A parks chatbot's hybrid BM25-plus-vector ranker is skewed by keyword-stuffed volunteer wiki pages. What RAG abuse pattern should security recognize?

Ranking manipulation against the civic knowledge base through keyword stuffing is the RAG abuse pattern to recognize. Hybrid BM25-plus-vector rankers still reward term frequency and related signals. Volunteer wiki pages stuffed with magic keywords can outrank quieter official parks pages, steering the chatbot toward low-trust content. Security and knowledge owners should detect stuffing, down-rank or exclude manipulable wikis, and monitor hybrid score distributions for gaming.

Inevitable GPU thermal throttling during hybrid search fails by concept: the stem describes ranking skew from stuffed pages, not hardware thermals. Proof that BM25 cannot be influenced by document text fails because BM25 is explicitly text-term based and highly influenceable by repetition. Calling the skew a benign side effect of HTTPS on the wiki origin fails; transport security does not validate content trustworthiness or ranking integrity.

Exam caveat: Hybrid retrieval inherits classic search-engine gaming. Municipal open contribution surfaces need anti-stuffing and trust tiers, not blind hybrid ranking.

Operational check: Score official vs wiki pages for stuffed query terms, add quality or trust priors that favor parks.gov sources, quarantine pages that spike keyword density, and re-test chatbot rankings after mitigations.