The Librarian behind the mail slot has read everything — every novel, every textbook, every encyclopaedia entry ever printed. That sounds like a superpower. And mostly it is. But there is a catch that catches almost every first-time user out, and it trips up quite a few developers too.
The Librarian finished reading a while ago. They do not get fresh newspapers slid under their door. And when you ask about something they are hazy on, they do not say "I don't know" — they say whatever sounds most like the right answer. Sometimes that is perfectly fine. Sometimes it is confidently, plausibly wrong.
This article is about the two errands the Runner can run to fix both problems.
The Librarian's two big weaknesses
Before the errands make sense, it helps to see the problem clearly.
Weakness one: the training cut-off. At some point, the Librarian stopped reading new material. Everything since then is a blank. Ask them about a news story from last month, a law that changed last year, or the price of something today, and they are guessing. They will not always tell you they are guessing.
Weakness two: hallucination. This is the word techies use when an AI confidently invents something that is not true. Hallucination happens because the Librarian's job is to write the next most-likely word. If they are not sure, they still write the next most-likely word — they cannot simply stop. The result can read like a real fact, a real quote, a real book title, that does not actually exist. It is not lying. It is pattern-matching without a reality check.
Both weaknesses come from the same root: the Librarian cannot look anything up. They only have what is already inside their head.
Errand one: fetch the right book from YOUR shelf
Here is the scenario. You are a parent reading through your child's school report, and you want to ask the AI a question about it. You slip a note through the slot: "What are the main concerns in this report?" But the Librarian has never seen the report. They will guess, or politely say they cannot help without it. Either way, not useful.
The fix is for the Runner to fetch the document and attach it to the note. Now the note reads: "Here is the report [full text]. What are the main concerns?" The Librarian can answer properly, because the actual information is right there on the note.
That is the simple version. But the real magic is what happens when your shelf has hundreds of documents — a whole folder of school reports, lesson notes, or your company's internal policies. You cannot staple all of them onto every note. The slot is only so wide.
So the Runner has to be clever about which pages to fetch.
Search by meaning, not by spelling
Here is where something genuinely interesting happens.
The obvious approach is to search for exact words. You ask about "a lonely dog" so the Runner hunts for documents containing the phrase "lonely dog." That works — until the relevant passage says "a puppy who misses its owner" instead. Same meaning, different words. A keyword search misses it entirely.
The Runner uses a different method. Before any of this starts, every document on your shelf is read and converted into a kind of map of its meaning — numbers that capture what the text is about, not just which words it uses. Techies call these numbers vectors, and the whole approach is called vector search or semantic search. Think of it as a super-librarian who has grouped every book on the shelf by what it is about — not alphabetically, not by spine colour, but by subject and feeling. "A puppy who misses its owner" ends up sitting right next to "a lonely dog" on that map, because the two passages are about the same thing.
When the Runner gets your question, they convert that into the same kind of map and look for the documents whose maps are closest. Then they fetch those pages and staple them onto your note.
The full name for this whole trick — convert documents into meaning-maps, retrieve the relevant ones, staple them onto the note, let the Librarian answer from real text — is RAG: retrieval-augmented generation. The Librarian's answer is augmented, topped up, by retrieved real material rather than by guessing.
RAG does not make the Librarian all-knowing. It makes the Librarian answer from actual evidence rather than memory. That is a very useful difference.
Why this matters for the "making things up" problem
When the Librarian answers from retrieved text that is sitting right there on the note, there is much less room to invent. The answer is grounded in something real and specific. Hallucination does not disappear entirely — the Librarian can still misread or misquote — but it drops dramatically compared to answering from bare memory about a topic they might only half-know.
This is why a lot of serious AI tools today are built around RAG rather than relying on the model's training alone. Customer service tools that pull from a real database. Study assistants that pull from your actual notes. Document analysis tools that pull from the file you uploaded. They are all different sizes of the same Runner-fetching-the-book errand.
Errand two: go outside to ask around
The second errand handles the cut-off problem: facts that are newer than the Librarian's training, or facts that simply change — today's exchange rate, yesterday's headlines, who currently holds a particular job.
The Runner goes outside. They take your question, run a web search, read the most relevant pages, and bring the key facts back in to staple onto the note. Now the Librarian is answering from fresh sources, not from memory of what the internet looked like a year or two ago.
Web search for AI works in almost exactly the same way as RAG for your own files. The difference is where the Runner goes to fetch: your private shelf, versus the whole public internet. The mechanism — retrieve relevant material, attach it to the note, let the model answer from real text — is the same in both cases.
This is why modern AI assistants can often tell you today's football scores, quote last week's government announcement, or summarise a news story from this morning. The model itself still has its old training. But the Runner fetched something real and put it on the note.
Two errands, one idea
Underneath both errands is a single principle: the Librarian does better work when there is real, specific evidence on the note. Left alone with only their memory, the Librarian guesses. Given the right pages, the Librarian reads and summarises. One of those is much more reliable than the other.
The Runner's job in both cases is to find the right thing quickly — by meaning, not just by spelling — and get it onto the note before it goes through the slot. The Librarian never even needs to know where the pages came from. They just see a longer note and answer from it.
If you have read what an AI agent is, you will recognise this pattern. The Runner handles the mechanical fetch-and-attach work. The Librarian handles the understanding-and-writing work. Neither could do the other's job well. Together, they produce something genuinely useful.
The one-line version
RAG and web search are the same errand in two different directions: the Runner fetches real, current, specific text and staples it onto the note so the Librarian answers from evidence rather than memory. That is why an AI that can search often makes fewer things up than one that cannot.
Next in the series: the Runner also needs a way to request specific actions from the outside world — not just fetch information, but do things. That is what function calling and MCP are about. Read Function calling and MCP explained next.
Back to the AI in Plain English series.
Jason runs aitutors.me — AI tutors for UK secondary students. Updated 5 June 2026.