Episode Details

When your podcast library hits 4,000 episodes, traditional tagging and search break down completely. Tags multiply into duplicates, context windows overflow, and users stop trusting the system. This episode unpacks the two-stage agentic pipeline that solves it: a map step that generates raw tags per episode, then a reduce step using embedding similarity and DBSCAN clustering to normalize everything into a clean, canonical taxonomy. No manual effort, no token limits, no drift.