No model reads your archive

The worry, spelled out

Years in reading rooms produce a body of sources that few other people have read. For a researcher that is the advantage. It is also fragile. If those pages are uploaded to a service that keeps copies or learns from what it is given, the advantage can leak into a model anyone can query. Nobody has to steal anything for that to happen. The archive just becomes part of the background. It is a new kind of risk. The tools that create it are the same ones that promise to make scans searchable.

So ask two questions of any tool: does it keep a copy of your document and does anything learn from it?

OCR is not a language model

Two quite different technologies get called AI in the same breath.

Text recognition (OCR) looks at the image of a page and decides which letters are printed where. Current engines, including Tesseract, the open-source engine Searchive uses, do this with a small neural network trained once on typefaces. It turns pixels into characters. It has no idea what a sentence means and can't answer a question about the page. It is a fixed file: Tesseract's models for German, French, English, Dutch and Italian are between one and six megabytes each and ship inside the Searchive installer.

A language model is what sits behind chat assistants. It is trained on enormous amounts of text and can reproduce, paraphrase and reason about what it has read. When people worry about an AI reading their journals they mean this. The worry is justified because text given to a service that trains on its inputs can come back out, in some form, to anyone.

So "Searchive uses no AI" would be the wrong way to put it. Searchive uses a fixed recognition model that runs on your machine, never changes because of your documents and is never connected to a language model or a server.

Where documents go with the usual tools

• Online OCR converters upload your PDF, process it on their server and send a result back. What happens to the copy is a paragraph in their terms. Some services delete it after a while and some don't say.
• Cloud drives with OCR, such as Google Drive, keep the document on their servers by design. Whether it is used to improve their products depends on the account type and on a policy that changes.
• Pasting a scan into a chat assistant is the most direct route from your archive into a language model. Whether the input is used for training differs by product, plan and setting. The defaults aren't always the safe ones. Read the current terms before doing this with anything unpublished.
• Transkribus and similar research platforms upload to a European cooperative under GDPR. It is a much better position, but still a copy on a server you don't control. Our comparison goes into it.
• Desktop OCR (Acrobat, FineReader, OCRmyPDF on the command line) recognizes text on your own machine, which is the right category. A firewall shows what else the application talks to.

Where documents go with Searchive: nowhere

Searchive is a desktop application for macOS and Windows. A document you import goes through four steps.

1. The file is copied unchanged into a library folder you chose and its checksum is recorded.
2. If it needs recognition, Tesseract runs on your machine with the models shipped in the app and writes a second copy with a searchable text layer next to the original.
3. The recognized text is stored in a SQLite database file in the same folder and indexed there.
4. Searching reads that database. Viewing renders the PDF from that folder.

No step involves a network. The app sends no request that carries document content, filenames, page text, search terms or notes, to us or to anyone else. The free tier makes no network request at all. A paid license is confirmed by a periodic check to searchive.io. The app keeps working offline for thirty days between checks. The request carries the license key and, as any web request does, your IP address and the app version in its header. There is no telemetry, no crash reporting that leaves the machine and no analytics inside the application.

Because nothing leaves, nothing can be learned from it. The recognition models on your disk on the day you install are the same files a year later. They don't adapt to your documents. We never receive your documents, so there is nothing for us to train anything on.

Check it yourself

You shouldn't have to take a web page's word for any of this. Two tests will show it, one quick and one thorough.

The unplugged test

Turn off Wi-Fi or pull the cable and start Searchive. Import a folder, choose the language, wait for it to recognize, search, open a page, pin it and export a CSV. Everything works. A tool that needed the network to read your documents would have stopped at the first step.

The firewall test

On macOS an application firewall such as Little Snitch or LuLu shows every connection a program tries to open. On Windows the built-in Defender Firewall can log outbound connections per program. Run Searchive through a full import and search, then look at the list. On the free tier it is empty. With a paid license it shows one destination, searchive.io, contacted occasionally with a small request. You can block it. The app then runs on its thirty-day grace period and tells you so.

Run the same test on any other tool you're considering. A product that processes files locally has nothing to fear from it.

What we won't build

Our product specification rejects cloud OCR services outright and says no embedding model or language model is added unless a user explicitly asks for one. If we ever add something like semantic search it will run entirely on your machine from a model stored in your library folder. It will be off by default. A feature that sends page text to a hosted model is not on the roadmap and won't be. For the people this tool is for, privacy is the reason to use it.

Frequently asked questions