How AI works · the company-knowledge trick
RAG · answering from your documents
30-second gist~30s read
RAG stands for retrieval-augmented generation. It's the trick that lets a generic AI answer from a specific corpus of documents — your company handbook, the school's policy library, a set of legal cases — without retraining the whole model.
The pattern is simple: when you ask a question, a search engine first pulls the most relevant passages from your documents; then the AI writes its answer using just those passages. Most enterprise "AI assistants that know about us" are RAG systems.
If you want more
Why RAG, not fine-tuning?
Three reasons. It's much cheaper than fine-tuning a model. You can update the documents any time without re-training. You can show your sources — the AI cites the passages it used, so the user can click through and verify.
How it can still go wrong
Two failure modes are common. The search step can pull the wrong passages — close-but-not-relevant — and the AI then answers from those. Or the AI can hallucinate inside the retrieved passages, blending real text from your documents with invented detail. The fix is the same as for any AI answer: verify the citation.