File deletions

[BernhardtMilan]

We are currently testing the pruning tool in our Open WebUI setup, and had some questions/feedback:

1. When using the ‘Preview Changes’ option, is there any way to see the names of the files which will be deleted? So far we can see the total number of files which will be deleted, but no information on what, exactly, these files are.
2. Regarding files which will be deleted as part of running the pruning tool, the tool identifies ‘orphaned files’, ‘upload files’ and ‘vector collections’. What exactly is the difference between these categories? Moreover, are they mutually exclusive? Finally, are these categories dependent on our setup somehow?
3. In our setup, we use Qdrant to store data uploaded to our instance of OWUI (either directly in chats or in knowledge bases). How exactly does the pruning tool interact with Qdrant? Are these files really also deleted from Qdrant itself?

[Classic298]

Answers

1. Seeing file names in preview

Currently the preview mode only shows aggregate counts, not individual file names. This is a valid feature request — adding a --verbose or --list flag that logs which specific files/records would be affected would be useful. For now, if you run with --verbose / -v flag on the standalone script, the DEBUG-level log messages will show some individual items as they're scanned, but there's no dedicated per-item listing in preview mode yet.

2. Difference between orphaned files, upload files, and vector collections

These are three distinct layers of data that Open WebUI creates when you upload a file. They are not mutually exclusive — a single file upload touches all three, and all three need cleanup:

Category	What it is	Where it lives
Orphaned files	Database records in the file table. This is the metadata entry (filename, hash, user_id, path, etc.)	PostgreSQL / SQLite
Orphaned uploads	The actual physical file on disk (the PDF, image, etc. that was uploaded)	Filesystem (data/uploads/ directory)
Orphaned vector collections	The vector embeddings generated from the file content, used for RAG search	Your configured vector DB (Qdrant, ChromaDB, etc.)

When the tool deletes an orphaned file, it actually cleans all three layers for that file: it removes the vector embeddings, deletes the DB record, and removes the physical file from storage.

The separate "orphaned uploads" and "orphaned vector collections" cleanup steps are a final sweep that catches anything that might have been left behind (e.g., uploads that exist on disk but have no matching DB record, or vector collections that no longer correspond to any known file or knowledge base).

These categories are not setup-dependent in the sense that the tool always checks all three. However, which vector database cleaner is used depends on your VECTOR_DB configuration (Qdrant, ChromaDB, PGVector, Milvus, etc.).

3. How the tool interacts with Qdrant

Yes, the tool deletes data from Qdrant directly. Specifically:

In standard Qdrant mode (non-multitenancy): Each file and knowledge base gets its own Qdrant collection named {prefix}_file-{id} or {prefix}_{kb_id}. The tool lists all collections with that prefix, compares them against active file/KB IDs, and deletes entire collections that are orphaned.

In Qdrant multitenancy mode: All data lives in 5 shared collections (memories, files, knowledge, web-search, hash-based), with a tenant_id field distinguishing different logical collections. The tool scrolls through each shared collection, identifies tenant_id values that don't correspond to any active file/KB/user, and deletes those specific points — it does not delete the shared collections themselves.

In both cases, the tool connects to Qdrant using the same client and credentials that Open WebUI uses (it imports the factory-created client). So yes, the embeddings are truly removed from Qdrant.

[Classic298]

Added export preview to csv feature