File Search System (ML)

Design a machine-learning powered file search system for Dropbox that lets hundreds of millions of users instantly find the right document, slide-deck, PDF, image, or code file among tens of billions of objects stored in multiple formats. The system must support both keyword queries (“quarterly revenue report”) and semantic intent (“Q3 financial summary”) even when no words overlap, return personalized and access-controlled results in <200 ms at p99, and continuously learn from user interactions to improve ranking. You are expected to detail the end-to-end architecture: ingestion & content-extraction pipeline (PDF, DOCX, PPTX, images, source code), indexing layer that marries inverted indices (BM25) with dense vector embeddings, hybrid retrieval & ranking (ANN + cross-encoder re-ranker), real-time personalization that boosts recently opened or team files, and enforcement of Dropbox-style sharing permissions so users can never see files they aren’t allowed to view. Discuss scale (10 B+ files, 4+ regions, 1 M+ QPS), storage & compute cost trade-offs, incremental indexing on every file edit, and how you would measure and iterate on search quality offline and online.