RAG

RAG adds a search step to a language model, retrieving matching passages from a company's own document collection before an answer is generated. Those passages are handed to the model together with the question, so the answer is grounded in specific, checked sources rather than only what the model learned generically during training. The approach exists because a model's training knowledge is time-limited and generic, while business documents need to stay current and specific.

In practice, a request first triggers a search across an indexed set of documents — manuals, contracts, internal wikis — before the language model formulates anything. The retrieved passages often appear as a source reference beneath the answer, so it stays traceable where a statement came from. When a document changes, the next lookup automatically reflects that, with no need to retrain the model.

A limiting case: RAG substantially reduces the risk of fabricated answers, but it does not eliminate it. If the needed information isn't in any indexed document, or the search misses it, the model can still produce something that sounds plausible but isn't backed by a source. A RAG answer is never better than the quality and completeness of the documents behind it.