Yelp tackled a common user frustration: buried answers on business pages. Their solution, Yelp Assistant, uses a Retrieval-Augmented Generation (RAG) system to provide direct, cited answers. The architecture separates retrieval from generation, sourcing evidence from reviews, photos, and structured attributes to answer precise questions like "Is the patio heated?" without hallucination.

Building this at scale required moving beyond a simple prototype. Yelp implemented a four-part data strategy. They established hybrid ingestion for real-time freshness on reviews but weekly batches for static menus. Critically, they split data stores: a search index for noisy reviews and photos, and a Cassandra database for canonical facts like hours. This prevents anecdotes from overriding verified information.

A key innovation was hybrid photo retrieval, combining caption text matches with image embedding similarity. This handles cases where a concept like "heated" appears only visually in a photo. Unified serving logic then queries both stores efficiently. The system's design prioritizes answer accuracy and source citation over a monolithic index, a necessary evolution for production scale.

Quick Fact: