# Retention and deletion

Treat uploaded sources, file-scoped derivatives, semantic caches, job results,
idempotency records, and infrastructure backups as distinct resource classes.
Deleting one is not a promise that all others have been erased.

- Uploads are organization-scoped and reusable while retained and unsealed.
- An extraction with `ephemeral: true` requires `file_id`, seals it for its final
  consumer, and removes the source and file-scoped derivatives after registered
  consumers finish. New consumers must not use a sealed upload.
- Classify's `ephemeral` option behaves like extraction's: it requires `file_id`,
  keeps classify's OCR and cache entry under the file, and removes the source and
  them when the call or its queued job finishes.
- `DELETE /v1/uploads/{file_id}` seals the upload at once (new consumers get 404),
  returns `409 upload_busy` with `Retry-After` while jobs or requests still use it,
  then deletes the source, its file-scoped derivatives, and the organization's
  content-keyed OCR text, raw Document AI output and extract/locate/classify/split/
  summarize cache entries for the same bytes. Entries written for those bytes later
  are dropped on write. Repeats return 204. It does not delete job records or
  results, idempotency records, review/resolve/infer caches, other uploads of the
  same bytes, or callbacks already sent.
- Locate's `ephemeral` option scopes its job artifacts to the file for later
  cleanup; it does not itself finalize/delete the upload and selects a queued job.
- Summarize supports ephemeral file sources. Their OCR derivatives are file-scoped
  and semantic result caching is disabled for requests with an ephemeral source.
- Previously created unscoped caches and retained results are not retroactively
  erased by a final consumer. Ephemeral is not a zero-data-retention mode.
- Terminal-job DELETE removes that job's stored resources; it does not cancel
  active work. Backup and bucket soft-delete retention may outlive logical deletion.

Passwords used to open PDFs are not retained. **Prepared/decrypted document
bytes may be stored** for asynchronous processing and reusable uploads.

Automatic expiry and backups depend on the deployment's configured lifecycle
rules. Ask the deployment operator for the applicable policy rather than assuming
all self-hosted environments share hosted retention durations. Legal retention
information is published at [Data retention policy](/legal/retention).
