Skip to content

Data erasure & GDPR

DioscHub gives administrators the tools that map directly to data-protection duties: find a person’s data, give them a copy of it (access, Art. 15, and portability, Art. 20), erase it (Art. 17), and delete it automatically after a period (storage limitation, Art. 5(1)(e)). All of them work on every tier, Free included. None is a paid feature, and no licence capability turns them on or off: they are governed by the data-privacy:read and data-privacy:execute permissions only.

This page states the exact scope of erasure: what it removes, how it behaves under failure, and what it retains by design. That last part matters. A compliance review is better served by a precise boundary than by a blanket “everything is deleted” claim — so the carve-outs below are stated, not hidden.

  • Find a person’s data. A subject-access search is scoped to a single Assistant: it returns content from within that assistant’s boundary, not across every assistant in the deployment.
  • Give a person a copy of their data. An Owner or DPO can build a zip of everything DioscHub holds about one person within one assistant. See Subject access export.
  • Erase at three grains. An operator can erase a single session, a single file, or an entire user. Erasing a session also erases the files held by that session — the files an anonymous visitor uploaded (see anonymous file upload). A signed-in user’s library files are not tied to a session; erasing the file or the user removes them. To erase a subject, find them by user id on the Erase a subject tab of the Data Privacy page. Anonymous visitor ids (anonymous, anon_…) can’t be looked up there, as they have no lasting identity; erase an anonymous conversation as a session from Search & act instead.
  • Delete old data automatically. The retention window deletes conversations and users’ files once they have gone unused for a period you set, through the same path as erasure. It is off by default.

Erasing a user scrubs the identifiers and content that could re-identify them, across the stores that hold them:

  • The conversations — the message history for the user’s sessions, and the session records themselves: name, user id, external id, the visited pages, the message preview, and the forwarded auth artifacts held for the session.
  • The derived data — the user id on usage and analytics records, the person’s linkage in aggregated signals, and the summary of each conversation (written by the language model) together with its search embedding. Analytics classification runs keep their record of when they ran and what they counted, but lose the link to the session and any error text that names it.
  • The operational records — error-log messages tied to the session, the user’s identifiers inside approval records, the user’s rate-limit records, and the search access rules that name the user.
  • Problem reports — the user’s problem reports for the assistant. The problem they belong to keeps its report count, and the Problems page shows how many reports were erased. A report.erased event is sent to every webhook sink that received a report; the sink’s own copy is yours to erase. Erasing a single session deletes that session’s reports the same way. The retention window does not delete reports: once it removes a conversation, that conversation’s reports stay, detached from the session, and this is also true when the window retires an inactive user. Only a user erasure removes them.
  • The files and their index — the user’s uploaded files (both private chat files and any they own), their stored bytes, the text extracted from them, and every version of the search-index chunks and permissions derived from them (see Files: stored text and indexing).
  • The identity — the user’s own external identifier is replaced with an irreversible, one-way tombstone (see below), not left in place.
  • Their exports — any access export of the person’s data that is queued, being built or ready to download is deleted, file and record.
  • Each attempt is transactional. The database work for an erasure runs as a single transaction and is retried as a unit, so a failure rolls back rather than leaving a half-scrubbed record behind.
  • It makes forward progress. Work already committed to one store is never undone because a different store failed — reverting a committed scrub would put personal data back. Removal from the search index and the file store runs as best-effort with retries, followed by a verification pass that confirms no re-identifiable data survived and raises an error loudly if any did. Treat this as forward progress plus verification — not a single atomic operation spanning every store at once.
  • The tombstone is one-way. So you can still answer “was this person erased?” without keeping their data, the user’s external identifier is replaced by a keyed one-way hash. It cannot be reversed to the original id, and it depends on a per-deployment erasure secret you configure (DioscHub refuses to start in production without it — see production hardening).

A file holds personal data in three places: the file itself, the text DioscHub extracted from it (kept in the file store, next to the file), and its search index, which can hold more than one version while the file is being re-indexed. A spreadsheet in a user’s library has three more: its workbook outline (sheet names, headers and a preview of cells), the tables the assistant loaded from it, and the pictures extracted from it (see Spreadsheets). Erasure and redaction reach all of them. Before either touches a file, it stops any indexing of that file that is still running, so a run that started earlier can’t write the old content back afterwards.

  • Indexing is stopped first. Erasing a file, a session or a user cancels the indexing of the files concerned and waits for any run in progress to stop, before anything is deleted or anonymised. The wait is bounded: 60 seconds by default (PRIVACY_INDEX_CANCEL_WAIT_MS, in milliseconds). If a run doesn’t stop in time, the erasure fails closed with the step stop_indexing failed as index_job_still_running. Nothing has been deleted, and you can run it again.
  • Erasing a file removes the stored file, its extracted text, every version of its search index, its permissions, its indexing jobs, and the record, and for a spreadsheet its outline, its loaded tables and its extracted pictures. Deleting a file from the Files dialog, its expiry and an anonymous session’s end remove the outline, tables and pictures the same way. For a knowledge-base file that was being replaced, the incoming version goes too. The verification pass checks that no stored file, no extracted text, no index of any version and no queued or running indexing job is left; any residue fails the erasure. The user’s storage usage stops counting the file.
  • Redacting a file stops indexing the same way, then deletes the file’s whole search index before it reads anything. It rewrites the extracted text, and a plain-text or JSON file itself, as a stream, replacing the terms. A binary file (PDF, Word, an image) has its original bytes deleted, and its redacted text stays as the file’s content. A file that was ready or being indexed is then re-indexed from the redacted text, and the verification pass checks the extracted text, the file and every version of the index for the terms.
  • Redacting a spreadsheet also deletes its loaded tables, its extracted pictures and its outline. The outline is deleted rather than rewritten, and the verification pass checks that it is gone. The redaction doesn’t build a new outline: for a redacted CSV, a Re-index builds one from the redacted file, and its tables are loaded again from the redacted file. An Excel workbook’s original bytes are deleted like any binary file’s, so the table tools answer that its source has been removed.
  • Subject-access search reads the search index and also the stored text of files that aren’t indexed: files whose search index couldn’t be built or stopped at the storage quota, and files still being indexed for the first time. One search reads the stored text of at most 500 such files (PRIVACY_DSAR_TEXT_SCAN_MAX_FILES). When there are more, or one of them can’t be read, the results say they are partial (storedTextTruncated: true) and the portal shows a notice; narrow the search to one assistant. Hits found this way are marked Stored text (not indexed) (source: "stored_text"). On a deployment that is still moving file text to the file store after an upgrade, the search, redaction and verification also cover the old database copy until it is gone.

Two limits to know. A file whose erasure or redaction failed with index_job_still_running can be left showing as being indexed until you run the erasure or redaction again. And after a redaction deletes a binary file’s original bytes, a download of that file fails with a server error rather than a clear “removed” response; the redacted text is what the assistant reads.

Erasing or redacting a file doesn’t reach conversations. Text and images the assistant already read from the file stay in the message history of every session that read them, until that session, or its user, is erased; redacting the session with the same terms replaces them in its text. See Read file content stays in the conversation.

Images, workbook pictures included, stop reaching the model once their file is gone. Before every model call, DioscHub checks the file behind each image already in the conversation. When that file has been erased, deleted, redacted, has expired or can’t be checked, the image is replaced with a short note that it is no longer available, so the model’s provider stops receiving it. The image itself stays in the stored message history until the session, or its user, is erased.

Erasure deliberately keeps a few things, and a review should know exactly what:

  • De-identified aggregates. Cross-user patterns that no longer point at any individual are kept; only the link between the person and the aggregate is removed. What remains describes a population, not a person.
  • The tombstone record. The user’s identity row persists under its irreversible tombstone — that is the erasure marker, holding none of the original data.
  • Non-subject operational events. Records keyed only to an IP or a network window, not to a person, are cleared on their own schedule rather than by a subject erasure.
  • The administrative audit trail. The record of who erased what is retained, and it holds no chat user’s data — see Audit trail.
  • Erasure acts on the live datastore. Your backups sit outside this operation; their contents are governed by your own backup and retention policy, not by an erasure you run against the running system.
  • Erasure is performed on request. An administrator runs it against a subject; it is the mechanism you invoke to meet an erasure obligation, applied when you invoke it. Deletion after a period is the retention window, which you turn on separately; erasure still removes conversations the window keeps, pinned ones included.

Next: Administrator roles & permissions.