Glossary: Glossary: LLM, RAG, embedding, agent, Model Routing, node and more.
The most important terms in business AI, explained briefly and clearly – from LLM, RAG and agent to node, routing, shared data core and Private AI.
In short: The glossary briefly explains key terms so you can compare quotes and technical statements more easily.
Explanation
Key terms for conversations about AI.
The table below explains the terms used on this website and in conversations with us. Where a term has its own page, it is linked.
Terms
Glossary.
| Term | Explanation |
|---|---|
| LLM (large language model) | A language model generates text from learned patterns. It only knows internal data if that data is provided to it. |
| RAG (retrieval-augmented generation) | A method that first searches your own documents for relevant passages and provides them to the model. Answers should name the sources used; the model is not retrained for this. |
| Embedding | A numerical vector that represents the meaning of a text passage. The basis of semantic search in RAG systems. |
| Agent | An AI system that handles a task: it uses tools, decides on intermediate steps, keeps track of its state and obtains approvals at defined points. |
| Human-in-the-loop / approval | A defined point in a workflow where a person confirms before an action is triggered – e.g. before sending or booking. |
| Model Routing | Assigning a task to the right model based on capability, data protection, speed, cost and availability (Swiss AI Factory). |
| AI / reverse gateway | Uniform access for applications to all models – local and external – through one interface. |
| Load balancing | Distributing the load across several nodes or models. |
| Fallback | A backup model or node that takes over if the primary one fails – according to rules that match the data class. |
| Node (AI node) | A compute node with models that receives tasks from the model layer. Additional load may require further components and testing. |
| Private AI | An operating model in which models, knowledge and agents run in an environment controlled by the company – on-premise, private cloud or hybrid. |
| On-premise | Operation on your own servers, on your own premises. |
| Hybrid | Sensitive tasks local, non-critical tasks external – governed by data classes and routing. |
| Data class | Classification of data (public, internal, confidential, particularly sensitive) that determines where a task may run. |
| Shared data core (Shared Business Core) | A central data store for customers, companies, people, projects, documents, users and roles that several modules use. |
| Fine-tuning | Adapting a model with example data for a specific task. Whether RAG or fine-tuning fits depends on the use case. |
| Distillation | A small model learns the behaviour of a large one – for specialised, low-cost tasks. |
| OCR | Text recognition from images and scans. |
| Swiss AI Studio / Agents / Factory | The three layers of the Swiss AI Platform: control, execution, model layer. |
Decision guide
How to use the glossary
- Before a quote: clarify the terms that appear in the quote.
- Before an AI potential assessment: understand LLM, RAG, agent and routing.
- For data protection questions: Private AI, on-premise, hybrid, data class.
Sources
What this article is based on.
- Own working definitions of Swiss AI Agency & Labs; no external study
Further reading
Products, Labs and case studies.
Frequently asked questions
Frequently asked questions: glossary
Is a term missing?
Write to us – we add to the glossary on an ongoing basis.
Next step
Have a specific process assessed.
The AI potential assessment examines the benefits, prerequisites and risks of the process.