prune-open-webui

KO | EN

gitlite — search

#javascript #python #react #ai #golang #android #deep-learning #linux #java #awesome #kubernetes #vue #awesome-list #mcp

prune-open-webui

(★ 134)

🧹 An automizable or interactive pruning tool for Open WebUI: clean up orphaned files, old chats, inactive users, stale vector data and more!

File Explorer

Download Latest Version (.zip)

- - bug_report.yml
  - config.yml
- PULL_REQUEST_TEMPLATE.md
.gitignore
COMPATIBILITY.md
LICENSE
prune.py
prune_cli_interactive.py
prune_core.py
prune_export.py
prune_imports.py
prune_models.py
prune_operations.py
README.md
requirements.txt
ruff.toml
run_prune.sh
standalone_prune.py
validate_structure.py

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# LEGACY

> [!IMPORTANT]
> # This Repository is Legacy Now and will not be maintained any further
> The approach used by the prune tool here had some downsides with Kubernetes setups and was not very handy to use in day-to-day scenarios since it always required CLI access.
> # Successor:
> **[The direct successor to this repository is the Prune Event Function found in my other repo here](https://github.com/Classic298/open-webui-plugins)**
> It is easy to install, has a UI, is automated, and strictly better than the version in this repository.
> Since maintaining both is not economical (and makes no sense when the replacement is strictly better) this repository is now LEGACY and will soon be archived in the future.
> You should no longer use this tool. As time passes, it will stop working with newer versions of Open WebUI.
> If you ever encounter any bugs with this tool, they will NOT be fixed. Only the [new Prune Event Function](https://github.com/Classic298/open-webui-plugins) will receive support (as much as I can).
> Again thank you to everyone who helped test this, submitted PRs here and helped with the continuous improvement. The development is not dead, it just shifted to the other repo into a new event function.
> Enjoy.

---

# Open WebUI Prune Tool

> [!IMPORTANT]
> **This is a community-built tool.** It is not an official part of the Open WebUI project. It is developed and maintained independently by community contributors on a best-effort basis. While we strive for correctness and safety, there are no guarantees of functionality, compatibility, or continued maintenance.

> [!CAUTION]
> **This tool performs irreversible, destructive operations on your database and file system.** Deleted data cannot be recovered. There is no undo. Always create a full backup or snapshot of your server, database, and data directory before running any pruning operations. Test on a staging environment before running against production. The authors and contributors of this tool accept **no liability** for data loss, corruption, or any other damage caused by its use, whether due to bugs, misconfiguration, or user error. **You use this tool entirely at your own risk.**

A standalone command-line tool for cleaning up your Open WebUI instance, reclaiming disk space, and maintaining optimal performance.

## Table of Contents

1. [Overview](#overview)
2. [Features](#features)
3. [Installation](#installation)
   - [Environment Variables](#environment-variables)
   - [Method 1: Docker (Recommended)](#method-1-docker-installation-recommended)
   - [Method 2: Systemd Service](#method-2-systemd-service-installation)
   - [Method 3: Pip](#method-3-pip-installation)
   - [Method 4: Git (Manual)](#method-4-git-installation-manual-install)
4. [Quick Start](#quick-start)
5. [Usage](#usage)
6. [Configuration Options](#configuration-options)
7. [Automation](#automation)
8. [Important Warnings](#important-warnings)
9. [Troubleshooting](#troubleshooting)
10. [Support](#support)

## Overview

The Prune Tool provides a safe and powerful way to clean up your Open WebUI instance by:
- Deleting old chats and conversations
- Removing inactive user accounts
- Cleaning orphaned data (files, tools, skills, automations, prompts, etc.)
- Pruning channels and channel messages (age-based, plus orphaned channels/messages)
- Removing old audio cache files
- Optimizing database performance

It runs independently of the web server and can be scheduled for automated maintenance.

## Features

✅ **Two Operation Modes**
- **Interactive Mode**: Beautiful terminal UI with step-by-step wizard
- **Non-Interactive Mode**: Command-line interface for automation

✅ **Complete Configurability**
- All configuration options are fully configurable
- Preview mode to see what will be deleted before execution
- Export detailed preview to CSV for auditing before execution
- Fine-grained control over what gets deleted

✅ **Database Support**
- SQLite (default)
- PostgreSQL
- Vector databases (community contributions for additional backends welcome!):
  - ChromaDB — full cleanup support
  - PGVector — full cleanup support
  - Milvus — full cleanup support
  - Qdrant — full cleanup support

✅ **Storage Backend Support**
- Local filesystem (`data/uploads`)
- Amazon S3 (and S3-compatible endpoints) — honors `S3_KEY_PREFIX`
- Google Cloud Storage
- Azure Blob Storage

Orphaned storage objects (files with no matching `File` row in the database) are detected and removed from whichever backend Open WebUI is configured to use.

✅ **Safety Features**
- File-based locking prevents concurrent operations
- Explicit `--execute` flag required — nothing is deleted without it
- Multiple confirmation prompts (interactive mode)
- Admin user protection
- Detailed logging of all operations

### Channel Pruning

Channels and channel messages can be cleaned up in three ways:

- **Age-based channel messages** (`--channel-message-max-age-days N`): delete messages older than N days from all channels while keeping the channels themselves. Pinned messages are preserved unless `--no-exempt-pinned-channel-messages` is given. Useful for keeping busy Support/Alerts channels fast and tidy.
- **Orphaned channels** (`--delete-orphaned-channels`, off by default): delete channels owned by a deleted user, along with their messages, reactions, members, file links and access grants.
- **Orphaned channel messages** (`--delete-orphaned-channel-messages`, on by default): delete messages whose channel no longer exists.

Files attached to pruned channel messages become unreferenced and are removed by the orphaned-file / upload / vector cleanup that always runs afterwards.

### Compatibility

See [COMPATIBILITY.md](COMPATIBILITY.md) for the full version compatibility matrix.

## Installation

### Prerequisites

- Open WebUI installation
- Python 3.11+
- Access to Open WebUI's Python environment and database
- All Open WebUI dependencies installed (including vector database libraries like chromadb)

### Folder Structure

The prune tool is designed to live as a `prune/` subfolder inside your Open WebUI installation directory. All commands in this README assume that structure:

```
open-webui/              # Your Open WebUI root
├── backend/
├── prune/               # This repository, cloned here
│   ├── prune.py         # Main entry point
│   ├── prune_cli_interactive.py
│   ├── prune_core.py
│   ├── prune_export.py
│   ├── prune_operations.py
│   ├── prune_imports.py
│   ├── prune_models.py
│   └── requirements.txt
└── ...
```

### Environment Variables

The prune tool reads the **same environment variables as Open WebUI** so it can locate the database, data directory, and vector store. The variables are identical across every install method; only *how* you supply them differs (see each method below).

| Variable | Required | Description |
|---|---|---|
| `WEBUI_SECRET_KEY` | Yes | Secret key for Open WebUI |
| `DATABASE_URL` | Yes | Database connection string (SQLite or PostgreSQL) |
| `DATA_DIR` | No | Data directory path (default: `/app/backend/data`) |
| `CACHE_DIR` | No | Cache directory path (audio cache and the lock file) |
| `VECTOR_DB` | No | Vector database type if using RAG (e.g. `chroma`, `pgvector`) |

**SQLite (default):**
```bash
DATABASE_URL=sqlite:////app/backend/data/webui.db
```

**PostgreSQL:**
```bash
DATABASE_URL=postgresql://username:password@host:port/database
VECTOR_DB=pgvector
```

> **PostgreSQL passwords with special characters** must be URL-encoded: `@` becomes `%40`, `:` becomes `%3A`, `/` becomes `%2F`, `?` becomes `%3F`, `#` becomes `%23`. So `password@123` becomes `password%40123`.

### Providing the variables

The prune tool reads variables from the current environment and, if present, from a `.env` file in the working directory. Use whichever method matches your setup:

- **Docker:** set them in your `docker-compose.yml` `environment:` block (or with `-e` flags on `docker run`). A correctly configured Open WebUI container already has them, and `docker exec` inherits them automatically, so no extra setup is needed.
- **`.env` file (best for repeated, native, or systemd use):** create a `.env` in the Open WebUI directory and it is loaded automatically.
  ```bash
  cat > /opt/openwebui/.env <<'EOF'
  DATABASE_URL=postgresql://user:password@localhost:5432/openwebui
  VECTOR_DB=pgvector
  DATA_DIR=/var/lib/openwebui
  CACHE_DIR=/var/lib/openwebui/cache
  EOF
  ```
- **Systemd service:** a unit's variables are not exported to your shell, so source them before running the tool:
  ```bash
  set -a
  source <(systemctl show openwebui.service -p Environment | sed 's/^Environment=//; s/ /\n/g')
  set +a
  ```

> **These errors mean the variables are not set:** "Required environment variable not found", "unable to open database file", or "Failed to connect to database". Provide them with one of the methods above.

---

## Method 1: Docker Installation (Recommended)

Running inside the Docker container is the **recommended approach** because all dependencies (including vector database libraries like chromadb) are already installed.

**Step 1: Download the prune tool** (one-time setup)
```bash
git clone https://github.com/Classic298/prune-open-webui.git prune
```

**⚠️ The trailing `prune` in the clone command is required** — it ensures the repository is cloned into a folder named `prune/` instead of `prune-open-webui/`.

**Step 2: Make the prune tool available in the container** (one-time setup)

**A volume mount is the recommended approach.** Mounting the `prune/` folder into the container (rather than copying it in with `docker cp`) keeps the scripts in sync with your local copy, so you never have to re-copy after pulling updates. Add the mount to your `docker-compose.yml`:

```yaml
services:
  open-webui:
    image: ghcr.io/open-webui/open-webui:main
    volumes:
      - open-webui:/app/backend/data
      - ./prune:/app/prune          # Mount the prune tool
```

Then recreate the container so the mount takes effect:
```bash
docker compose up -d
```

> **Alternative (one-off copy):** If you cannot edit your compose file, copy the folder in instead, but you will have to re-copy after every update.
> ```bash
> # Find your Open WebUI container name
> docker ps | grep open-webui
>
> # Copy the prune folder into the container
> docker cp prune <container-name>:/app/
> ```

**Step 3: Run the prune script**

**Option A: Run interactively inside container (Recommended)**
```bash
docker exec -it <container-name> bash
cd /app
python prune/prune.py
```

**Option B: Direct execution from host**
```bash
docker exec <container-name> python /app/prune/prune.py --days 90 --execute
```

**Option C: Preview mode from host**
```bash
docker exec <container-name> python /app/prune/prune.py --days 90 --dry-run
```

**Important Notes:**
- **Always execute inside the container** to ensure all dependencies are available
- Environment variables are automatically inherited from the container
- All vector database dependencies (chromadb, etc.) are pre-installed in the container
- Data persists in your Docker volumes

> A correctly configured Open WebUI container already has the required variables and `docker exec` inherits them, so no extra setup is needed. See [Environment Variables](#environment-variables).

---

## Method 2: Systemd Service Installation

If Open WebUI runs as a systemd service, its environment variables (like `DATABASE_URL`) are **only available to the service**, not to your terminal session.

Install the prune tool itself as shown in [Method 3: Pip](#method-3-pip-installation) or [Method 4: Git](#method-4-git-installation-manual-install). Because the service's variables are not in your shell, you must provide them before running the tool, for example with a `.env` file or by sourcing them from the unit.

See [Providing the variables](#providing-the-variables) for the `.env` file and systemd-sourcing approaches, and [Environment Variables](#environment-variables) for the full list.

---

## Method 3: Pip Installation

**Important:** Ensure all Open WebUI dependencies are installed, including vector database libraries (chromadb, etc.). The prune script requires the same dependencies as the main Open WebUI application.

```bash
# Activate environment where open-webui is installed
source venv/bin/activate

# Ensure all dependencies of Open WebUI are installed
pip install -r backend/requirements.txt
# Also install all dependencies of the prune script
pip install -r prune/requirements.txt

# Find installation location
pip show open-webui | grep Location

# Run from that location
cd <location>

# Set the required environment variables first (see Environment Variables above)
python prune/prune.py
```

---

## Method 4: Git Installation (Manual Install)

**One-time setup:**
```bash
cd ~/path/to/open-webui        # Navigate to your Open WebUI directory
git clone https://github.com/Classic298/prune-open-webui.git prune
source venv/bin/activate        # Activate your Python virtual environment
pip install -r prune/requirements.txt  # Install prune dependencies
pip install rich                # For interactive mode (optional)
```

**⚠️ The trailing `prune` in the clone command is required** — it ensures the repository is cloned into a folder named `prune/` instead of `prune-open-webui/`.

**Ready to use:**
```bash
cd ~/path/to/open-webui        # Must run from repo root
source venv/bin/activate
python prune/prune.py          # Launch interactive mode
```

## Quick Start

### Interactive Mode

```bash
python prune/prune.py
```

Features beautiful terminal UI with:
- Step-by-step configuration wizard
- Visual preview of what will be deleted
- Multiple safety confirmations
- Progress bars and status updates

### Non-Interactive Mode

```bash
# Preview what would be deleted (safe, no changes)
python prune/prune.py --days 90 --dry-run

# Delete chats older than 90 days
python prune/prune.py --days 90 --execute

# Full cleanup with optimization
python prune/prune.py \
  --days 90 \
  --delete-inactive-users-days 180 \
  --audio-cache-max-age-days 30 \
  --run-vacuum \
  --execute
```

### Export Preview to CSV

Before running a destructive operation, export a detailed list of every item that would be deleted to a CSV file for auditing:

```bash
# Export preview to CSV (non-interactive)
python prune/prune.py --days 90 --dry-run --export-preview preview.csv
```

The interactive mode also offers export after each preview — with size estimation and a progress bar.

The CSV contains one row per item with columns: `category`, `id`, `name`, `owner_id`, `size_bytes`, `reason`. The `size_bytes` column is populated for physical files (uploads and audio cache); other categories leave it empty.

## Usage

### Basic Patterns

**Preview Mode / --dry-run (Safe):**
```bash
python prune/prune.py --days 60 --dry-run
```

**Conservative Cleanup:**
```bash
python prune/prune.py \
  --days 180 \
  --exempt-archived-chats \
  --exempt-pinned-chats \
  --exempt-chats-in-folders \
  --execute
```

**Orphaned Data Only:**
```bash
python prune/prune.py \
  --delete-orphaned-chats \
  --delete-orphaned-knowledge-bases \
  --execute
```

**Inactive Users:**
```bash
python prune/prune.py \
  --delete-inactive-users-days 180 \
  --exempt-admin-users \
  --exempt-pending-users \
  --execute
```

**Retention Policy — expire old knowledge bases (DESTRUCTIVE):**
```bash
# Deletes KBs created more than 365 days ago, even if their owner is active
# and the KB is still in use. Always preview with --dry-run first.
python prune/prune.py \
  --delete-knowledge-bases-older-than-days 365 \
  --dry-run
```

## Configuration Options

The deletion flags fall into three categories, ordered from safest to most destructive by *what they target*:

1. **Orphaned Data Cleanup** — removes data whose owner no longer exists. This is the tool's core, safe purpose: it only touches data nobody can reach anymore.
2. **Age-Based Deletion** — removes data by age with a bounded blast radius (a user's own old chats, inactive accounts, derived audio cache). Time-based, but self-contained.
3. **Retention Policy** — removes **live, owned, in-use** data purely by age. Destructive and opt-in: it deletes things an active user still relies on.

A fourth group (**Execution & Output**) controls preview, execution, optimization, and logging.

### 1. Orphaned Data Cleanup (owner no longer exists)

| Option | Type | Default | Negate with | Description |
|--------|------|---------|-------------|-------------|
| `--delete-orphaned-chats` | flag | True | `--no-delete-orphaned-chats` | Clean orphaned chats |
| `--delete-orphaned-knowledge-bases` | flag | True | `--no-delete-orphaned-knowledge-bases` | Clean orphaned KBs |
| `--delete-orphaned-kb-metadata` | flag | True | `--no-delete-orphaned-kb-metadata` | Clean KB search-metadata embeddings whose KB no longer exists |
| `--delete-orphaned-memories` | flag | True | `--no-delete-orphaned-memories` | Clean memories an active user deleted that the vector store kept |
| `--delete-orphaned-prompts` | flag | True | `--no-delete-orphaned-prompts` | Clean orphaned prompts |
| `--delete-orphaned-models` | flag | True | `--no-delete-orphaned-models` | Clean orphaned models |
| `--delete-orphaned-notes` | flag | True | `--no-delete-orphaned-notes` | Clean orphaned notes |
| `--delete-orphaned-folders` | flag | True | `--no-delete-orphaned-folders` | Clean orphaned folders |
| `--delete-orphaned-chat-messages` | flag | True | `--no-delete-orphaned-chat-messages` | Clean orphaned chat_message rows |
| `--delete-orphaned-automations` | flag | True | `--no-delete-orphaned-automations` | Clean orphaned automations and automation runs |
| `--delete-orphaned-tools` | flag | False | — | Clean orphaned tools |
| `--delete-orphaned-functions` | flag | False | — | Clean orphaned functions |
| `--delete-orphaned-skills` | flag | False | — | Clean orphaned skills |
| `--delete-orphaned-channels` | flag | False | — | Delete channels owned by deleted users (with their messages/files) |
| `--delete-orphaned-channel-messages` | flag | True | `--no-delete-orphaned-channel-messages` | Delete channel messages whose channel no longer exists |
| `--orphan-file-grace-hours N` | int | 24 | — | Never treat files younger than N hours as orphaned (`0` disables the grace window) |

**How orphaned files are determined:** a file is deleted only when **nothing references it** — no knowledge base, chat, folder, channel message or model. Who uploaded it plays no role: a file uploaded by a since-deleted user into another user's live knowledge base is still in use and is kept. Files younger than the grace window are always kept, because uploading a file and referencing it (sending the message, adding it to a KB) are two separate requests in Open WebUI — without the grace window, a prune running in that gap could delete an upload the user is still attaching.

### 2. Age-Based Deletion (bounded blast radius)

| Option | Type | Default | Negate with | Description |
|--------|------|---------|-------------|-------------|
| `--days N` | int | None | — | Delete chats older than N days (by last-updated) |
| `--exempt-archived-chats` | flag | True | `--no-exempt-archived-chats` | Keep archived chats even if old |
| `--exempt-pinned-chats` | flag | False | — | Keep pinned chats even if old |
| `--exempt-chats-in-folders` | flag | False | — | Keep chats in folders even if old |
| `--delete-inactive-users-days N` | int | None | — | Delete users inactive N+ days (cascades all their data) |
| `--exempt-admin-users` | flag | True | `--no-exempt-admin-users` | Never delete admins (RECOMMENDED) |
| `--exempt-pending-users` | flag | True | `--no-exempt-pending-users` | Never delete pending users |
| `--audio-cache-max-age-days N` | int | None | — | Clean audio cache files older than N days |
| `--channel-message-max-age-days N` | int | None | — | Delete channel messages older than N days (channels are kept) |
| `--no-exempt-pinned-channel-messages` | flag | (pinned kept) | — | Also delete pinned messages during age-based channel pruning |

### 3. Retention Policy (DESTRUCTIVE — deletes live, owned, in-use data)

**These flags delete data that is still owned by an active user and may be in active use.** This is a data-retention/expiry policy, not orphan cleanup. There is no undo. Only enable it if your policy genuinely requires expiring live content by age.

| Option | Type | Default | Negate with | Description |
|--------|------|---------|-------------|-------------|
| `--delete-knowledge-bases-older-than-days N` | int | None | — | Delete knowledge bases older than N days, **even if the owner exists and the KB is in use** |
| `--knowledge-bases-age-field` | `created_at`/`updated_at` | `created_at` | — | Timestamp used to measure KB age |

When a knowledge base is deleted this way, the tool mirrors Open WebUI's own KB deletion: it removes the KB's vector collection, deletes the KB record, removes the KB's search-metadata embedding from the shared `knowledge-bases` collection, and **de-references the KB from any model that points at it** (so referencing models keep working). The KB's now-unreferenced files, uploads, and per-file vector collections are reclaimed by the orphan sweep that runs immediately afterwards.

### Execution & Output

| Option | Type | Default | Negate with | Description |
|--------|------|---------|-------------|-------------|
| `--run-vacuum` | flag | False | — | Run database optimization on main and vector DBs (locks DB!) |
| `--dry-run` | flag | True | — | Preview only (default) |
| `--execute` | flag | — | — | Actually perform deletions |
| `--verbose, -v` | flag | — | — | Enable debug logging |
| `--quiet, -q` | flag | — | — | Suppress non-error output |
| `--export-preview PATH` | str | None | — | Export detailed preview to CSV (requires `--dry-run`) |

### Using Flags

Every flag that defaults to **True** can be disabled with the `--no-` prefix. This lets you selectively run only specific cleanup categories. Flags that default to **False** (e.g. `--delete-orphaned-tools`, `--delete-orphaned-functions`, `--delete-orphaned-skills`) do not have a `--no-` variant — they are already off unless you explicitly enable them.

```bash
# Disable all default-on orphan cleanup except chats — process one category at a time
python prune/prune.py \
  --no-delete-orphaned-prompts \
  --no-delete-orphaned-knowledge-bases \
  --no-delete-orphaned-models \
  --no-delete-orphaned-notes \
  --no-delete-orphaned-folders \
  --no-delete-orphaned-chat-messages \
  --no-delete-orphaned-automations \
  --execute

# Delete archived chats too (overrides default exemption)
python prune/prune.py --days 60 --no-exempt-archived-chats --execute

# Include admins in inactive user deletion (NOT RECOMMENDED!)
python prune/prune.py --delete-inactive-users-days 180 --no-exempt-admin-users --execute
```

## Automation

### Cron Job Example (Native / Systemd / pip Installs)

```bash
# Edit crontab
crontab -e

# Weekly cleanup every Sunday at 2 AM
0 2 * * 0 /path/to/run_prune.sh --days 90 --audio-cache-max-age-days 30 --execute >> /var/log/openwebui-prune.log 2>&1

# Monthly full cleanup with VACUUM (first Sunday at 3 AM)
0 3 1-7 * 0 /path/to/run_prune.sh --days 60 --delete-inactive-users-days 180 --run-vacuum --execute >> /var/log/openwebui-prune-monthly.log 2>&1
```

### Docker Automation (Host-Side Cron)

Standard Open WebUI Docker containers **do not include cron**. The recommended approach is to schedule the cron job on your **host machine** and use `docker exec` to run the headless prune script inside the running container.

This works because `docker exec` runs a command inside an already-running container, inheriting its environment variables, Python dependencies, and database access — everything the prune script needs.

**Step 1: Mount the prune folder into your container**

Make sure your `docker-compose.yml` mounts the prune directory:

```yaml
services:
  open-webui:
    image: ghcr.io/open-webui/open-webui:main
    volumes:
      - open-webui:/app/backend/data
      - ./prune:/app/prune          # Mount the prune tool
    environment:
      - WEBUI_SECRET_KEY=${WEBUI_SECRET_KEY}
      - DATABASE_URL=sqlite:////app/backend/data/webui.db
    ports:
      - "3000:8080"
```

> **Tip:** Mounting the folder (instead of `docker cp`) keeps the prune scripts in sync with your local copy and avoids having to re-copy after updates.

**Step 2: Verify it works manually**

```bash
# Preview mode first (safe, no changes)
docker exec open-webui python /app/prune/prune.py --days 90 --dry-run

# If the preview looks good, run with --execute
docker exec open-webui python /app/prune/prune.py --days 90 --execute
```

**Step 3: Add a cron job on the host**

```bash
# Edit the HOST machine's crontab (not inside the container)
crontab -e

# Weekly cleanup every Sunday at 2 AM
0 2 * * 0 docker exec open-webui python /app/prune/prune.py --days 90 --audio-cache-max-age-days 30 --execute >> /var/log/openwebui-prune.log 2>&1

# Monthly full cleanup with VACUUM (first Sunday at 3 AM)
0 3 1-7 * 0 docker exec open-webui python /app/prune/prune.py --days 60 --delete-inactive-users-days 180 --run-vacuum --execute >> /var/log/openwebui-prune-monthly.log 2>&1
```

> **Note:** Replace `open-webui` with your actual container name. Find it with `docker ps | grep open-webui`.

**Optional: Host-side wrapper script**

For cleaner cron entries, create a small wrapper script on the host:

```bash
#!/bin/bash
# /usr/local/bin/openwebui-prune-docker
# Wrapper to run prune inside the Open WebUI container from the host

CONTAINER_NAME="${OPENWEBUI_CONTAINER:-open-webui}"

# Check if the container is running
if ! docker inspect -f '{{.State.Running}}' "$CONTAINER_NAME" 2>/dev/null | grep -q true; then
    echo "ERROR: Container '$CONTAINER_NAME' is not running" >&2
    exit 1
fi

# Run prune script inside the container, forwarding all arguments
docker exec "$CONTAINER_NAME" python /app/prune/prune.py "$@"
```

Make it executable and use it in cron:

```bash
chmod +x /usr/local/bin/openwebui-prune-docker

# Now cron entries are cleaner:
0 2 * * 0 /usr/local/bin/openwebui-prune-docker --days 90 --execute >> /var/log/openwebui-prune.log 2>&1
```

### Best Practices for Automation

1. **Always test manually first** with `--dry-run`
2. **Schedule during low-usage hours** (2-4 AM)
3. **Create backups before pruning**
4. **Monitor logs regularly** for errors
5. **Start conservative**, adjust gradually
6. **Never use VACUUM** during active hours

## Important Warnings

### ⚠️ User Deletion Cascades

When you delete a user, **ALL their data is deleted**:
- Chats and messages
- Files and uploads
- Custom tools, functions, and skills
- Knowledge bases and embeddings
- Automations and automation run history
- Prompts, models, notes, folders
- Everything they created

**Recommendations:**
- Use long periods (180+ days minimum)
- **Always** exempt admin users
- Test on staging first

### ⚠️ VACUUM Locks Database

When `--run-vacuum` is enabled:
- **Main and vector databases may be locked** during operation
- **All users will experience errors**
- Can take **5-30+ minutes** (or longer)
- **Only use during scheduled maintenance windows**

### ⚠️ Preview First

**Always run with `--dry-run` first** to see what will be deleted:

```bash
# Safe preview
python prune/prune.py --days 60 --dry-run

# Review output, then execute
python prune/prune.py --days 60 --execute
```

## Troubleshooting

### Error: "Failed to import Open WebUI modules"

**Solution:** Must run from Open WebUI root directory

```bash
cd /path/to/open-webui  # Go to repo root, NOT prune/
python prune/prune.py
```

### Error: "Failed to connect to database" or "unable to open database file"

This is the **most common issue** for systemd and pip installations.

**Problem:** The script tries to connect to SQLite but you're using PostgreSQL, **OR** environment variables from your systemd service file are not available in your terminal session.

**Symptoms:**
```
sqlite3.OperationalError: unable to open database file
Failed to connect to database
```

**Root Cause:** Environment variables like `DATABASE_URL` are set in your systemd service file but are **NOT** exported to your shell session.

**Solution:** See [Method 2: Systemd Service Installation](#method-2-systemd-service-installation) for detailed instructions on configuring environment variables for your setup.

### Error: "Operation already in progress"

**Solution:** Remove stale lock file (if >2 hours old)

```bash
# Check lock file age
ls -la cache/.prune.lock

# Remove if stale
rm cache/.prune.lock
```

### Error: "No module named 'rich'"

**Solution:** Install optional dependency for interactive mode

```bash
pip install rich
```

### Error: "NoneType object has no attribute 'lower'" (VECTOR_DB issue)

**Problem:** Vector database dependencies are not installed.

**Solution:** Run inside Docker container (recommended), or install all Open WebUI dependencies:
```bash
pip install -r backend/requirements.txt
```

For non-Docker installations, ensure `VECTOR_DB` matches your configuration and the corresponding library is installed (chromadb, pgvector, etc.).

### Performance Issues

If operations are very slow:
- Check database size: `du -h data/webui.db`
- Run during off-hours
- Consider breaking into smaller operations
- Monitor with `htop` during execution

## Technical Details

### What Gets Deleted

**By Age:**
- Chats older than specified days (based on `updated_at`, ensuring only long-unused chats are deleted)
- Users inactive for specified days (based on `last_active_at`)
- Audio cache files (based on file `mtime`)

**By Retention Policy (destructive, opt-in):**
- Knowledge bases older than specified days (by `created_at` or `updated_at`), **even when the owner is active and the KB is in use**. Referencing models are de-referenced; the KB's vector collection, search-metadata embedding, record, and now-unreferenced files are removed.

**Orphaned Data:**
- Chats/tools/skills/automations/prompts/etc. from deleted users
- Automation runs from deleted automations
- Chat messages (analytics metadata) from deleted chats — see note below
- Files referenced by nothing (no chat, knowledge base, folder, channel message, or model — regardless of who uploaded them) and older than the upload grace window (`--orphan-file-grace-hours`, default 24h). Storage objects are deleted from the configured backend — local, S3, GCS, or Azure
- Vector collections for deleted files/KBs
- Knowledge base search-metadata embeddings (in the shared `knowledge-bases` collection) whose knowledge base no longer exists. Open WebUI stores one small embedding per KB (its name and description) for semantic search across knowledge bases; this removes the leftover entry when a KB was deleted outside the tool or by an older version. Toggle with `--delete-orphaned-kb-metadata` / `--no-delete-orphaned-kb-metadata`.
- Memories (in each active user's `user-memory-{id}` collection) whose memory row no longer exists. When a user deletes an individual memory, the vector point can be left behind and keep getting injected into chat context via RAG; this reconciles each active user's memory collection against the `memory` table and removes the orphans. Toggle with `--delete-orphaned-memories` / `--no-delete-orphaned-memories`.
- Storage objects (local uploads, S3/GCS/Azure blobs) whose path does not match any `File` row in the database

> [!NOTE]
> SQLite does not enforce `ON DELETE CASCADE` by default. This means deleting a chat may leave behind orphaned rows in the `chat_message` table (used by the Analytics feature). The prune tool detects and removes these automatically. PostgreSQL users are not affected.

**Preserved:**
- Active user accounts
- Referenced files
- Valid vector collections
- Recent data (within retention period)
- Exempted categories (archived, pinned, folders, admins)
- Knowledge bases, unless age-based retention (`--delete-knowledge-bases-older-than-days`) is explicitly enabled

### Vector Database Support

- **ChromaDB**: Full cleanup — SQLite metadata, directories, and FTS indices
- **PGVector**: Full cleanup — PostgreSQL tables and embeddings
- **Milvus**: Full cleanup — standard and multitenancy modes
- **Qdrant**: Full cleanup — standard and multitenancy modes
- **Others**: Safe no-op (does nothing)
- **Adding support for a new backend**: Subclass `VectorDatabaseCleaner` in `prune_core.py`, implement the abstract methods, and register it in the `get_vector_database_cleaner` factory. Community contributions are welcome!

### Storage Backend Support

The prune tool uses Open WebUI's existing `Storage` abstraction and its already-configured clients — no extra credentials or configuration are required beyond what Open WebUI itself needs. Whichever backend `STORAGE_PROVIDER` is set to, the prune tool will:

1. List every object in that backend
2. Compare each object's path against the set of active `File.path` values in the database
3. Delete objects whose path does not match any active `File` row

| `STORAGE_PROVIDER` | Scan target | Notes |
|---|---|---|
| `local` (default) | `data/uploads/` | Walks the directory with `Path.iterdir()` |
| `s3` | S3 bucket | Honors `S3_KEY_PREFIX` — objects outside the prefix are never touched |
| `gcs` | GCS bucket | Scans the entire bucket |
| `azure` | Azure container | Scans the entire container |
| unknown | — | Orphan storage scan is skipped with a warning |

> [!WARNING]
> For GCS and Azure, Open WebUI does not expose a key-prefix option. The prune tool scans the **entire** configured bucket/container and will flag any object with no matching `File` row as orphaned. If you share the bucket with other applications, either give Open WebUI a dedicated bucket or run with `--dry-run` / `--export-preview` first to review what would be deleted.

### Safety Features

- **File-based locking** prevents concurrent runs
- **Explicit execution** requires `--execute` flag — nothing is deleted without it
- **Admin protection** enabled by default
- **Stale lock detection** automatic cleanup
- **Error handling** per-item with graceful degradation
- **Comprehensive logging** all operations tracked

## Support

- Review this README and the [Troubleshooting](#troubleshooting) section
- Check logs: `tail -f /var/log/openwebui-prune.log`
- Test in a staging environment first
- **Questions, feature requests, and general discussion**: [Discussions](https://github.com/Classic298/prune-open-webui/discussions)
- **Reproducible bugs only**: [Open an issue](https://github.com/Classic298/prune-open-webui/issues)

---

> [!CAUTION]
> **With great power comes great responsibility.** Always preview first, create backups, and start with conservative settings. This tool is not responsible for your lack of backups.

prune-open-webui

File Explorer

Today's picks

Was this content helpful?

Recent Feedback

prune-open-webui

File Explorer

Today's picks

Was this content helpful?

Recent Feedback

Download README