What Is Retrieval-Augmented Generation (RAG)?

Retrieval-augmented generation (RAG) is an AI method in which the system first retrieves relevant documents and then writes its answer using them. The term comes from a 2020 research paper by Patrick Lewis and colleagues. For a health practice, RAG is the reason a well-written page can change what an AI tool says about you today, without waiting for the model to be retrained.

How retrieval-augmented generation works

Start with the problem RAG solves. A large language model is the software behind tools like ChatGPT. It learns from a very large body of text up to a certain date, and what it learned is fixed after that. Ask it about a clinic that opened last year and it has little to work with. In the paper, the authors call this built-in knowledge the model's parametric memory, meaning facts stored inside the model itself.

RAG adds a second kind of memory that the model can search. Lewis and colleagues paired a language model with a dense vector index of Wikipedia. A dense vector index is a way of storing text so that passages with similar meaning sit close together, which lets a system find relevant passages even when the exact words differ. The system retrieves the closest passages and the model writes its answer from them.

Today's search tools use the same pattern with the live web. OpenAI describes ChatGPT search in similar terms. It can turn a question into search queries, retrieve results and use them to answer, with links to the sources. Each tool has its own details, and the vendors do not publish their full recipes.

For a practice, the system has two steps, and you can fail at either one.

• Step one is retrieval. The tool has to find your page. That depends on whether the page is public, indexed and a close match for the question.
• Step two is use. The tool has to find a usable answer on the page and decide it is good enough to quote. That depends on how clearly the page is written.

The paper itself reported that RAG produced responses that were more specific, diverse and factual than a model working from its built-in memory alone. That is one reason search-enabled tools lean on it. It is also why a vague page loses. Retrieval works on pieces of text, so it matches a patient's question to a paragraph, not to your whole website.

A simple way to hold this in your head is three layers. Retrieval picks the material. The language model writes the answer. The platform decides what the reader sees, for example whether sources appear as a visible list or sit behind an icon. A practice can do well in the first two and still be hard to notice in the third. Display choices change often and you do not control them, so I would not build a plan around them.

What a model learned in training is a fixed snapshot until a new version ships (see training data). The lever you control day to day is what gets retrieved.

Access still matters. Google Search Central says there are no additional requirements to appear in AI Overviews or AI Mode, and no special optimizations are necessary, but a page has to be indexed and eligible to show with a snippet. For ChatGPT search, OpenAI says OAI-SearchBot is the crawler that surfaces sites, and its settings are independent from GPTBot, which crawls for training. A site can allow one and block the other.

RAG is closely tied to grounding, the goal of tying answers to real sources. Retrieval is the most common route to it.

One more piece sits in front of retrieval. When a tool breaks one question into many searches before it retrieves anything, that is query fan-out. More searches means more chances for a focused page of yours to match one of them.

Example: retrieval-augmented generation at a health practice

Consider a naturopathic clinic in a mid-size city. Its "Thyroid care" page opens with two sentences. The first says the clinic evaluates thyroid symptoms with a full lab panel and a one-hour intake. The second names the naturopathic doctor who leads the visit and the city. Below that, short sections cover what labs are ordered, how often follow-up happens and what the first visit costs.

A patient asks, "What should I expect at a naturopath for thyroid problems?" This example is illustrative, not a transcript, and here is how the two steps could go.

• Retrieval is the first step. The tool searches and finds the clinic's page, along with a few others. The page matches because its headings use the same words patients use.
• Use is the second step. The tool lifts the intake description and the lab list into its answer, then links the page as a source.

Now take a second clinic in the same city. Its thyroid information sits inside a slideshow, and its page title is "Services." The tool either cannot read the text or cannot tell it apart from a dozen other clinics. Retrieval may fail at step one.

Here is the real data that frames this. Lewis and colleagues found that retrieving passages before writing made the model's responses more specific, diverse and factual than memory alone. That finding is from 2020 and from a research setup, so it describes the method, not any single product today. OpenAI's help page describes ChatGPT search using queries, retrieval and links in a similar way. The takeaway is practical. Tools that retrieve are reading the web in the moment, so a page you publish can be used once it is crawled and indexed.

How to check yours in 60 seconds

1. Pick one question a patient would ask that your page answers, such as "What happens at a first functional medicine visit?"
2. Ask it in a search-enabled tool, such as ChatGPT with search, Perplexity or Google AI Mode.
3. Open the cited sources and look for your page, or a page that lists you.
4. If it is missing, ask again using your page's own wording and see whether it appears. If it appears only then, your headings do not match how patients ask. Save a screenshot of the sources either way, so next month you have something to compare against.

What to do about it

1. Give each important patient question one focused page, with a title that states the question or topic.
2. Put the answer in the first two sentences, before any story or background. Leading with the answer makes the section easier to match and quote.
3. Use headings that match how patients phrase questions, and keep each section understandable on its own.
4. Keep the page crawlable, with real text on the page and no login wall. Confirm your robots.txt file does not block OAI-SearchBot.
5. Add the plain facts that make a page easy to match, such as provider names, credentials, city and what a visit includes. Ask your web person to confirm the page is indexed.
6. Retest every month with the same questions and record whether your page was retrieved and whether it was cited.
7. Refresh the page when anything changes, including providers, prices, labs and hours, since a tool reading it today will repeat what it finds.