Zarya
How we handle documents in the Laboratory
Edition of 7 October 2026. Below is what happens to your documents step by step, what is sent out and what this processing does not do.
In short, step by step
- You upload PDFs. The files are encrypted and stay on our server in Russia.
- Our platform extracts the text. Pages without a text layer, that is scans and photos, are recognised by Yandex Vision OCR. The platform then hides passport details, bank accounts, cards and personal contacts in the text.
- Only the prepared fragment text and your question are sent out, to Moonshot AI, the Kimi model.
- The draft memo comes back to your cabinet, and you check every statement yourself.
What happens after upload
Uploaded PDFs are encrypted on arrival and stored on our side. Text is extracted on our own server, located in Russia, in a separate process. If a page has no text layer, its image is sent to Yandex Vision OCR for recognition. Platform code then hides values found by formal patterns: passport details, bank account and card numbers, personal contacts. Each hidden value is replaced by a label, and the table linking labels to the original values stays on our side and is never sent out. Values are hidden by our own code, with no third-party services involved.
What is sent out and to whom
Only the prepared fragment text, your question, the side you stated, the decision description and the review date are sent out.
There are two external recipients. The first is Yandex Vision OCR, Yandex.Cloud LLC. It receives only images of pages without a text layer and returns the recognised text. The image is sent whole, before values are hidden, because values cannot be hidden on a page that has not been recognised yet. The service runs in Russia, request logging on the Yandex side is switched off, and images are not passed on beyond Yandex.
The second recipient is Moonshot AI, the Kimi model (api.moonshot.ai). It prepares the draft analytical memo for an executive from the prepared text. The service runs outside Russia, so the prepared text is processed abroad.
Under that service’s policy the text sent to it may be used to improve its models, and the platform offers no separate opt-out. That is exactly why we hide personal data before sending anything, and never send the original documents at all.
There are no other external recipients in the route. Yandex translation and text-generation services are not called: values are hidden by our own code on our own server, so there is no need to send documents anywhere else for that.
What stays with us
Not sent out: the original PDFs, the filenames, the table linking hidden labels to original values, your credentials and your company record in our system. We record the exact payload and its checksum, so the content of the transmission can be verified.
What hiding values does not guarantee
Hiding values reduces the amount of sensitive information but is not anonymisation and does not guarantee that a fragment cannot identify the parties. It works on formal patterns, so an unusually formatted identifier may remain in the text. Public company details — name, corporate tax ID, registration number — are not hidden: the analysis would be meaningless without them. Send personal data only to the extent your question requires.
Materials that are not accepted
Do not upload materials covered by attorney-client privilege, state secrecy, a formally established trade-secret regime, or materials that may not be processed outside your own organisation. If such a category is declared, the analysis stops on our server before anything is sent to the external service. Such materials require separately agreed arrangements.
Human review is mandatory
The memo is prepared automatically and may contain errors. Every statement carries the document, page and source fragment, and you confirm it yourself. The final conclusion is confirmed separately, and editing the text resets earlier confirmations. Until the review is complete the memo cannot be exported. No internal machine check counts as proof of correctness.
Retention and deletion
Documents, intermediate materials and memos are stored encrypted. With pilot access they are kept for 30 days from the upload of a document or the creation of a memo, and a notice is sent seven days before deletion. On paid plans materials are kept for 12 months, and taking out a plan extends the storage of materials already uploaded. When the period ends, materials are deleted automatically. Your workspace holds 200 MB. You can delete your own documents in the cabinet at any time. Revoking access closes the workspace entirely. Content is visible only to members of your workspace. Administrators see the access terms, status, our own costs and quality signals, not documents or memo text. There is one exception: an analysis you have chosen to share with a Zarya specialist.
Quality review
After each analysis we keep quality signals: the codes of the automatic checks of the memo, the number of documents and statements, your decisions on the statements and the reasons you marked for rejecting them. They contain no document text, memo text or your comments. If a memo got something wrong, you can share that analysis with a Zarya specialist using the “Share with a specialist” button. Sharing lasts 30 days, you can revoke it at any time, and every opening of the materials is recorded. The specialist works out the error and changes the checking rules. Information from your documents is not carried over into training materials.
What we accept for analysis
We accept PDFs in Russian, including scans and photos of pages: up to 20 pages per set and one main question per set. The number of analyses depends on your access terms, and pilot access includes three. Scan recognition is included in every access. Quotations from recognised pages are marked in the memo: handwriting and poor photos may be read with errors. Re-generating a memo is not available to the user, and a technical failure does not consume an analysis.