RAG (Retrieval Augmented Generation): definition and framing
Definition
RAG stands for Retrieval Augmented Generation. Before answering, a language model is handed matching passages from your approved documents and relies on them instead of general training knowledge. The answer then rests on current company knowledge and can name its source. Comparable to a case handler who opens the right file before giving information.
Why the Term Matters
For a roadmap, RAG is often the point where the viability of an AI initiative is decided. As a neutral advisor we first check whether the needed knowledge is sufficiently ordered, who may see which sources, and whether a verifiable answer is legally required. Who understands RAG cleanly does not buy a system prematurely, but first confirms that the data situation supports the approach.
When It Is Relevant
- Answers should rest on your own, current knowledge, not on model assumptions.
- The origin of a statement must be traceably documented.
- The knowledge changes, such as price lists, policies or technical data.
- Access rights matter, because not every role may see every source.
When Not
When only style or wording is at stake and no current company knowledge is needed, RAG is the wrong lever. For a very narrow, always identical task, fine-tuning may fit better. And without reasonably sorted sources, RAG only answers as well as the documents it can find. Then data preparation comes before the build in the roadmap.
Industries That Use It
- Technical wholesale: answer questions on product data and supply terms from your own catalogue, with a reference to the data source.
- Utilities and energy providers: give guidance on tariffs and grid rules based on the version currently in force.
- Tax and audit advisory: research across mandate and statute texts with a clear reference for each statement.
- Property management: clarify questions on assets, contracts and house rules from your portfolio with a verifiable source.
Typical Use Cases
- An internal assistant that returns answers linked to the source file used.
- A search across contracts that cites the matching clause instead of paraphrasing it.
- An onboarding helper that guides new staff along the applicable policies.
- A proposal assistant that pulls specifications from the product catalogue.
- An assistant that hands off honestly to a human when no source fits.
Decision Guide
| Signal | RAG fits | Alternative |
|---|---|---|
| Current company knowledge needed | RAG with an ordered knowledge base | fine-tuning adds no knowledge |
| Evidence required per answer | RAG with a source reference | plain prompting rarely enough |
| Only tone or format | rather not | fine-tuning or prompt |
| Sources are chaotic | clarify the data first | preparation before the build |
Example
A wholesaler wants to give internal sales and customer service faster access to product knowledge. In the potential analysis we check which catalogues and price lists are current and legally cleared, and we frame the case in the EU AI Act. Only when the data situation holds do we recommend building with a partner. This describes the method, not an asserted client metric.
Service Connection
Whether RAG fits your data is clarified by the AI readiness check or the AI potential analysis. For legal questions on traceability, the EU AI Act consulting helps.
Sources
The description follows the established RAG pattern from the technical literature. Quality, permissions and data protection depend on sources, architecture and the chosen model. Statements on the EU AI Act are general orientation and not legal advice and must be checked case by case.