Shipped a semantic recommendation engine. No keywords. Content → vector embedding → pgvector. User reads something → interest_vector updates → cosine search drives the next feed. All processing happens in the background. Zero impact on response time.