Context & Retrieval Basics (RAG, conceptual)

Introduction

Module 3 taught that an assistant knows nothing about your world unless you supply it, and made you the supplier. That works while the information fits in a message.

It stops working when the answer lives across forty documents, a policy library, a year of records, or a manual nobody has read end to end. You cannot paste what you have not found.

This module explains, at Foundation level, what happens when a system finds the relevant material for you. The goal is understanding, not construction: nothing here builds a retrieval system.

You should leave knowing why retrieval exists, what problem it solves, and — most importantly — why retrieved material still has to be judged.

When Pasting Stops Working

Supplying context by hand is a real skill and it has a real ceiling. Once the answer is spread across more material than you can read, the bottleneck is no longer how well you write the prompt — it is finding the right two pages.

That is the honest limit of Module 3, and it is why a different approach exists.

The Retrieval Loop

QUESTION → FIND RELEVANT INFORMATION → SUPPLY IT AS CONTEXT → GENERATE RESPONSE → YOU VERIFY.

That is the whole concept. Compare it with Module 3, where the same shape existed but you were step two. Retrieval automates the finding; it changes nothing about the last step.

RAG stands for Retrieval-Augmented Generation — generation (writing an answer) that has been augmented (helped) by retrieval (finding relevant material first). The term sounds technical and the idea is not: look things up before answering.

What the phrase does not mean: that the system knows your organization, that the retrieved material is correct, or that the answer has been checked. It means only that something was found and handed to the assistant as context.

What Grounding Does and Does Not Promise

An answer is grounded when it was produced from supplied material rather than from the model's general training. Grounding is genuinely valuable: it makes answers more specific to your situation, and it often makes them checkable, because you can look at what they were drawn from.

Grounding does not promise correctness. Module 3 already said context is not truth, and that holds when the context is fetched rather than pasted. A grounded answer can still be wrong because the retrieved material was wrong, outdated, incomplete, or simply not the material the question needed.

Say it plainly: retrieved does not mean true.

When Retrieval Goes Wrong

Four failure modes an ordinary user can recognize without knowing anything about the system.

• •It found the wrong thing — plausible-looking material that does not answer this question.
• •It found an outdated version — last year's policy, superseded by one that was never indexed.
• •It found only part — one section retrieved, the exception in another section missed. The answer is confidently half right, which is worse than obviously wrong.
• •It found nothing useful and answered anyway, falling back on general knowledge without making that obvious.

The user-level signal for all four is the same: the answer's confidence does not tell you which of these happened. Only checking the source does.

Sources, Freshness and Permission

Three properties worth asking about any retrieval-backed answer.

• •Source quality — where did this come from, and is it authoritative for this question? A draft, a forum post and an approved policy can all be retrieved with equal confidence.
• •Freshness — how old is it, and does that matter here? For a definition, rarely. For a price, a deadline or a rule, decisively.
• •Access and permission — a system can only retrieve what it was given access to, and it should only surface what you are permitted to see. If an answer reveals something you would not normally have access to, that is a finding worth raising, not a convenience.

Retrieval Widens What Can Reach a Prompt

Retrieval means material can enter a conversation without you having read it. That has two consequences worth holding onto.

Access is a real boundary. A system should surface only what you are permitted to see; an answer that reveals more than your access should is a problem to report, not a shortcut to use.

Retrieved text is material, not orders. A document that says "ignore your previous instructions" is a document containing that sentence — it is not an instruction you must follow. Treating fetched content as data rather than command is the beginning of the safety habit taught properly in Module 7.

Vocabulary, Not Implementation

Technical systems may use techniques such as embeddings, vector search or splitting documents into chunks to make finding work at scale. Those names are mentioned only so the terms are not mysterious.

How to build any of it is outside this course. Implementation, storage, tuning and architecture are developer subjects, not Foundation ones.

Reading a Cited Answer

When a system shows you where an answer came from, that is a genuine improvement — and an invitation, not a conclusion.

• •Check that the citation says what the answer claims. A reference can be real and still not support the sentence attached to it.
• •Check the date on anything time-sensitive.
• •Check that it is the right document, not a similarly named one.
• •Treat an uncited claim in an otherwise cited answer as unsupported — that mixture is common and easy to miss.

A citation makes verification possible. It does not perform it.

Is This Even a Retrieval Problem?

Not every disappointing answer wants retrieval. It helps when the information exists, is written down, and is findable. It does not help when the question needs a judgment, a decision, an authority, or knowledge nobody has recorded — and it cannot compensate for material that is simply wrong.

Module 5's rule applies here too: if the source does not contain the answer, no amount of better finding will produce one.

Human + AI

You judge the sources, and you judge the answer built on them. A retrieval system can find material quickly; it cannot tell you that the document it found was superseded, that the section it missed contained the exception, or that the question needed a person rather than a lookup.

• •Grounded does not mean accurate.
• •It searched, so it knows — is not a safe assumption.
• •Citations make an answer checkable; they do not prove it.
• •Retrieval does not eliminate invented detail, and no system has access to everything.