Which Apps and Products Use RAG?
Far more than you’d guess — and you almost certainly used one this week. Any product that answers questions from a specific body of material rather than from a model’s memory is doing Retrieval-Augmented Generation, whether the marketing says so or not. That covers AI search engines, customer support chatbots, “chat with your PDF” tools, coding assistants that read your repository, note-taking apps with an ask-your-notes feature, and virtually every internal company help desk built in the last couple of years.
Feature names and product line-ups change constantly in this field, so what follows is deliberately about patterns rather than a checklist of vendors. Learn the pattern and you can identify RAG in a product you’ve never heard of.
The pattern to look for
Four tells, none requiring technical knowledge:
It cites sources. Footnotes, link cards, “from your documents,” page numbers. Citations are only possible when the system knows which passage it used, which means retrieval happened.
It shows a searching or reading step. A “searching the web…” or “reviewing 12 files…” status line is the retrieval step announcing itself.
It knows things newer than the model. Correct discussion of last week’s events means it looked something up.
It knows things private to you. Your files, your company, your codebase. The model was never trained on any of that, so if it’s accurate, it read them.
Any one of those means retrieval. Now the categories.
AI search engines and answer engines
The purest consumer example. You ask a question, the product searches the web, opens the top results, and writes a synthesized answer with citations. That’s retrieve-augment-generate with the web as the library, and it’s why every answer arrives wearing footnotes.
This category includes the AI answer panels appearing at the top of mainstream search results, standalone conversational search products, and the browsing modes inside general assistants. The distinguishing feature is that the library is public and enormous, so the hard part is judging which of a billion pages to trust.
General assistants, in some of their modes
Mainstream chatbots are hybrids: sometimes closed-book, sometimes retrieval-backed, switching per question. Web browsing, file uploads, connected drives, and custom assistants with attached reference documents are all retrieval features; a plain answer from memory is not. Does ChatGPT use RAG? goes through how to tell which mode you’re in.
Customer support bots
The commercial heartland of RAG. A company indexes its help center, product documentation, and past ticket resolutions, then puts a chatbot in front of it. Ask about a billing error and it retrieves the relevant help articles and answers from them.
You can usually spot these by the linked articles under the answer — and by the way they gracefully hand off to a human when nothing relevant is found, which is the retrieval step reporting an empty shortlist.
Internal company assistants
Same machinery, private library. The documents are the HR handbook, IT runbooks, sales playbooks, engineering wikis, meeting notes. Employees ask in plain language instead of hunting through a shared drive.
This is where most enterprise RAG spending goes, largely because the alternative — teaching a model your policies by retraining it — is impractical for information that changes monthly. How do company chatbots know company policies? covers this in detail.
Coding assistants
An AI that suggests code fitting your project has to know your project, and no general model was trained on your private repository. So these tools index your codebase and retrieve relevant files, function definitions, and related snippets before generating a suggestion.
They also often retrieve from library documentation, which is why a good one can use a package released after the model’s training cutoff. It’s RAG over code rather than prose, and the retrieval step matters at least as much as the model.
Document and note apps
“Chat with your PDF” tools, research assistants over paper libraries, and note apps with an ask-your-notes button all follow the identical recipe: index the user’s material, retrieve on question, answer from the retrieved passages.
Worth knowing a subtlety here. For a short document, some of these skip retrieval entirely and just feed the whole thing to the model, since it fits. Purists don’t call that RAG, because nothing was selected. It only becomes retrieval when the collection outgrows what the model can hold at once.
Places you’d never think to look
Site search on documentation-heavy websites. Many “ask a question” boxes on developer docs, government sites, and knowledge bases are now RAG under the hood.
Email and meeting assistants. Summaries that reference “the thread from Tuesday” are retrieving from your history.
Shopping and travel assistants that answer from current catalogues, inventory, or listings rather than the model’s memory of prices.
Healthcare and legal research tools where citing the source document is not optional — a category that essentially requires retrieval to be usable at all.
Voice assistants doing the boring thing they’ve always done, now with a language model writing the reply: look up the answer, then say it.
What is not RAG
The contrast sharpens the pattern:
- Image generators. Nothing is retrieved; the output isn’t an answer to a factual question.
- Translation, rewriting, summarizing text you pasted. You supplied the material — no search step.
- Autocomplete and grammar suggestions. Pattern completion, no library.
- Brainstorming and creative writing. Retrieval would only get in the way.
- Chatbots answering purely from training knowledge, however impressive.
Why the label is often invisible
You rarely see “powered by RAG” on a product page, for the sensible reason that it describes an implementation rather than a benefit. Products advertise the outcome — “answers from your documents,” “always up to date,” “with sources” — and leave the acronym to engineering blog posts.
Which is a useful thing to know when evaluating tools. Every vendor claim of “trained on your data” deserves a follow-up question, because most such products don’t train on anything: they retrieve. That distinction matters for how your data is handled, how updates propagate, and what happens when you delete a document. What is the difference between an LLM and RAG? unpacks it.
The takeaway
RAG lives inside AI search engines, support bots, internal company assistants, coding tools, document chat apps, and a long tail of “ask a question” boxes across the web. The acronym almost never appears in the interface; the tells do — citations, a visible search step, knowledge of recent events, or knowledge of your private files. Spot those and you’ve spotted retrieval.