| Data | Location | Priority |
|---|---|---|
| Open WebUI database — config, accounts, conversation history, assistant (model) definitions, custom prompts | DATA_DIR\webui.db |
Critical |
| Uploaded documents | DATA_DIR\uploads |
Critical (unless originals kept elsewhere) |
| Vector DB | DATA_DIR\vector_db |
High (rebuildable by re-ingesting, but costly) |
| Knowledge register + test sets | your backup folder | High |
| This repo (scripts/docs/version register) | git remote | Covered |
Not backed up: model files (%USERPROFILE%\.ollama\models) — replaceable via
ollama pull; the model register (names + tags) is what you keep, not blobs.
scripts/backup.ps1 stops nothing (SQLite copy is safe when WebUI is idle — run
it after stop-ai.ps1 for a guaranteed-consistent copy). It zips DATA_DIR +
the version register to Backups\localai-backup-<date>.zip under your user
profile (change with -Destination). Weekly is a sensible cadence, plus always
before any update.
Backups contain confidential documents and chat history — store them on an encrypted/access-controlled location only (BitLocker-protected drive or encrypted external disk). Never a public cloud folder by default.
Restore: install same versions from the register → stop stack → replace
DATA_DIR contents from the zip → start → verify a chat, an assistant, and one
RAG query.
Never blind-update the whole environment at once.
Before any major update: 1) run backup; 2) record current working versions; 3) read the component’s release notes; 4) update one component at a time; 5) re-run a mini test: one chat + three RAG questions (incl. one refusal case); 6) roll back if degraded (pin previous version, restore data backup).
Disable/ignore auto-update mechanisms where possible: launch Open WebUI
version-pinned (uvx open-webui==X.Y.Z serve ...); update Ollama deliberately
via its installer, not whenever the tray nags.
| Component | Version | Date verified | Notes |
|---|---|---|---|
| Windows 11 build | |||
| Ollama | |||
| Open WebUI | pinned version | ||
| LLM default | qwen3:4b @ digest | ollama list shows digests |
|
| LLM advanced | |||
| Embedding model | bge-m3 @ digest | re-index required if changed | |
| Python | |||
| Intel Arc driver | |||
| Docker/WSL2 | only if used |