You will be able to explain retrieval and web search as ways of putting fresh information into the context window.
Ask a search-enabled assistant whether part-time workers in Singapore get annual leave, and you will usually get a tidy answer with small numbered markers after some sentences. Click one and a Ministry of Manpower page opens, or a law firm's blog, or sometimes a forum thread from years ago. The answer looks researched, but the model behind it was trained long before some of those pages were written.
So where did the pages come from? Lesson 5.1 gave you half the answer: the model can only use what is in its context window. This lesson covers the other half, which is how fresh text gets into that window in the first place.
The method is called retrieval. Before the model writes anything, a separate search step runs. It takes your question, sometimes rewritten into better search terms, and looks for relevant material. That might be a web search, or a search through a set of documents such as a company's HR policies or the files you uploaded. The search step picks out the passages that seem most relevant and pastes them into the context window, along with an instruction along the lines of "answer the question using these sources and say which one each point came from".
Only then does the model write. It is still doing what lesson 3.2 described, predicting one token after another. The difference is that the text in front of it now contains passages that actually discuss your question, so the likeliest next tokens tend to repeat what those passages say. The numbered markers are the model's way of pointing back to the passage it drew on.
This pattern has a technical name you will see in product pages and job ads: retrieval-augmented generation, usually shortened to RAG. Retrieval is the search step, augmented means the results are added to the context, and generation is the model writing its reply.
Once you know the pattern, you will spot it in many places. Assistants with a web search option use it every time they search. Tools that let you chat with a PDF, a folder of contracts or a company knowledge base use it too, except that they search your documents instead of the web. Many customer service chatbots on bank, telco and insurer websites work the same way, searching the company's help articles before they reply.
With documents, the tool usually cuts each file into chunks of a few paragraphs ahead of time. When you ask a question, it finds the chunks that match it most closely in meaning, not only those that share your exact words, and passes those to the model. It does not hand over the whole file. That is worth remembering when a chat-with-your-documents tool misses something: the relevant chunk may never have reached the model.
Retrieval solves two real problems. The model can answer about things that happened after its training cutoff, which lesson 4.4 showed it otherwise gets wrong. And it can answer from private material it was never trained on, such as your company's leave policy. Because the answer is grounded in text it can see, it invents facts less often than it does from memory alone.
Less often is not never. The model can still misread a passage, merge details from two sources into one claim that neither makes, or drop a condition that changes the meaning. It can put a citation marker next to a sentence that the cited page does not support. The marker shows which passage was nearby when the sentence was written. That is not the same as proof that the passage says it.
The answer can only be as good as what the search step found, and the model has no way of knowing what a better search would have turned up. If the search pulls up an outdated guide, a page about the rules in Malaysia, or a forum post from someone guessing, the model writes from that material in the same assured tone it would use for the official source.
Take an example. Mei Ling, an admin manager at a small design firm, asks about the notice period for ending a staff member's contract. The search finds a law firm article written before a later change and a forum thread. The reply cites both and reads well. Neither source is the current official guidance, and nothing in the wording warns her of that.
So the citations are your way in, and you have to open them. Check that each source is official or at least reputable, that it is current, that it is about Singapore when your question is, and that the cited passage really says what the answer claims. Lesson 6.3, A five-minute routine for checking any answer, turns this into a habit. The quickest way to see why it matters is to run a search-enabled assistant on a local rule you care about and follow every one of its markers back to the page.
Ask a search-enabled assistant a question about a Singapore rule, open every source it cites, and note whether each source actually says what the answer claims.
Junxiong-WFG Organisation is an authorised representative of AIA Financial Advisers Private Limited (Reg. No. 201715016G).