use with multi-step logic.
- Workflow Automation
- Automating several tasks in a chain — often with tools like n8n, Zapier, or Make. AI is one of several steps in this process.
Hardware
- GPU (Graphics Card)
- Not for gaming here, but as a computing chip for AI. Professional GPUs (NVIDIA H100, RTX 6000 Ada) are today the standard for local AI workstations.
- VRAM
- Memory on the GPU. Determines how large a model can be. Rule of thumb: 1 billion parameters require approx. 1 GB VRAM (4-bit quantized).
Term missing or unclear?
Write to us briefly — we are continuously updating the glossary.
Schedule an initial consultation
Frequently asked questions
What is a token?
A German word like "Werkstattauftrag" is broken down into 3-4 tokens. One page of text has approx. 500-700 tokens.
Is RAG or Fine-Tuning the better choice for SMEs?
RAG is usually the better choice for SMEs.
What is a hallucination?
The AI invents plausible-sounding but false content.
What is bias?
If the training data was one-sided, the answers are one-sided.
Legal notice and privacy policy: Impressum & Datenschutz (German only; the German version is binding).