Pinterest built Manas, an embedding retrieval system that handles a billion embeddings at scale. The core challenge is supporting multiple approximate nearest neighbor search algorithms while keeping query latency low and allowing real-time updates so new content becomes searchable within seconds. This matters because embeddings power recommendation and search features across the platform, and doing this efficiently at Pinterest's scale requires careful engineering choices. The team focused on quantization to shrink the memory footprint without tanking search quality. They applied product quantization (PQ) to embeddings stored on disk while keeping full precision for the centroids — a deliberate deviation from the original SPANN paper. The payoff was significant: using SIMD intrinsics to run 8-bit integer operations in parallel cut computing resources per query by 10–15%. This means faster searches without buying more hardware. The real win here is the specific tweak of mixing precision levels. By quantizing the bulk of the data while preserving precision where it matters most, they got both accuracy and throughput gains. That's not a minor optimization — it's the kind of decision that lets a platform handle billions of queries without becoming prohibitively expensive to run.