SilverTorch: Index as Model — A New Retrieval Paradigm for Recommendation Systems

SilverTorch is a unified neural network-based recommendation retrieval system that replaces traditional microservices architecture with a single "Index as Model" paradigm.

• •Achieves 23.7x higher throughput and 20.9x cost efficiency improvements compared to traditional approaches while improving recommendation accuracy
• •Replaces microservices mesh with a single PyTorch model where all retrieval components (ANN search, eligibility filtering, neural reranking, scoring) become tensors and nn.Module operations
• •Eliminates latency from inter-service communication, version inconsistencies between components, and siloed development between ML and infrastructure engineers
• •Redesigns GPU-native retrieval primitives like Bloom index filters and fused Int8 ANN search to optimize for GPU memory behavior and tensor execution