Unit 01.01: Where the answer has to come from

"Where did that come from?" is the question a reviewer, an auditor or an annoyed customer will eventually ask. A system that cannot answer it has a problem no amount of accuracy will fix.

Three sources, three risk profiles

Any answer a system produces comes from one of three places: the model's training data, documents retrieved at query time, or the user's own message. They look identical in the output and behave completely differently under scrutiny. Training data cannot be inspected, dated or corrected. Retrieved documents can be. The user's message is trustworthy about what the user wants and unreliable about anything factual.

The example below makes the second source concrete. A three-line corpus, a crude retriever, and three questions - two that the corpus covers and one it does not.

CORPUS
  policy-1   Refunds are allowed within 7 days of purchase.
  policy-2   Enterprise contracts are negotiated per account.
  policy-3   Support responds within one business day.

QUESTION                        WHAT RETRIEVAL FINDS
When does support reply?        policy-3
Can I get a refund?             policy-1
What is the office address?     NOTHING

The third question finds nothing, and that is the correct result -- there is
no address in the corpus. An assistant that answers it anyway is answering
from somewhere other than the evidence.

The third question finds nothing, and that is the correct result: there is no address in the corpus. The value of retrieval is not that it always finds something. It is that when it finds nothing, you know, and you can act on knowing.

Compare that with an answer sourced from training data. There is no empty result to observe, no id to log, and no document to correct when the answer turns out to be wrong. The retrieved source is the only one of the three you can put under version control.

The mistake this prevents

The mistake is assuming a system "has" RAG and therefore every answer is grounded. Retrieval running does not mean retrieval was used. A model handed three irrelevant chunks will happily answer from its weights and cite nothing, and the output looks like every other output. You need the citation check from Module 5 to tell the two apart.

Takeaway

Every answer has a source, whether or not the system tells you which. Designing a knowledge assistant means making that source explicit for every claim, so a wrong answer can be traced to a fixable document rather than to a model you cannot edit.

Mark complete