AviAvi privacy policy
Last updated: 2026-10-10
Avi is a Discord moderation bot. A server's admins add it, and it moderates that server by the server's own rules. This policy explains what Avi stores, why, for how long, and who else sees it.
What Avi stores
| Data | Why | Kept for |
|---|---|---|
| Messages in channels Avi can see: text, author id, channel, reply target, attachment and link info | Context for judging a reported or flagged message, undo, the weekly report | 14 days |
| The server's rules, compiled from its rules channel, and any drafted rulebook | Judging by the server's own rules | While Avi is in the server |
| Moderation cases: the rule, the action, the evidence message ids, the reason, appeals and their outcome | Undo, appeals, mod review, learning how this server's mods rule | While Avi is in the server |
| Precedents: past rulings with pseudonymized message excerpts and an embedding of them | Ruling consistently with the server's past decisions | While Avi is in the server |
| Member statistics: join date, message counts, past cases, and how often pairs of members talk to each other | Telling banter between friends from harassment of strangers, spotting raids | While Avi is in the server |
| Scam indicators: domains, message fingerprints, image hashes and user ids reported as scams | Blocking scams | While Avi is in the server |
| Server settings and model usage (tokens and cost per call, no content) | Running the bot and enforcing budgets | While Avi is in the server |
Avi does not read direct messages. When a member reports a scam DM with /report-dm, a model describes the screenshot and the description is kept in the case. The image itself is not stored.
Models
Avi uses large language models on DigitalOcean Serverless Inference to screen messages, judge flagged ones, compile rules and read /report-dm screenshots. Before text is sent, members and their mentions are replaced with labels such as U1 and U2. The rest of the message text is sent as written. Avi never uses server data to train models.
The scam network
A server can opt in to the scam network with /avi. When it does, scam indicators its mods confirm (domains, message fingerprints, image hashes and the scammer's user id) are shared with other opted-in servers, along with the ids of servers that confirmed them. No message text is shared. Network flags only raise suspicion; they never ban anyone on their own. Indicators other servers rely on stay in the network after a server leaves.
Who sees what
Server mods see cases, evidence and reports in the server's mod-log channel. Members see their own notices and appeal results. The Avi operator can access the database to run and fix the service. Data is not sold or shared with anyone else.
Deletion
When Avi is removed from a server, everything about that server is deleted 30 days later, except the shared network indicators described above. Admins can ask for earlier deletion, and members can ask for their data to be deleted, by emailing support@sorv.dev.
Changes
Changes to this policy are dated at the top.