Why grounding beats knowing

P1 was about asking well. This module is about changing what the model is answering from.

Everything in Explorer assumed the model knew things. It does — a great many — but what it knows is a compressed impression of text it saw during training, with no index, no dates and no way to tell you which part it is drawing on. E7·L1 named the consequence: fluency and accuracy are produced by different machinery, and they come apart without a visible signal.

Grounding is the structural fix. Instead of asking a question and hoping, you put the relevant passage in the request and ask the model to answer from that.

The shift, in one comparison

“What is our refund window?”

The model answers from an impression of every refund policy it has ever read. The answer will be plausible. It may even be right, by coincidence. You cannot tell which.

“Here is section 4 of our returns policy. What is our refund window? Quote the sentence you used.”

Now the model is reading. The task has changed from recall to comprehension, and comprehension is the thing these systems are genuinely good at.

What this actually buys you

Currency. A model’s training has a cutoff. Your policy changed last Tuesday. No amount of prompting fixes that; supplying the current text does.

Specificity. “Our” refund window is not a fact about the world. It is a fact about one document, and only that document contains it.

Checkability — the one that matters most. A grounded answer comes with a place to look. E7·L3’s two-minute method becomes possible: open the source, read the sentence in place. An ungrounded answer offers nothing to check against except your own memory.

What it does not buy you

Grounding is not a truth serum, and the failure it does not fix is worth naming early because it is the one people meet in production.

A passage that mentions the topic is not a passage that answers the question. Retrieval finds text that looks relevant. Whether that text actually contains the answer is a separate question, and nothing about supplying it forces the model to notice the difference.

This is not hypothetical. Lrnon’s own assistant retrieves from these lessons, and asked “what is the capital of France?” it legitimately surfaces E1·L6 — because that lesson contains the phrase “the capital of France is ___” as a next-word-prediction example. The retrieval is correct. The passage does not answer the question. Only an explicit rule to refuse when the passages fall short, plus a check that discards an answer citing nothing, closes that gap.

You will build both in L7 and L8.

The shape of the thing you are building

Four steps, and the rest of this module is each one in turn:

Split your documents into passages small enough to be useful (L2)

Find the few that bear on the question (L3, L4)

Supply them, clearly fenced, with the question (L5)

Refuse when they do not answer it, and prove what was used (L7, L8)

Step four is the one that gets skipped, and it is the one that decides whether the thing you built is trustworthy or merely impressive.

Try it now (5 minutes)

Take a question you already ask an assistant, where the answer depends on a document you control — a policy, a spec, a set of notes.

Ask it cold. Then paste the relevant page in and ask again, adding “quote the sentence you used.” Compare the two answers, and notice specifically whether the cold answer was wrong or merely unverifiable. They are different problems, and the second is the more common one.

Check your understanding

1. The main advantage of putting a source passage in the request is that: It makes the model respond faster The answer becomes checkable against something you control It reduces the number of tokens the request uses

2. A retrieved passage mentions your question's topic but does not answer it. What should a well-built system do? Answer from the passage anyway, since it is on topic Say the passage does not contain the answer, and stop Fall back on the model's general knowledge for that one

3. Grounding an answer in a supplied document means the answer is: Guaranteed correct, because it came from your own material Narrower in what it can be wrong about, and easier to check No more reliable than an ungrounded one

Recap

Grounding replaces “what does the model know” with “what does this passage say”. It fixes currency and specificity, and — most importantly — it makes an answer checkable. It does not make the model honest about passages that fall short of answering, which is the failure the rest of this module is built around.