BreakpointDaniel Andrade Groppe explains how Yelp built its Assistant UI using a hybrid native and server-driven architecture. The post covers separating view configuration from conversation data, optimistic message updates, quick replies, client actions, and failure handling across mobile platforms.
Yelp engineering blog
Yelp describes how its Nrtsearch Inference Plugin embeds machine-learning ranking inside the Lucene-based search engine, avoiding network overhead from a standalone scoring service. The post covers feature extraction, model loading, custom scorers, testing, deployment, and Prometheus-based monitoring.
Yelp describes how it built Menu Vision, a cross-platform feature that uses on-device text recognition, computer vision, prefetching, and distributed dish data to identify menu items in real time. The post details its data pipeline, discovery architecture, matching improvements with substring and Jaro-Winkler similarity, and evolution from a hackathon prototype to a production visual experience.
Yelp details its three-year migration of a 1.4-million-line JavaScript monorepo from Flow to TypeScript. The post explains the incremental Flow/TypeScript interoperability toolchain, dependency-graph strategy, automation, rollout, and outcomes, including improved type coverage and developer productivity.
Yelp details its transparent migration from deprecated Apollo Tooling to GraphQL Codegen across a 500-plus-package React monorepo. The post explains compatibility-focused plugin changes, concurrency fixes, upstream collaboration, and incremental rollout practices that preserved existing imports and generated types.
Yelp describes Training Orchestrator, a configuration-driven DAG framework that standardizes Spark-based model training. The system decouples training logic from execution, enabling local testing, type-safe Pydantic validation, MLflow tracking, reproducible runs, and unified monitoring across teams.
Yelp details how it replaced a rigid support chatbot with an LLM-assisted system using workflow routing and a RAG pipeline. The post explains metadata-based embeddings, FAISS retrieval, validation against hallucinated links, daily knowledge-base updates, and a 94% recall@5 result that helped double chatbot resolution rates.
Yelp explains how partition-level access visualizations exposed analytics data usage patterns, enabling targeted Apache Iceberg migrations and a 33% reduction in S3 storage costs. The post details storage-class selection, access-based retention, and an AWS server-access-log aggregation architecture.
Yelp explains how its Konbini code-generation system connects the CHAOS server-driven UI framework with the Cookbook design system across Web, iOS, Android, and Python. The post details shared JSON interfaces, serialization, generated client libraries, design tokens, and versioning with migrations for backward compatibility.