A deep dive into how Perplexity serves search results at scale: embeddings for ranking, GPU-based model inference, request batching, running inference servers, and handling latency/throughput trade-offs.