Yelp’s business pages overflow with reviews, photos, menus, hours and community posts. That wealth of content makes it hard to pull out a single fact—say, “Is the patio heated?”—without wading through dozens of reviews. To fix this, Yelp rolled out an AI assistant on business pages. It parses a user’s question, grabs relevant reviews or structured facts, then uses a large language model to spit back a brief, cited answer—yes or no, a list of five highlights, even a tailored three-course menu suggestion. Under the hood, every question follows the same path. First, the system tags the query for intent and safety, then it picks which data to search—reviews, photos, or business fields. Retrieval hits three near-real-time indices (reviews, photo captions and embeddings, external site content) plus a Cassandra store of structured facts. Next, a prompt combining retrieved snippets goes to the LLM, which generates a concise answer with citations. Finally, the answer and any relevant images stream back to the user interface, with logs and metrics feeding ongoing quality checks. Building this end-to-end flow took nine months. The prototype relied on Redis snapshots of warehouse dumps; the production version uses streaming ingestion for reviews, photos and core business data, keeping those fresh within ten minutes. Websites, menus and community posts still update weekly. A single content-fetch API unifies all sources into “LLM-friendly” blobs. Key takeaways: an EAV schema in Cassandra lets them add fields on the fly, keyword search plus LLM expansion hit quality targets quickly, and batching only the slow-moving data keeps the streaming pipeline more manageable.