Part 1 of 5 — RAG Explained
The model passed the interview. It failed the handbook.
General-purpose AI is trained on public text up to a cutoff. It does not hold your employee handbook, your ticket history, or the clause in contract 447-B unless you put those documents in front of it at ask-time. Without that step, fluency masquerades as knowledge. That is not a personality flaw. It is architecture.
Enterprise AI starts when you stop asking “how smart is the model?” and start asking “what corpus is this answer allowed to cite?”
Training knowledge vs approved corpus
Two different sources of “truth” get conflated in sales decks:
Parametric knowledge is whatever weights absorbed from pre-training. Useful for grammar, common patterns, and public facts — with errors.
Approved corpus is what you index: wikis, PDFs, tickets, runbooks, CRM exports. RAG retrieves from the second bucket and conditions generation on what came back.
The shift from general AI to practical enterprise AI is mostly that second bucket: who may read it, how often it is re-indexed, and what happens when retrieval returns the wrong chunk.
Hallucination vs grounding
Hallucination here means the model states a detail that is not supported by the context you intended it to use. Grounding means the next tokens are tied to retrieved passages you control.
RAG does not eliminate hallucination. It moves the failure mode. Wrong chunk, confident answer. Stale index, outdated policy. Empty retrieval, model fills the gap anyway. Those are ops problems — which Part 3 takes up — not reasons to pretend retrieval is optional.
Why now
Three pressures push RAG from nice-to-have to default:
- Compliance — Answers about HR, finance, or customer data need traceable sources.
- Freshness — Product names, prices, and policies change faster than retraining cycles.
- Scope — Internal knowledge is not on the public internet the base model saw.
You do not need a agent loop to get value from the first fix. Index, retrieve, generate is enough for many Q&A and draft workflows. Part 5 is where agents enter — and where they should not.
Example
A support lead asks: “Do we still offer 90-day returns for EU customers?”
Without your corpus, the model may invent a plausible policy from e-commerce blogs. With RAG over your current policy PDF, retrieval either surfaces the EU section or returns nothing — and a well-written prompt can say “not found in approved docs” instead of guessing.
That is not magic. It is retrieve-then-generate with an approved store. Part 2 draws the diagram.
Takeaway: Enterprise RAG begins with a corpus you approve — not with a bigger base model.
Hub: RAG Explained
Part 2: Architecture