Skip to content

Subject access export

When a person asks for a copy of their data (a subject access request, GDPR Art. 15), an administrator can build it from the Data Privacy page as one zip file. The files inside are JSON, so the same export also answers a request for data portability (Art. 20).

The export works on every tier, Free included. No licence capability turns it on or off. Like erasure, it is governed by permissions only.

Making an export, following its progress and downloading it all need the data-privacy:execute permission. Among the built-in roles, Owner and DPO have it (see Administrator roles & permissions). An administrator who can only read the Data Privacy page doesn’t see the Export action, and the hub refuses their export requests.

  1. Open Data Privacy in the admin portal and go to the Erase a subject tab.
  2. Pick the Assistant, enter the External user id (the id your application sends for the person), and select Find.
  3. When the person is found, select Export data. The page shows the progress, for example Exporting conversations (3 of 10). The export carries on if you leave the page: find the person again and select Export data to get back to it.
  4. When it’s ready, the page shows its size and the time it is available until, a reminder to review it, and Download zip.

One export covers one person within one assistant, the same unit as Find and Erase. Preferences are the exception: they are stored per person, not per assistant, so the export includes all of them. If a request covers several assistants, export from each one.

What else you may see:

  • Nothing found. Nothing found for this subject on this assistant. No export was made. Check the assistant and the user id.
  • Already erased. This subject has been erased. There is nothing to export. No file is made.
  • An export already exists. While an export for the same person is queued, running or ready, asking again shows that export instead of starting a second one.
  • Anonymous visitors (user ids anonymous or starting anon_) can’t be looked up or exported. They have no lasting identity, so there is no way to check that a requester is the person in the data.

The export runs as a background job. In a cluster, any instance can build it, and it is written to the deployment’s configured file storage, so it can be downloaded from any instance.

FileWhat it holds
conversations.jsonEvery conversation the person had with the assistant: its id, the assistant (id and name), when it started (createdAt), when it was last updated (lastMessageAt, which moves on any change, such as pinning), its last message activity (lastActivityAt, the clock the retention window counts from), its status, whether it is pinned, and the messages in order, both sides, with the tools the assistant called and their results
files.jsonThe files the person uploaded: name, type, size and when
files/The original bytes of those uploads
analytics.jsonThe per-conversation analysis: topic, outcome and summary
feedback.jsonThe ratings and comments the person gave on answers
approvals.jsonThe actions the person was asked to approve and what they decided
usage.jsonThe usage recorded for the person’s conversations
errors.jsonThe technical error records of the person’s conversations, where a tool server or the language model failed: the tool or model, the error category and status, and the error text (redacted)
problem-reports.jsonThe problem reports the person filed with the assistant, including ones whose conversation retention has since deleted: title, summary, steps, expected, actual, excerpt, page path, failed calls (redacted), source and the assistant’s assessment
preferences.jsonThe person’s saved preferences, for every assistant
audit.jsonPrivacy actions taken on the person’s data, with the administrator named by role only
profile.jsonWhat the assistant holds about the person: name, email, roles, attributes, first and last visit
README.txtThe person, the assistant, the export time (UTC), each file with its item count, what is not included and why, and the review note

Each JSON file has the same shape: { "subject", "exportedAt", "items": [...] }.

  • Tool calls are redacted. Tool arguments and results go through the same redaction as the hub’s logs: sensitive parameters and credential-like text are replaced before they are written.
  • File names are made safe. Names in files/ have path separators, .., drive or root prefixes and control characters removed, are cut to 200 bytes (the extension is kept), and duplicates get (2), (3). A name is rewritten, never refused.
  • A missing original is reported. If an upload can’t be read from storage, files.json still lists it and README.txt says how many are missing from files/.
  • Credentials. DioscHub stores no sign-in or token for the person. The auth your backend forwards is held only for the life of a session and never written to the approval record (see The credential-blind guarantee), so the export has none to include.
  • Administrator identities. Audit entries name the administrator who acted by role only, never by name, account or address.
  • Derived copies. Extracted text, embeddings, workbook outlines, sidecars and the search index are left out. The originals they came from are included.
  • The assistant’s setup. Prompts, roles and knowledge-base files belong to the assistant, not to the person.
  • Other people’s data. Only this person’s rows within this assistant are included.
  • Operational records. Per-turn timings, short-lived tool-call bookkeeping, the daily usage roll-up, knowledge-gap links, rate-limit counters, and the record of the export request itself. README.txt lists each one with its reason.

Tool results can mention other people: records the assistant looked up while helping the person, for example. DioscHub doesn’t screen the export for third-party data. The page shows a warning above every download: review the conversations and remove anything about other people before you send the export to the person.

  • Download needs your admin session. There is no public link and no token in the URL. A download serves only that export’s file.
  • It is a snapshot. An export holds the data as it was when it was built, and is kept at most 24 hours. A conversation deleted or erased after that moment can still be in an export made before it.
  • It is available for 24 hours after it is ready. Then the zip and its record are deleted by an hourly cleanup, and the page says the export has expired: request a new one. An export that failed or got stuck is removed after 24 hours too.
  • Erasure removes it. Erasing the person deletes any queued, running or finished export, and so does the retention window when it removes an inactive person.
  • Size limit. One export holds at most 4 GB or 65,535 files. A larger export fails with The export is too large for one zip file; nothing is cut short to make it fit.

Each step is written to the data-privacy audit trail: the request (export_requested), completion (export_completed) or failure (export_failed), every download (export_downloaded), and the deletion (export_deleted, with the reason: expired, erasure or failed). The entries name the person by their one-way erasure tombstone, never by the raw user id.


Next: Data erasure & GDPR.