Arcjet extended its sensitive information detection SDK with an optional local NER model backend using the open source Rampart model (14MB ONNX). The post details the engineering decisions: a pluggable SensitiveInfoBackend interface to keep the rule stable across model changes, offset reconstruction from BERT-style subword tokens back to original character positions, hybrid detection combining deterministic regex/Luhn recognizers for structured data with the NER model for names and addresses, a span merge algorithm that resolves overlaps by length and source priority, lazy model loading with failure-safe caching, overlapping 480-character chunking to handle inputs beyond the 512-token model window, and Next.js bundling guidance to avoid packaging native ONNX binaries. The feature ships in server runtimes (Node.js, Bun, Deno) with p50 latency around 6.5ms.
