Data export

Everything your organization holds on Akumi can be taken out, by you, without asking us. This page is the whole answer to "how do we get our data out", whether you are evaluating the platform, answering a due-diligence question, or leaving.

The short version

Go to Settings > Data export and press Prepare export. It runs in the background, we email you when it is ready, and you download one file from the same page.

That file covers every workspace in your organization: the facts held about each of your end-users, every conversation thread with its messages, your collections, and the metadata of every document you have ingested.

What you get

One gzipped, newline-delimited JSON file (.ndjson.gz). Each line is one record, so you can stream it without loading the whole thing into memory:

{"type":"manifest","organization":{...},"counts":{"facts":1284,"threads":340,...},"excludes":[...]}
{"type":"workspace","workspace_id":"...","name":"Production",...}
{"type":"fact","workspace_id":"...","user_ref":"usr_abc123","content":"...",...}
{"type":"thread","workspace_id":"...","thread_ref":"conv_42","messages":[...]}
{"type":"document","workspace_id":"...","title":"Handbook","checksum":"...",...}

The first line is a manifest carrying the row counts and the list of what is deliberately excluded. Compare its counts against what you parsed to confirm you received all of it, rather than assuming a transfer that ended was a transfer that finished.

What is included

What is not, and why

Your document files are not in it. Documents are listed as metadata: title, collection, ingestion date, size, and a SHA-256 checksum of the text that was ingested. You already hold the originals you uploaded, and the checksum is what lets you match the two up and confirm nothing was altered. Documents ingested before checksums were recorded return null for that field, which is stated rather than hidden.

The knowledge graph is not in it. Its topology, the embeddings, and the entity-resolution data are how Akumi processes your data rather than data you gave us. What comes back is what you supplied, plus the personal data extracted from it.

How long it lasts

A finished export is deleted 24 hours after it is ready. It is a complete copy of your organization's data, so we do not keep one sitting in storage longer than you need it. Ask for another whenever you want; there is no limit beyond a small hourly cap to stop a loop.

While it exists it is stored encrypted, is never publicly linkable, and is streamed to you through the application behind your session rather than handed out as a shareable link.

Who can do it

Owners and admins. It returns everything the organization holds, so it is not a member-level action.

Both preparing an export and downloading one are recorded in your audit log, so you can see afterwards who took a copy and when.

It stays available while you are leaving

If your organization is scheduled for deletion, the export is still reachable during the grace period, from the page explaining the deletion. That is the window in which you are most likely to want it, so it would be a poor time for it to disappear.

Why this is not an API endpoint

Deliberately. Exporting an organization is an administrative act by a person, not something your application does on a request, so it lives in the dashboard next to the other things only an owner or admin can do.

The per-end-user export is on the API, and the difference is the point:

See Recall for the per-end-user export.

Everything else you can take with you

The data export covers Recall. The rest of what we hold is exportable too:

Inference itself is OpenAI-compatible, so moving away is a base_url change rather than a rewrite. We would rather say that plainly than have you discover it.