yubingz

Chinese-context desensitization for Hermes chats. Replaces PII and business-sensitive entities — person names, company names, addresses, file paths, and magnitude figures — with reversible placeholders before a message reaches the model, then restores them in the reply. Two layers: an optional local/OpenAI-compatible LLM for semantic entity judgment, plus a sub-millisecond regex fallback (phone, ID card, email, IP, path) that keeps the conversation working when the model is unavailable. Disclosure — the semantic layer sends the pre-redaction text to the configured LLM, because recognising a name requires seeing it: with the default provider=ollama that is the local Ollama host and nothing leaves the machine, but provider=openai without an explicit base_url falls back to https://api.siliconflow.cn/v1 (remote); set base_url, or stay on ollama, to keep it local. Also rewrites inbound user messages and outbound assistant replies, so transcript content differs from what you typed; placeholders are code-generated and may over-match (18-digit ID-card-shaped strings, names that collide with the public-entity allowlist); company names are matched greedily and a modifier such as 原北京某某有限公司 is absorbed into the placeholder; magnitude masking drops the numeric value and only fires on the configured trigger words, so a domain term absent from that list is left unmasked without warning; the placeholder mapping is held in memory only and never written to disk.