Frequently asked questions

It is sent to our server only to run the extraction you asked for. It is held in memory for at most 30 minutes and then discarded. It is never written to disk or saved otherwise. Your browser might remember the key locally so you do not have to retype it the next time you visit Metalens. However, that copy of your key lives in your browser only and not on our server.

The PDF is uploaded to our server, which renders its pages and sends either the page images or the extracted text, together with the prompt, to the model provider you chose. The whole content of the document therefore reaches the provider.

If you provided your own API key, this runs through your own provider account under that provider's data policy. If you use Metalens credits via an account, it is our account at Google, and the Gemini API terms apply. Provider policies differ and change; check the current one before uploading sensitive material. If nothing may leave your machine, run the software locally with a local model (see below).

• The PDF, its page images and the model's raw response, so you can review every value against the highlighted source.
• The extracted records, the verbatim evidence quotes with page numbers, and your corrections and verification history.
• The text layer of the PDF, keyed by the file's hash.

If you are not logged in

Everything you upload is deleted automatically two hours after your last activity. Download your results as JSON or CSV from the review screen before then; no account is needed for that. You can also delete everything at once with the link below.

If you are logged in

Your data stays until you delete a paper or a dataset. Deleting removes the PDF, the page images, the raw model response and the records.

The hosted service runs in the United States (Fly.io, Virginia) with files in Cloudflare R2 object storage. If you are bound by an institutional data agreement that rules this out, you can run a local version of this yourself. You can find the Metalens code at github.com/johanna-einsiedler/metalens and a Python package for running MASEMiner locally at github.com/johanna-einsiedler/maseminer.

Not directly. In the hosted version the PDF is read and the model is called by our server, not by your browser. An address such as http://localhost:11434 therefore points at our server, where no model is running, and the connection fails.

• A model server with a public address, for example a university GPU server or a cloud vLLM instance, works: choose "Local / self-hosted" as provider and enter its URL and model name.
• A model on your laptop behind a tunnel (ngrok, Cloudflare Tunnel) also works, but the content of the PDF still passes through our server, so it brings no privacy benefit.
• To keep everything on your machine, run the software locally (next question). There localhost works and neither the PDF nor the text leaves your computer.

Yes. The software is open source (MIT). The maseminer command starts the same engine locally with an embedded database and no account; PDFs and data stay in a folder on your machine. Point it at a local model server (Ollama, vLLM, LM Studio) and nothing leaves the computer at all. Installation instructions are in the repository's README.

No. Papers, datasets and presets are private to you until you publish a dataset. Publishing shares the extracted records, the evidence quotes, the preset and your verification history, but not the PDFs or page images.

It opens a pull request in the public datasets repository on GitHub; once merged, the dataset is listed in the data catalogue with a citation that points at the GitHub copy. A merged dataset is in git history and in any fork, so it cannot be withdrawn afterwards; corrections are published as a new version.

You are responsible for that, which is why the upload step asks you to confirm it. Two rules of thumb:

• Published articles you have lawful access to can normally be analysed for research under the research and text-and-data-mining exceptions of your jurisdiction, but your publisher's licence terms may restrict automated analysis to the publisher's own tools. Check them for commercial use.
• Manuscripts under peer review, confidential reports and anything you received under a non-disclosure obligation may not be uploaded to a third-party model provider. Reviewer agreements and funder rules forbid it.

If a PDF contains personal data, data-protection law applies to the transfer to the model provider and to our servers.