What local AI processing does and does not mean
Local inference keeps inputs out of a processing API. The website still downloads code and public model files.
Distinguish model tools from helpers
Cards identify local neural models, trainable classifiers, statistical methods and fixed rules. A title template is not generative writing. Extractive summaries return original sentences, and contact masking covers selected patterns rather than every kind of personal information.
Understand the download
MiniLM needs about 47 MB of model/runtime files on first use. OCR typically needs 5–15 MB depending on the selected language and runtime. Files are served by this site and may remain in the normal HTTP cache. API fees are absent, but internet and hosting bandwidth still exist.
Understand what the host can see
Text, training examples and OCR files are not uploaded for processing. The host receives ordinary requests for pages, code and weights and may record IP addresses and request paths. This is different from anonymous browsing or a site that makes no network requests.
Clear and verify your results
Use Clear inputs or leave the page to clear input and result state. Files you download stay on your device. Review OCR errors, approximate similarities and pattern-based masking before publishing or sharing a result. No generative chatbot, model translation or fact verification is included.