Frequently asked questions

The workspace where your company keeps what it knows. Documents live as Markdown files inside a folder that belongs only to your company, organized by department, with per-person permissions. On top of those files runs an AI agent that reads and writes them under the same rules a teammate follows, and every bit of usage is metered company by company.

Three concrete things. The agent works inside your files: it opens the department's main note, its memory and its open tasks, and writes back in the same place. What it learns stays in your vault as dated, sourced text rather than in a vendor's opaque memory. And every call is metered and audited per company, with a hash chain that makes any tampering with the record obvious.

No. Each agent chat is confined to its own project's file subtree. Parent projects and connected projects are not mounted: the agent knows them only through the summary we inject, which is the description, the memory, the open tasks and a file listing. We hardened this after finding and fixing a case where a session with no assigned path mounted the whole company folder.

Yes. Search runs two queries at once, one by meaning over vectors and one by exact term, then fuses the two rankings. That is how the minutes that mention early termination surface when you typed cancel the contract sooner. Both queries carry your company identifier inside the SQL, not as a filter applied afterwards. Indexing and searching consume no credits, because the embedding model runs on our own server.

Two tiers. The company role, from owner down to viewer, and the seat you hold in each department. A manager of a parent department also manages its children, because authority inherits down the tree. Sharing a single file only ever adds access, never removes it, so an exception cannot accidentally lock someone out.

No. A personal note is private from your company's owner and from its administrators. You see it, and the platform administrator only when there is a support case in play. If you want to share it, you do that yourself, with per-person permissions.

Yes. A scheduled task takes a cron expression or a named interval: hourly, daily, weekly or monthly. It runs the prompt you wrote acting with your real role, so it cannot touch anything you could not touch, and it consumes credits like any other agent turn.

One credit is a tenth of a US cent of real model spend. It is not an invented currency with a hidden margin, it is the unit we use to measure what your usage actually cost. On our own installation an agent working conversation has averaged USD 0.19, which is about 190 credits. So the 40,000 credits on Basic are roughly 210 conversations of that kind, and the 90,000 on Pro roughly 470.

Only the agent's work consumes credits. Uploading documents, indexing them, searching them, walking the graph, re-reading past conversations and messaging a colleague consume nothing, because all of that runs on our own server without calling a paid model.

Today usage is recorded and shown, but it does not cut off automatically. You see the budget overrun in Settings, Usage / AI, and we tell you. An owner requests more credit from that same screen and a platform administrator approves or denies it with a note. The automatic cutoff is written and tested but not yet wired into the agent path, and we would rather say that here than promise the opposite.

Not by itself. Your plan's monthly credit is applied to your account by a person. No automated job tops it up on the first of the month, and we are not going to claim one exists until it does. If your usage is steady we load it ahead of time so you never have to think about it.

Not for the Trial. Signing up charges nothing and asks for no card: you leave your company details and a work email, verify the email, and a person creates the account and writes back. On Basic and Pro payment is made by card in the checkout, off this site, and while that checkout is not enabled the button leaves you on the same form and we settle it by email.

No, and it is blocked in four independent layers. Every row carries your company identifier. PostgreSQL enforces row level security in FORCE mode across 23 tables, keyed on the identifier taken from the verified session and never from the request body. Your company's folder is mode 0700. And the graph uses one namespace per company. A dedicated test suite tries to cross those boundaries on every change.

No, and we would rather say so. No SOC 2, no ISO 27001, no external third-party audit. What we do have can be inspected: four-layer per-company isolation, argon2 password hashing, instant session revocation on logout, encrypted secrets that no endpoint ever returns, code execution in containers with no network and all capabilities dropped, and 55 test suites that run on every change, one of them dedicated to breaking cross-company isolation.

Passwords are hashed with argon2. A session uses a short access token of fifteen minutes and a refresh token of seven days, signed with distinct secrets. The refresh token carries a version number, so logging out invalidates every live session on that account instantly. Login and refresh are throttled per IP against brute force, and public self-registration is disabled.

In a rootless container created for that run and destroyed when it ends. No network, read-only root filesystem, all capabilities dropped, no privilege escalation, 512 MB of memory, one CPU, a process cap, an unprivileged user and a twenty-second wall-clock kill. A viewer or a member without permission never even starts the container: the request is refused first.

Yes. Every tool the agent dispatches leaves a row with the hash of its arguments, whether it finished cleanly, how long it took, and the previous row's hash chained into this one. Editing or deleting a row breaks the chain and shows. Raw arguments are never stored, only their hash. Separately, every AI call appends a line to a monthly log file on disk that you can download.

In a folder exclusive to your company on the server, at mode 0700. No other tenant on the platform can read it, not even at the filesystem level. Inside there is standard Markdown and the original files you uploaded. The search index and the graph are rebuildable views of that folder: delete them and they regenerate.

Yes, and it is a consequence of the design rather than a contractual promise. Your notes are .md files with wiki links and YAML metadata, readable in any text editor. The database is standard PostgreSQL and there is a daily pg_dump backup. We hand you the folder and the dump, with no paperwork and no exit fee.

Ofivia does not train models, not on your documents and not on anyone's. Indexing and search run on an embedding model hosted on our own server, so that text never leaves the machine. When the agent works, the portion of content needed to answer that query does go to the model provider. The processing agreement annex lists the processors involved and why, and if your company needs the no-training commitment in writing before signing, ask ventas@ofivia.com.

Several nets. Deleting a folder is a soft delete and it lands in the trash. A project's main note cannot be deleted, moved or renamed, it is protected by design. There is a daily database backup with fourteen-day rotation, plus hardlink snapshots of the vault so each snapshot costs only what changed that day. And every agent action stays in the chained log.

We do not promise data residency in a specific country, because that promise depends on where each installation is hosted and on the model provider. If jurisdiction is a requirement for your company, the honest route is self-hosting: install Ofivia on the server and in the country you choose. Tell sales where the data needs to sit and we will confirm in writing whether the managed installation meets it, before anything is signed.

Leave your company details and a work email on the signup form. A verification link arrives. Once you confirm it, a person creates the company, the first owner and the starter credit, and replies with your access. It is deliberately not self-serve: we would rather the first conversation happen.

The account is ready in under one business day. The rest is what takes time: deciding the department tree, who sees what, and which documentation is worth loading first. At a mid-sized company that is usually one or two weeks of shared work, and it is the part that decides whether the agent is useful or just talkative.

Yes. Upload files and folders and the ingest queue handles the rest: it chunks, indexes, builds the graph and drops byte-identical duplicates. An unchanged file is never reprocessed. What does not exist today is an automatic connector for Drive or SharePoint, so the initial load happens through the interface.

Yes, that is the Enterprise tier. The whole platform comes up with one idempotent command that provisions the database, the embedding server, the cache, the graph and the backup timers. Every service binds to 127.0.0.1 only and is published behind a reverse proxy whose configuration ships with it.

Email soporte@ofivia.com. In practice we answer the same business day, and much faster when the problem stops someone from working. We do not publish a service level agreement with guaranteed hours unless your contract includes one, because we will not put up a number we cannot hold.