When an AI assistant answers using manuals, procedures or company documents, you often hear that it has been “trained on our data”. In most RAG-based projects, however, that is not what happens.
RAG means retrieving before generating
RAG stands for Retrieval-Augmented Generation. The idea is simple: before asking the model to formulate an answer, the system searches authorized sources for the content most relevant to the question. That content is then provided to the model as context.
The process in five steps
- The user asks a question.
- The system searches for relevant information in the available repositories.
- It selects relevant documents or fragments.
- It inserts them into the context sent to the model.
- The model generates the answer based on the question and the retrieved material.
Question → Search → Relevant content → LLM → Answer
Why it isn’t training
In training, a model’s internal parameters are modified through a learning process. In RAG, by contrast, documents are retrieved when needed and temporarily inserted into the context. The underlying model can remain unchanged.
Why it matters to companies
This approach makes it possible to update sources without continuously retraining the model, and to apply access controls to content. But it does not eliminate the quality problem: if the search retrieves outdated or incorrect documents, the answer can be wrong too.
In short
Giving documents to a chatbot doesn’t necessarily mean teaching them to it. With RAG, the system consults them at the right moment. It is a technical difference with important consequences for updating, governance, security and answer quality.



Leave a Reply