What is RAG (retrieval-augmented generation)?
Plain-English definition · Updated 2026-10-06. Numbers dated; verify with the vendor.
AITop is an independent guide. Prices and features change fast — always check the vendor's page.
RAG is "look it up, then answer." Instead of relying on what the model memorized in training, the tool first retrieves relevant passages — from a search index, your company wiki, your PDF pile — hands them to the model, and generates an answer grounded in them, usually with citations. It is the architecture behind AI search engines, "chat with your documents" products and most enterprise assistants. RAG doesn't eliminate hallucination; it gives you the receipts to catch it.
What it actually means
Break the acronym apart and the whole idea falls out. Retrieval: a search step finds the handful of passages most relevant to your question. Augmented: those passages are injected into the model's context alongside your question — the model reads them like an open-book exam. Generation: the model writes an answer based on that material, ideally quoting and citing as it goes.
The crucial insight is that a model's training knowledge is frozen and unattributable, while retrieval is fresh and checkable. A model trained in March cannot know about an October price change; a RAG pipeline that searched yesterday can. That is why the products where being current matters — AI search, support bots, document analysis — are all RAG underneath, whatever the landing page calls it.
Why it matters when you're picking a tool
Because RAG quality is a product differentiator hiding behind a shared buzzword. Two tools can both say "grounded in your documents" and differ enormously: one retrieves the right five passages and cites them precisely; the other retrieves sloppily and cites confidently. The model gets the glory, but the retriever is where quality lives — garbage retrieval produces confident garbage no matter how good the model is. When you evaluate any "AI search" or "chat with your files" tool, you are mostly evaluating a search engine wearing a language model's jacket.
The 2026 reality check
AI search made RAG the default consumer-facing architecture: Google AI Mode generates directly on top of Google's index, Perplexity built citations into the product itself, and ChatGPT Search bolts retrieval onto a chatbot — three genuinely different implementations we compare in our AI search showdown. The arms race has moved to trust: ad slots inside AI answers, AI-search-optimisation spam planting fake sources, and citation formats that make sources harder to verify all appeared in 2025–2026. The healthy user habit is now the same as the healthy architecture: click the citation, read the passage, confirm it supports the claim.
Quick checklist
- Click citations before trusting them — confirm the linked source actually supports the sentence it's attached to.
- Test freshness on purpose: ask about something that changed this month and see what the tool's index knows.
- For document chat, test retrieval with a detail buried mid-document, not just the summary page.
- Prefer tools that cite every factual claim by default over tools that cite only when pressed.
Where you'll hit it
RAG is the load-bearing wall of Perplexity vs Google AI Mode vs ChatGPT Search and Perplexity vs ChatGPT, and it is the feature that makes AI legal tools trustworthy enough for filing work. Related terms: context window, AI hallucination — or back to the full glossary.
The bottom line
RAG is the difference between a model that knows and a tool that checks. It is not magic and not optional — it is the open-book exam every serious AI product gives its model before answering. Judge the retrieval, click the citations, and the buzzword collapses back into the only question that matters: can this tool show its work?