Unit 01.02: Retrieval does not make a model honest

The most common misreading of RAG is that adding retrieval makes a model truthful. It does not. It changes what the model is likely to say, which is a much weaker guarantee than it sounds.

Two responses, one citation each

Retrieval supplies context. The model still decides what to do with it - and "ignore it and answer anyway" is one of the options. So is "summarise it slightly wrong." So is "combine two chunks into a claim neither one makes."

The example below below hands the same context to two functions. One stays inside it; the other does not. Both attach the same source id.

CONTEXT SUPPLIED TO THE MODEL
  [policy-1] Refunds are allowed within 7 days of purchase.

ANSWER A   "Refunds are allowed within 7 days."           cites policy-1
ANSWER B   "Refunds are allowed within 30 days."          cites policy-1

Both cite policy-1. Only one is supported by it.

A citation is a claim about where text came from. Nothing checks that claim
unless you build the check.

Both outputs cite policy-1. Only one is supported by it. The citation is a claim about where the text came from, and nothing in the pipeline verifies that claim unless you write the verification.

That is the asymmetry worth carrying into the rest of the course: adding a source id is a formatting change, while checking that the cited chunk contains the claim is a real check. Module 5 builds the check, and Module 8 measures how often it fires.

The mistake this prevents

The mistake is treating the architecture as the guarantee. A diagram with a vector store in it says nothing about whether answers are grounded. What says something is an eval set with cases that should refuse, run after every change, with retrieval scored separately from the answer.

Takeaway

Retrieval improves the odds. It does not enforce anything. The enforcement comes from the refusal path, the citation format, and the checks you run against them.

Mark complete