# RAG in Plain Language: How Bots Answer From Your Documents Ask a plain AI chatbot about your refund policy and it will answer confidently — using average knowledge of the internet, which is another way of saying it will invent your policy. That confidence-without-knowledge problem has a standard fix, and it is simpler than the acronym suggests. It is called RAG: Retrieval-Augmented Generation. Here is the whole idea in one sentence — **the bot first searches your documents, then writes its answer using only what it found.** ## The three steps, no math ### 1. Retrieve Your documents (price lists, policies, manuals, FAQs) are indexed once — chopped into passages and stored so they can be searched by *meaning*, not just keywords. When a customer asks "can I return a used item?", the system pulls the two or three passages that actually discuss returns. ### 2. Augment Those passages are attached to the customer's question as context. The instruction to the model is effectively: *answer using these passages; if they don't contain the answer, say so.* ### 3. Generate Only now does the model write — grounded in the retrieved text. Good implementations go one step further and attach the source, so every answer carries its receipt. ## Why this matters more than "smarter models" A bigger model still doesn't know your prices from last Tuesday. RAG gets you three things raw models can't: - **Correctness about your business** — answers come from your documents, including last week's changes. - **Citations** — customers (and you) can check where an answer came from. [NabuDoc](/products/nabudoc) is built around exactly this. - **Privacy** — your documents stay in your deployment; the model reads passages, not your whole archive. ## Where you already meet RAG in the Nabux suite - [NabuDesk](/products/nabudesk) answers support tickets from your documentation — with an 85% auto-resolution target and a clean handoff when the answer isn't there. - [NabuDoc](/products/nabudoc) is the "ask your documents" tool: policies, manuals, contracts — cited answers in Arabic and English. - [NabuQuiz](/products/nabuquiz) even uses your material to generate assessments, so training content and the exam always match. ## When RAG is the wrong tool If the answer must come from judgment ("should we enter this market?") or from data that lives in someone's head, retrieval won't save you. RAG excels at *documented* knowledge — policies, specs, procedures, prices. Start there; [we can help you scope it](/contact).