What is RAG in generative AI?

Documents are split into passages and indexed. A question is used to retrieve a handful of passages. Those passages are placed in the prompt, and the model is told to answer from them and to say when they are not enough. The index is not the product. The product is the answer a specific person was allowed to receive, with a way back to the source.

Vectors are a common way to search by meaning. Filters on department, date, and document type still do a lot of the work. A RAG system that is “just a vector database” will mix last year’s policy with this year’s.

The right paragraph was never retrieved, so the model fills the gap. Two documents disagree and the model blends them. A scanned PDF was indexed as garbage. A user retrieves a file their colleague should not have shared. None of these are fixed by a more confident prompt. They are fixed by the corpus, the permissions, and a test set of real questions.

Fine-tuning does not replace this. Fine-tuning changes style or behaviour. It does not keep a living policy library up to date. Most companies should spend the first budget on retrieval and evaluation.

One collection with an owner. Permissions copied from the source. Citations on every answer. A refusal when retrieval is weak. A pilot group and a list of failures. The knowledge assistant sample shows that shape. It is a concept, not a client claim. The RAG development page is how we deliver it.