Glossary
Also: Retrieval Augmented Generation
A language model is handed relevant excerpts from your own data before answering, instead of relying on its training.
The flow has two halves: retrieval and generation. Answer quality depends almost entirely on the first half. An excellent model with poor retrieval produces confident wrong answers.
The work therefore sits not in the model but in the preparation: how documents are split, which metadata is attached, how permissions are honoured, and how you notice that the retrieved passages do not fit.
How you notice it
Frequently asked
Because the data situation changes, not the model. The prototype has a hundred clean documents, production has ten thousand contradictory ones with permissions. What is missing is almost always the same: splitting along document structure, reranking of results, and a measurement of answer quality.