Page 126 - CyberSecurityEssentialsEnglish

reader does not need to memorize these measures or their formulas, only to distinguish two
situations: the system never found the appropriate evidence, and the system found the
appropriate evidence but used it or phrased the answer wrongly.

In the generation stage, evaluate the correctness of the answer, its coverage of the question,
its willingness to abstain, and citation faithfulness: does the passage actually support the

claim, or is the link correct while the sentence goes beyond what it says? A measure of
citation completeness can be added: does every important claim have support? Then
evaluate the full experience: the time, the cost, the reindexing, and the review of failing

cases.

Failures in retrieval systems are not only linguistic. Documents may be old, duplicated, or
contradictory. OCR may fail on a number, a table may be split away from its heading, or the

search may return a passage from a revoked version. Address this with version and date
management, the removal of unapproved copies, the recording of the source and page, and
a clear policy when documents conflict.

There is also the security of retrieval. A document may contain instructions aimed at the
model, or the knowledge base may be injected with misleading content. Retrieved text must

be treated as untrusted data rather than as commands, access permissions must be applied
before retrieval, the sources of the index must be inspected, and the passages that were
used must be logged. A user should not see a result drawn from a document they have no
right to access merely because the model is able to summarize it.

125