v1.2.4

[v1.2.4] - 2026-07-04

Compatible with Open WebUI v0.10.0

Caution

This release fixes three bugs that could destroy live user data (a storage-wipe path, shared
knowledge bases being gutted when their uploader's account was removed, and live KB search
embeddings being deleted on every real run) plus a bug that made age-based chat pruning
silently ineffective on SQLite, the default backend.
Upgrading is strongly recommended for all deployments. Update before your next run.

Important

⚠️ prune-open-webui will be slowly PHASED OUT in favor of an Event Function for native Open WebUI compatibility ⚠️

This allows for:

1. a UI inside Open WebUI
2. automatic deletions (e.g. if a chat gets deleted -> go check if any other chats can be deleted based on age fully automatically)
3. easier setup
4. more compatibility across deployment types (e.g. Kubernetes pods)
5. less hassle when installing, as it is a plugin and doesn't require special setup
6. decide between automatic deletions, on-demand deletions or both!

➡️ You will be able to find the Event Function here when it releases.

Added

• --orphan-file-grace-hours N (default: 24, 0 disables): files younger than N hours are never
  classified as orphaned. Uploading a file and referencing it (sending the message, adding it to a
  knowledge base) are two separate requests in Open WebUI, so a prune running in that window could
  previously delete an upload the user was still attaching. Applies to the standalone script, the
  interactive CLI and the preview counts.

Changed

• Orphaned-file rule is now reference-based only. A file is deleted only when nothing references
  it (no knowledge base, chat, folder, channel message or model). Previously a file was also deleted
  whenever its uploader's account no longer existed, even if the file was still part of another
  user's live knowledge base or chat: offboarding a user could silently gut shared knowledge bases.
  Who uploaded a file no longer plays any role in whether its data is still needed.
• Inactive-user pruning no longer deletes the user's files eagerly. The reference-based rule
  applies instead: uploads backing another user's live knowledge base, chat or channel survive
  account deletion, and the user's own unreferenced files are reclaimed by the orphan pass of the
  same run.
• Vector collections now follow file rows, exactly like storage bytes. Any file row that
  survives the sweep (referenced, inside the grace window, or deferred to the next run) keeps its
  embeddings. Previously a grace-protected upload kept its row and bytes but silently lost its
  vectors in the same run, leaving a file that looks intact but returns nothing in retrieval.
• Storage-orphan sweep preserves every existing file row. The upload cleanup now protects the
  storage object of every file row in the database (unreferenced rows are removed by the database
  pass first). Paths are matched raw, normalized, and by basename, so a non-canonical UPLOAD_DIR
  or a migrated data directory (pip to Docker, moved volume) can no longer cause every ref to
  mismatch and classify all uploads as orphans.
• Synchronous vector-store and storage calls inside async code paths (per-file collection drops,
  bucket listing, physical deletion, preview counting) now run in worker threads, so the engine no
  longer stalls the event loop when embedded in a server process.

Fixed

• A transient database error during the storage sweep could delete every upload. The active-path
  lookup swallowed all exceptions and returned an empty preservation set, causing the sweep to treat
  every object in local/S3/GCS/Azure storage as orphaned. It now aborts the pass instead.
• SQLite: attachments of deleted chats were never reclaimed. Open WebUI does not enable SQLite
  foreign-key enforcement, so deleting a chat leaves its chat_file rows behind, and the reference
  scan trusted them: the attachments stayed pinned as "in use" forever, no matter how many sweeps
  ran. Dangling chat_file/knowledge_file/channel_file rows are now cleaned before every sweep
  and the reference scans verify the parent row still exists.
• Knowledge-base embedding cleanup on file deletion never actually ran. It called a method that
  exists on no cleaner class; the error was swallowed at debug level, so deleted files left their
  chunks behind in every referencing knowledge-base collection. Deletion now goes through the
  underlying vector client.
• Completed the shared knowledge-bases collection preservation fix (#37, #39) for the Milvus and
  Qdrant multitenancy modes, which were still deleting the search-metadata embeddings of live
  knowledge bases on every real run.
• Pruned users left their login credentials in the database. Inactive-user deletion went through
  a code path that removes the user row but not the auth row, so the email and password hash of
  every pruned account survived forever. Deletion now uses the same path as Open WebUI's own admin
  delete, which removes both.
• Legacy AI-generated images in live chats were deleted. Older chats reference generated images
  only by URL (/api/v1/files/{id}/content) with no id field, which the reference scan could not
  see. URL-style references are now recognized, and the channel_file junction is scanned too, so
  channel uploads whose message has not been posted yet are no longer classified as orphans.
• Preview counts now match execution exactly. Rows with a NULL owner were deleted at execute
  time but never counted in the preview; on large instances the per-table counts could exceed
  SQLite's bind-parameter limit and silently report zero while execution deleted normally; and
  messages inside channels deleted as orphaned in the same run were missing from the totals. All
  counts now stream rows with the same filters execution uses, and the age-based and orphan chat
  passes only count chats whose deletion actually succeeded.
• Orphan sweeps silently under-deleted behind paginated API calls: only the 30 most recent knowledge
  bases and 50 most recent notes were ever examined, and a single legacy prompt row (NULL tags)
  aborted the entire run. These passes now stream raw rows.
• Orphaned knowledge bases (owner deleted) are now de-referenced from every model's knowledge list,
  as the age-based KB path already did, so shared models no longer point at a deleted collection.
  The interactive CLI also gates the KB row deletion on the vector drop succeeding, matching the
  other frontends.
• GCS/Azure storage objects with nested names (containing /) are skipped during orphan cleanup:
  the deletion helper extracts keys naively and could have removed a different, live blob in a
  shared bucket.
• Standard-mode Milvus cleanup failed on a fresh process because the pymilvus default connection
  alias was never established; it is now connected on demand.
• The --help epilog advertised a --delete-orphaned-files flag that does not exist.