tedk

yt-grounded-qa: answers you can trust, or no answer at all. A user has a problem and asks the AI - but how can they trust the answer? That's where this started. This pipeline structures the question, extracts keywords, searches YouTube, scores subtitle evidence with a transparent rubric (recency, cross-video agreement, numeric sourcing, caption quality; -15 for seller channels), builds a concept-relation ontology where every edge carries video ID plus timestamp, and answers with citations. Below 60 points it withholds the answer instead of guessing, then retries once with relaxed keywords. Verified live: answered ventilation (2 independent sources) and price questions; correctly withheld on floor-noise and used-booth questions. No API key needed (ego-browser + yt-dlp).