On-Device Commercial Intent Retrieval Under Size, Latency, and Privacy Constraints: A 3 MiB Retrieval System with Typed Egress Boundaries
The research developed a system for retrieving commercial intent directly on a device under strict limits of size (under 3 MiB), latency (under 20ms p95), and privacy. The system uses a small, quantized model and a custom retrieval path over Korean commerce terms. It proves that achieving functional intent retrieval without sending any raw data externally is possible, though performance varies based on the type of commercial information requested.