← Safety

核验标记表示资料完整度，不是产品评分。下列未确认项可能影响选型；请按对应来源逐项确认。The confidence mark describes evidence completeness, not a product score. Unconfirmed items below may affect your choice; check them against the relevant sources.

English summaries and interface labels are available. Detailed source notes below are currently in Chinese.

模型与开放条件Model & openness

它本身不带生成模型：analyzer 页写明「Predefined recognizers leverage regex, Named Entity Recognition and other types of logic to detect PII in unstructured text.」，识别面由正则、NLP（spaCy / Transformers 类）、规则与校验和组合， 上下文词用来加权 score，自定义 recognizer 有官方 tutorial。可选 NLP 引擎与模型档位的完整清单：本批次未逐条清点。

It carries no generative model of its own: the Analyzer page states "Predefined recognizers leverage regex, Named Entity Recognition and other types of logic to detect PII in unstructured text." — the recognition surface combines regex, NLP (spaCy / Transformers class), rules and checksums, with context words weighting the score, and custom recognizers covered by an official tutorial. The full list of selectable NLP engines and model tiers: not itemised in this pass.

运行位置Where it runs

全部跑在你自己的边界内：文档首页的安装口径是 pip 包或 Docker REST API，并写明支持离线使用、数据留在你自己的部署里。 官方未提供托管服务面（未取到该口径）。镜像标签与版本对应关系本批次未逐字核验。

Everything runs inside your own boundary: the docs home page gives the install surface as pip packages or a Docker REST API, and states that offline usage is supported with data staying inside your own deployment. No hosted service surface is offered officially (no such statement captured). Image tags and their version mapping: not verified word for word in this pass.

本地文件Local files

它不读你的工作区，只处理你显式交给它的文本与图像，结果在你的进程或你的服务里返回。recognizer 依赖的语言模型文件的 下载位置、缓存路径与离线打包方式：本批次未取到逐字官方说明，记未知。

It does not read your workspace; it processes only the text and images you explicitly hand it, and returns results inside your own process or service. Where the recognizers' language-model files are downloaded and cached, and how they are packaged offline: no word-for-word official statement captured, recorded as unknown.

关机后的任务Tasks after shutdown

没有托管后台任务：一次 analyze / anonymize 是一次调用或你自己服务里的一次处理，批量与重试由你的编排层负责。 官方文档目录（Getting Started / Analyzer / Anonymizer / Image Redactor / Samples / API Docs）未出现队列或续跑概念。

There is no hosted background task: one analyze / anonymize call is one invocation, or one pass inside your own service, with batching and retries owned by your orchestration layer. The docs table of contents (Getting Started / Analyzer / Anonymizer / Image Redactor / Samples / API Docs) contains no queue or resume concept.

工具与扩展Tools & extensions

三面 API：AnalyzerEngine（识别）、AnonymizerEngine（变换）、Image Redactor（图像遮挡）。变换算子按文档 anonymizer 页 在册：replace、redact、hash、mask、encrypt、keep、custom，经 OperatorConfig 的 type 与 params 配置，多个算子可组合。 MCP 接入方式：官方文档未说明，记未知。

Three API faces: AnalyzerEngine (recognition), AnonymizerEngine (transformation), Image Redactor (image masking). The operators on record from the Anonymizer page are replace, redact, hash, mask, encrypt, keep and custom, configured through OperatorConfig's type and params and combinable. MCP transport support: not described in the official docs, recorded as unknown.

上下文与记忆Context & memory

这里的「上下文」是字面机制：recognizer 用上下文词对候选实体加权，因此同一段文本里邻近词会改变结果；文档首页把项目 表述为面向 text and images 的「fast identification and anonymization modules」。它是无状态的一次处理，不存在跨会话记忆， 批量场景的上下文窗口管理属于你自己的调用方。

Here "context" is a literal mechanism: recognizers weight candidate entities with context words, so neighbouring words in the same text change the result; the docs home page describes the project as "fast identification and anonymization modules" for text and images. It is a stateless one-pass handler with no cross-session memory — batch context-window management belongs to your calling code.

权限与限制Permissions & limits

许可见 license 字段（MIT，依 PyPI 元数据；LICENSE 正文未直读到）。归属要记两条实测事实：旧址 microsoft.github.io/presidio 当日显示「This page has moved」，中间站 data-privacy-stack.github.io 再 301 到 presidio.dataprivacystack.org，文档首页指认项目由 Data Privacy Stack 治理。数据流向：识别与脱敏在你自己的部署内完成、 支持离线，即文本不必发给第三方。红线口径：analyzer 页原文在册「Scores are not probabilities」，所以 score 阈值 不是检出概率，本站不写检出率、漏检率，也不做任何产品排名；它是否满足某法域的匿名化义务，官方不作法律结论。

See the licence field (MIT, taken from the PyPI licence metadata; the LICENSE text was not read directly). Two measured facts about ownership: the old address microsoft.github.io/presidio displayed "This page has moved" on the verification date, and the intermediate host data-privacy-stack.github.io 301s to presidio.dataprivacystack.org, whose home page states the project is governed by Data Privacy Stack. Data flow: recognition and anonymisation complete inside your own deployment and offline use is supported, meaning text need not be sent to any third party. Red line: the Analyzer page states on record "Scores are not probabilities", so a score threshold is not a detection probability; this site publishes no detection rate, no miss rate and no product rankings. Whether it satisfies an anonymisation obligation under any given law is not a legal conclusion the project makes.

适合什么任务Best for

适合：把个人数据识别与变换插进自己的管线（RAG 入库前、日志落盘前、发给云端模型前）、需要算子级定制与自定义 recognizer、要求文本不出自己的网络边界、要把整套东西离线跑。不适合：要托管控制台与合规交付物；要文本与图像之外的 模态（官方在册面止于这两类）；把 score 阈值当合规证明的团队。

Fits: inserting personal-data recognition and transformation into your own pipeline (before RAG ingestion, before logs are written, before a call to a hosted model); needing operator-level composition and custom recognizers; requiring text not to leave your network boundary; needing the whole chain to run offline. Not a fit: teams wanting a hosted console and compliance deliverables; modalities beyond text and images (the surface on record stops there); anyone treating a score threshold as proof of compliance.

本页由This page is generated from the ai-handbook 数据层生成（CC BY 4.0）。data layer (CC BY 4.0). 方法论与坐标系定义见仓库内For the methodology and coordinate definitions, see METHODOLOGY.md.