为什么值得关注

它使用两阶段本地模型管道（区域检测+相关性过滤）进行带语义替换的往返脱敏，不仅保留回答质量，还显著减少 PII 暴露。

适合谁使用

使用云端 LLM 进行工作或个人查询的注重隐私的专业人士
构建需要匿名处理用户数据后再调用外部 AI API 的开发者
处理敏感数据并希望在不牺牲回答质量的前提下减少 PII 泄露的组织
探索 LLM 使用中的披露控制技术的研究人员或爱好者

典型使用场景

讨论个人搬迁计划（城市、公司）时向云端 AI 咨询，同时不透露真实名称或位置
询问医疗或财务细节时保持特定实体名称私密
使用第三方 LLM 服务进行内部业务数据分析前匿名化数据
在本地开发环境中测试隐私保护的 LLM 工作流

项目优势

逼真的假实体替换比简单打码 [已删除] 更好地保留模型推理能力
相关性模型避免过度删除，只替换非必需的实体以保留回答所需信息
本地 PII 检测离线运行，仅脱敏后的提示离开设备
支持 Anthropic 和 OpenAI，通过可配置 API 密钥设置

使用前须知

并非隐私保证——有足够多查询的恶意对手仍可能重建上下文（披露控制，而非匿名化）
需要安装 Ollama 并下载本地模型（默认量化版本约 3.8 GB），增加部署复杂性
模型工件使用非商业许可（Qwen 研究许可）；商业用途需单独从阿里云获取许可

README 快速开始

Cloak by Praxis

Chat with cloud LLMs while a local model swaps your PII out before the prompt leaves your machine — and rehydrates the answer back to your real entities.

The idea

Every time you ask a cloud AI about something personal — a city you're moving to, a company you work for, a name — you're sending that data to an external server. Cloak intercepts each prompt, runs a local PII detector, replaces sensitive entities with realistic stand-ins (not [REDACTED] — actual fake names and cities), sends the scrubbed prompt to the cloud, then maps the answer back to your real entities before you see it.

The round-trip substitution is the differentiator: the cloud model sees a coherent, answerable prompt; you get a useful, real answer; and the actual PII never left your device.

This is disclosure control, not a privacy guarantee. The pipeline reduces PII exposure significantly — but a determined adversary with enough queries could still reconstruct context. The goal is to make routine LLM use substantially less leaky, without sacrificing answer quality.

Quickstart

Prerequisites: Ollama installed and running.

# 1. Install Ollama (if not already):
#    https://ollama.com/download

# 2. Clone and launch:
git clone https://github.com/praxis-society/praxis-cloak.git
cd praxis-cloak
./scripts/run.sh

run.sh will:

Download praxis/spanfinder-3b and praxis/relevance-3b from HuggingFace and build them into Ollama locally (first run only, ~3.8 GB total — no Ollama account needed)
Create a Python virtual environment and install dependencies
Build the frontend (requires Node; skipped if npm is not found)
Start the local server at http://127.0.0.1:8765 and open it in your browser

Settings: Paste your Anthropic or OpenAI API key in Settings to route scrubbed prompts through your preferred provider.

Models

Model	Weights	Size (q4_k_m)	Role
`praxis/spanfinder-3b`	HuggingFace	~1.8 GB	Detects PII spans in the prompt
`praxis/relevance-3b`	HuggingFace	~1.8 GB	Decides which detected entities to scrub vs. keep (needed to answer the question)

run.sh downloads these GGUFs from HuggingFace and builds them into Ollama for you (the relevanc

项目描述

On-device PII scrub + rehydration for cloud LLMs — chat with Claude/GPT while a local model swaps your PII out before the prompt is sent, then maps the answer back to your real entities. Built with Qwen.

praxis-cloak