Guides / AI models and where they run
Where your text goes
Two explanation cards: control of a text comes down to who can read it, where it is processed and under which rules, with the GDPR as the example; AV gives controls and records decisions, and the decisions are yours.
Before you start
- N/a (for local translation: an AI model server such as Ollama on your PC).
Who reads your text, and where
Two explanation cards: control of a text comes down to who can read it, where it is processed and under which rules, with the GDPR as the example; AV gives controls and records decisions, and the decisions are yours.
Step by step
Three questions about your text
Keeping control of a text comes down to three questions: who can read it, where it is processed, and under which rules.
GDPR as the example
This guide uses the GDPR’s ideas as the example. AV gives you controls for each and records some of your decisions; the decisions, and whether they are enough, are yours.
The controls in AV
The controls for where text goes: an AI model server on your own PC connected under Integrations and used to translate locally; the GPU Region for rented GPUs; Classify Integrations as a record of where each sends content; each project's recorded confidentiality decision, with masking shown going out and coming back; local PDF routes; and where AV keeps its data.
Step by step
Open Integrations: You start it or AV starts it
Under Integrations, each configuration says who starts the server: You start it for a server on your own PC or network, AV starts it for a rented GPU (Guide R). API providers and their keys are in Guide C.
Click Create New Configuration, You start it, An AI model server on this PC or your network
If you run an AI model server on your own PC or network (Ollama, LM Studio, llama.cpp’s server), connect it here: Create New Configuration, You start it, then An AI model server on this PC or your network.
Choose Local Machine, enter the address, click Test, choose an AI model, Finish & Save
Give it a name and choose Local Machine. The address is the server’s own, here http://localhost:11434. Test asks the server for its models; choose one and Finish & Save.
Click the plug to connect
The plug connects AV to the server. Nothing is started or billed: the server is yours.
Check Running services: level Internal
Running services shows it, at its localhost address. Its level is Internal: the text stays on your PC.
In the project, choose vLLM, Custom (self-hosted) and press Ctrl+Shift+M
In the project the roles are on vLLM → Custom (self-hosted), so Ctrl+Shift+M sends the segment to the server on your PC and nowhere else. A small model may translate less well than the large API models: here it wrote Bitte bitten Sie. Check it on your own text.
Set the GPU Region for rented GPUs
For rented GPUs, the GPU Region decides where AV may start a server. EU/EEA only is the default; it includes datacenters in Iceland and Norway, which are in the EEA but not the EU. API providers use their own hosts: AV has no region setting for them.
Use Classify Integrations to record where each sends content
Classify Integrations records where each one sends content: Internal, External · private (a provider you contract), or External. It is a record for you: AV does not act on it. What is masked is set per project.
Check the project's decision in Privacy Configuration
Each project records a confidentiality decision under Privacy Configuration. The claim uses an anonymizer profile for names and addresses; further down, a pattern profile masks emails, phone numbers and IBANs (Guide E).
A decision to send as written: No anonymization, with the reason
A public price notice needs no masking. The project records why: No anonymization, with the reason and a note. It is a statement: AV records it, and blocks nothing.
Translate: sent masked, restored in the grid
In the claim, the names, the address, the email, the phone number and the IBAN go to the AI model as stand-ins and placeholders, and come back restored in the grid.
Click Copy anonymized to see what the AI model received
What the AI model received. Copy anonymized in the editor puts the same masked text on the clipboard, so you can check it before anything is sent.
PDF Preparation routes that stay on your PC
PDF Preparation has routes that run on your PC or network once you set them up, and upload nothing (Guide P). The cloud routes send the whole PDF, unmasked: each card says which kind it is.
What AV keeps, and where
Everything AV keeps is on your PC. Only the key store is encrypted, so protect the storage folder as you protect the documents.
What still leaves
Explanation cards listing the connections AV makes that carry no document text, and those that carry text by design, followed by what remains the user's responsibility: AV does not make anyone compliant.
Step by step
What leaves without your text
Some connections are made whatever you do. None carries document text.
What leaves with your text
And what goes out with text, by design. A guide that only listed protections would not be telling you everything.
What is still yours to do
AV gives controls and keeps a record. It does not make anyone compliant. This guide is not legal advice.
In short
Video recorded on 5 November 2026 with AV 0.2.11. What AV sends and stores is as of that version; check the provider’s own terms before you rely on them.
Recorded on 5 November 2026 with AV 0.2.11. The videos replay AI model answers recorded earlier, so what an AI model writes for you may differ.