Solution 01
Score source authority and provenance before generation
01 — Root cause addressed
This addresses the retrieval layer surfacing low-authority joke or satirical material because it matched a data-void query topically.
02 — Implementation
Add authority and content-type signals before a passage reaches the generator. Candidate sources should carry a domain-authority assessment and classifications for satire, sarcasm, forum content, and other provenance cues. Down-weight or exclude low-scoring passages when the output would present a confident factual claim, and preserve the source context needed to distinguish a joke from genuine advice.
03 — Prevention
Evaluate retrieval sets, not only final prose, with cases where highly relevant passages are sarcastic, satirical, or user-generated and authoritative coverage is sparse.
04 — Trade-offs
Reliability and satire classification are imperfect and require labelled examples. Strong filtering can exclude genuine community expertise, while extra classification adds latency and compute at search scale. This is the wrong choice when a narrowly controlled corpus already contains only reviewed authoritative sources.