How do you back up a private AI system? Six kinds of data, three copies, one offline, and a restore you have actually tried.
Short answer. Back up everything you could not easily recreate: user accounts and permissions, chats and records, uploaded documents, the audit log and configuration, plus the keys that decrypt them. The search index and model files can be rebuilt, but that takes time, so many offices copy them too. Follow the 3-2-1 rule: at least three copies, on two devices, one of them offsite, with at least one copy kept offline so ransomware cannot reach it. Encrypt the copies, and test a restore on a spare machine on a schedule you write down, because a backup that has never been restored is only an assumption.
01
What to back up
Item
Example location
Rebuildable?
Note
Accounts, chats, settings
The main database (a SQLite file or PostgreSQL)
No
The most important item. Use the database’s own tools, not a copy of a live file.
Uploaded documents
The uploads folder
Only if you keep the originals
Keep originals in your normal file store as well.
Search index (vector database)
A vector_db folder or a database
Yes, by embedding again
Rebuilding takes time that grows with the size of the document set; copying the index saves it.
Audit and model-call logs
A log file or the gateway’s log store
No
Needed for investigations and for any retention rule.
Model files
The runtime’s models folder
Yes, by downloading again
Record exact versions so a rebuild gets the same model.
3-2-1: at least three copies, on two devices, one offsite. The UK’s NCSC describes this as the usual method for resilient backups.
Keep at least one copy offline or otherwise out of reach of the live system. NCSC has seen ransomware encrypt connected drives and cloud storage holding backups, and CISA advises maintaining offline, encrypted backups.
Encrypt every copy, and keep the key somewhere other than the backup itself.
Set frequency from your recovery point objective: if you can afford to lose a day of work, a daily backup is enough.
Write the restore steps down: who does them, in what order, and where the keys are.
Test restores. NIST says backups should be tested regularly to confirm the data is stored correctly and can be retrieved without errors.
03
Backing up a database that is running
Open WebUI’s backup tutorial, a community contribution, lists what its data folder holds: the audit log, a cache, uploaded files, the vector database and the main database file. It warns that the database download in the admin settings is not the recommended route, because with SQLite in write-ahead-log mode recent writes can sit in a separate file, so a copy of the main file can miss them. It suggests copying the database as its guide describes, or using pg_dump on PostgreSQL, and notes that backups are safest on a stopped system at the cost of some downtime.
PostgreSQL’s own documentation names three fundamentally different approaches: an SQL dump, a file-system-level backup and continuous archiving. Choose by how much data you can afford to lose.
04
Backups hold personal data too
A backup is another copy of personal data. Rule 6 of the DPDP Rules, 2025 lists reasonable security safeguards. They include measures for continued processing if data is destroyed or access is lost, such as data backups. They also include keeping logs and personal data for one year to detect and investigate unauthorised access, unless another law requires otherwise. Under the Rules’ commencement clause, rules 5 to 16 start eighteen months after publication on 13 November 2025, so as of October 2026 they are not yet in force.
Decide how long backups live, so that a deletion request does not leave a copy behind for years. This is general information, not legal advice.
05
What ANVE says
ANVE keeps the office’s documents, meeting transcripts and the log of model calls on the box in the office, which is why a backup plan matters. ANVEAI has not yet published ANVE’s backup design: what is copied, where, how often, whether copies are encrypted and how a restore is done. Ask for it in writing, and ask to see a restore performed during a paid pilot.
06
Questions people ask
Can we back up to the cloud?
Yes, if the copy is encrypted before it leaves and you accept that the data leaves the building. NCSC’s guidance on cloud backups says to keep at least one copy offline and to protect the account that can reach it. A fully local design can use a second box or an external drive kept off the network.
Do we need to back up the models?
Not if you can download them again, but record the exact versions and licences. An office that runs offline should keep a copy, because the original download may not be there later.
How often should we test a restore?
Set a schedule and keep to it. Neither NIST nor NCSC gives one number; NCSC says backups should be tested regularly. Test after any major change as well.
07
What is not real yet
No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.
Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.
