What happens to a document you upload

If you are putting a client contract, a patient file, or an unreleased filing into a tool you did not build, the question is not whether the AI is good. It is what happens to the file. This page answers that directly. The binding legal version is the privacy policy; nothing here extends or narrows it.

Your documents are not training data

Documents you upload are processed to deliver the feature you asked for and nothing else. They are not used to train AI models. That holds on every plan, including the free one.

You choose which provider sees a request

To answer a question, the content of your request goes to the model provider serving the model you picked. Nothing switches provider on your behalf. The default is fixed, and a request that arrives without naming a model falls back to a fixed model rather than to whatever happens to be cheapest or fastest that day.

What matters here is who operates the endpoint, not who trained the model. Those are different companies more often than not. The open-weight models coming out of Chinese labs, Kimi and GLM and MiniMax among them, are mostly served from the United States by American operators, so choosing one of those models does not send your request to China. The table below sets out which operator receives a request for which models, and which country each operates from.

A few operators do run from outside the United States. They are visible on every plan and anyone can select one, but they are never the default and never the fallback. A request reaches one only when you switch to it yourself. Paid plans can also run requests against your own provider account using your own API key.

Two caveats on that column. It says where an operator is based and whose law it answers to, which is not the same as a guarantee about a named datacenter, and most providers reserve the right to process across regions. Novita AI is the one case where even the country is unknown: it does not publish where its serverless API runs, its GPU fleet spans regions worldwide, and its policy permits transfers outside your region under standard contractual clauses. A request to a model it serves can therefore land somewhere it does not name. If that is a problem for your work, pick a different model or ask us about Enterprise data residency.

The client that makes those calls is mohdel, written and maintained by AI For Verticals and published under the MIT licence. It calls each provider's own API using that provider's official SDK. Nothing resells the request or routes it through a broker in between, and because the client is ours and its source is public, what leaves our infrastructure can be read rather than taken on trust.

Encryption and isolation

• Encrypted in transit with TLS, and at rest with AES-256
• Row-level security on every database table, so a query carries the identity of the account making it rather than relying on application code to filter
• API keys and credentials held in encrypted secret storage
• Regular security reviews and access audits

Row-level isolation is the one worth understanding, because it is the difference between "our code remembers to check who you are" and "the database refuses to return another tenant's row even if the code forgets". The check sits under the application, not inside it.

Deletion is yours to trigger

A document is retained until you delete it or close your account. Delete it and it goes; there is no archived copy kept for analysis.

Around that, the policy sets outer bounds: account data is deleted within 30 days of closing your account, usage logs are kept up to 12 months and then aggregated or deleted, and payment records are held as long as tax and accounting law requires. Free accounts that go inactive have their stored documents cleared on the schedule shown below.

Who else touches your data

Running the service means a short list of vendors sees some part of it. Each is bound by a confidentiality agreement, and each does one job:

• OVHcloud provides the object storage your documents sit in and the dedicated servers that process them. The compute is bare metal rather than shared cloud instances, so the machines running your documents are not co-tenanted with anyone else's workload
• Supabase hosts the Postgres database holding accounts, billing records, and document metadata
• Google Cloud Document AI reads the pages you send through Enhanced OCR, the paid per-page pass for scans the free one could not make sense of. It sees nothing unless you run it
• Cloudflare serves this website
• Postmark delivers transactional email such as sign-in links and receipts
• Stripe processes payments and holds card details, which never reach us
• Model providers receive the content of a request when you ask a question, as described above

Automatic OCR is the exception worth naming. Every scanned upload gets read on arrival, on every plan, and that pass runs on our own servers using our own software. Nothing leaves for it. A scanned document only reaches an outside OCR service if you deliberately run Enhanced OCR on it afterwards.

Where your data is processed

Data is processed primarily in the United States. If you are outside the US, your information is transferred there. Enterprise plans can arrange data residency options, and Enterprise also supports SAML single sign-on so account access is governed by your own identity provider.

Reporting a vulnerability

No system is fully secure. If you find something, send it to [email protected] and it will be looked at.

At a glance

Which operator serves which models

A model's creator is not always the company that runs the endpoint. Open-weight models from Chinese labs are mostly served from the United States by American operators, so picking one does not mean your request goes to China. The middle column says where each operator is based.

Per-plan figures are listed on compare plans. The binding legal text is the privacy policy and the terms of service.