You are putting privileged client material into this application. You are entitled to know precisely what happens to it. This page says so in specifics rather than in adjectives.
Nyayika runs on dedicated infrastructure with a self-hosted Postgres database and a private object store. There is no third-party backend-as-a-service holding your files, and no shared multi-tenant vendor account between us and your documents.
Every table that holds your work — documents, chunks, clients, matters, chat history, usage — has row-level security enabled and a policy restricting access to the authenticated owner. This is enforced by the database itself, not by application code that could be bypassed by a bug in a route handler. Uploaded files live in a private bucket under a per-user folder, with storage policies scoped to that folder.
A separate set of constraints stops a request from attaching your work to another user's matter or client record, even if a caller tampers with the identifiers it sends.
When you run a Quick Tool or send a chat message, the text of that request — including the case files you attach to it — is sent to a third-party AI provider to generate the response. Nothing else is sent, and nothing is sent at all until you run something.
Requests go to OpenAI first. If OpenAI is unavailable, the request falls back to OpenRouter and then to NVIDIA's hosted API, in that order, so the tools keep working through an outage. OpenAI does not train on data submitted through its API. The fallback providers are free-tier services and their terms are less restrictive — treat a fallback response as having been processed under weaker guarantees. The interface tells you when a response came from a fallback provider rather than hiding it.
If that trade-off is unacceptable for a particular matter, do not attach the file. We would rather tell you this plainly than claim a guarantee we cannot enforce on someone else's infrastructure.
Your documents, drafts, and client records are never used to train any model by us, and are never sold, shared, or used to build a shared corpus across accounts.
Deleting a document removes both the stored file and its search index entries. The database is backed up on a schedule; backups are retained for a limited window and then discarded, so deleted material can persist in a backup for that window before it ages out.
Everything the AI produces is a draft for a qualified advocate to check. Citations in particular can be wrong or invented. The tools are instructed to flag authorities they are not certain of, but that is a mitigation, not a guarantee. Verify against the bare act and the reported judgment before filing.
Questions about any of this? See what each plan includes.