RAG (retrieval-augmented generation)
RAG is a way of using a language model in which the system first searches the company's documents for matching passages and only then asks the model to answer from them. The answer comes with a link to the source, so it can be checked.
On its own a model knows nothing about your price lists, customer contracts, work instructions or order history. Training a model on that data is expensive and goes stale with the first price change. RAG reverses the order: the documents stay where they are, the system searches them on every question, and a corrected procedure takes effect as soon as it is re-indexed.
The mechanism is simple. Documents are split into passages and indexed, the user's question becomes a search, and the few best-matching passages go to the model with an instruction to answer only from them. The answer carries the document and the part of it the content came from, so the employee can click through and read the original.
The quality of such an assistant depends on the search more than on the model, and that is where most of the implementation work sits. Scans with no text layer, three versions of the same instruction in different folders, a price list in a spreadsheet with merged cells: with sources like these the model gets the wrong passage and answers from it, again in a confident tone. Permissions matter too, because the index has to know who is allowed to see which document.
You will recognise the problem when questions like "what discount does this customer get" or "what is the complaints procedure" all land on one person, because only they know which file to open, and the whole company notices when they take leave.