Does UniAI send my documents anywhere?
Yes. This is what separates UniAI from the rest of UniDoc, so here it is bluntly.
UniPDF, UniOffice and the UniHTML client all process documents on hardware you control. UniAI does not. It is a hosted endpoint: you upload the PDF to UniCloud, and the text taken out of it is passed to a third-party language model provider under contract to UniDoc in order to produce the structured output.
If your documents cannot leave your infrastructure, UniAI is the wrong tool. There is no self-hosted build and no on-premise option.
What is sent where
| Step | Where it happens | What moves |
|---|---|---|
| Upload | Your client to UniCloud | The PDF file, over HTTPS |
| Text extraction | UniCloud | Nothing leaves; UniPDF reads the pages |
| Structuring | Third-party model provider | The extracted text, plus page images if is_scanned=true |
| Response | UniCloud to your client | The JSON output |
The distinction in the third row matters for a review. The PDF binary is not forwarded to the
model provider. What goes is the text extracted from it, and, only when you set
is_scanned=true, images rendered from the pages. On a born-digital document with the flag
off, the provider sees text and no images.
That is still your document’s content. It is a smaller surface, not a private one.
What UniCloud records about the request
Separately from the content, each call is logged for metering and abuse control. The record includes a generated reference id, the page count charged, the uploaded file’s name, the operation called, a timestamp, and the request’s hostname and IP addresses.
The filename is the one people are surprised by. If your filenames carry case numbers, client names or personal identifiers, that string is stored even though the document body is not. Rename before upload if that matters.
Retention
The metering records above persist, since they are what billing is computed from.
Page images rendered for is_scanned=true are written to temporary storage during the request
and deleted once read. The extraction handler does not write the PDF or the extracted text to
the database.
For contractual retention commitments, including those of the model provider, ask support rather than relying on this page. Behavior described here is what the service does; retention is a matter of contract.
Personal data
Resume extraction is processing personal data almost by definition, and invoices frequently carry names, addresses and tax identifiers. If you are subject to GDPR or a similar regime, that means you need a lawful basis for the processing, a data processing agreement with UniDoc, and the model provider disclosed as a subprocessor in your own records. Support can supply the paperwork.
If nothing may leave your machine
Use a local library and build the structuring yourself. The trade is real work in exchange for the content never crossing your network boundary.
| Need | Local option |
|---|---|
| Text out of a PDF | UniPDF text extraction |
| Tables out of a PDF | UniPDF table extraction |
| Text from scans | UniPDF OCR |
Those give you the text in your own process. Turning it into fields is then your code, which for a fixed set of vendor layouts is often a smaller job than it sounds, and it is auditable in a way a model is not.
Your API key is a credential
It is the only thing standing between the internet and your credit balance, and the endpoint has no other authentication. Keep it in a secret store, not in client-side code, and rotate it in the UniCloud dashboard if it leaks.