AI-Knowledge-Assistant

Open WebUI Deployment — Method Decision, Install, Verification

Deployment method decision (16 GB machine)

Method	RAM overhead	Verdict
Native Python (uv/pip)	~0.5–1 GB (WebUI process only)	Chosen default
Docker Desktop + WSL2	+1.5–2.5 GB idle (WSL2 VM, vmmem, Docker daemon)	Fallback only; if used, cap WSL2 RAM via .wslconfig (memory=3GB) and stop Docker when the stack is down

On a 16 GB laptop the Docker overhead is roughly the cost of an entire second 4B model. Native wins unless the audit shows Python cannot be installed.

Install (native, localhost-only)

scripts/04-install-openwebui.ps1 automates this. Manual equivalent:

# Requires Python 3.11 or 3.12 (audit checks this). Then:
pip install uv
# Data lives in %USERPROFILE%\open-webui\data (set DATA_DIR to control it)
$env:DATA_DIR = "$env:USERPROFILE\open-webui\data"
uvx --python 3.12 open-webui@latest serve --host 127.0.0.1 --port 8080

--host 127.0.0.1 is mandatory — the default binds 0.0.0.0 (all interfaces, LAN-visible). Ollama already defaults to 127.0.0.1:11434; leave it that way.

First launch: open http://localhost:8080, create the admin account immediately (first signup becomes admin) with a strong password, then in Admin Panel → Settings disable open signups (Default User Role: pending or disable Enable New Sign Ups).

Connect to Ollama: Admin Panel → Settings → Connections → Ollama API = http://127.0.0.1:11434 (usually auto-detected).

Verification checklist (Gate 4)

• http://localhost:8080 loads; admin account created; signups closed
• Ollama connection green; qwen3:4b appears in the model selector
• New chat streams a response; conversation history persists after restart
• File upload works (any small PDF attaches to a chat)
• Workspace → Knowledge can create a collection (RAG plumbing present)
• Arabic prompt renders correctly RTL in both the input box and the answer
• scripts/05-privacy-check.ps1 shows 8080 and 11434 listening on 127.0.0.1 only

Service notes

• Stop with Ctrl+C in its console, or scripts/stop-ai.ps1.
• scripts/start-ai.ps1 starts Ollama (if needed) and Open WebUI in a background window. Deliberately not installed as an auto-start Windows service: heavy AI services must not load at every boot on this machine (Section 29 of the master prompt). Revisit only if benchmarking proves the impact acceptable.
• Updates: uvx open-webui@latest fetches the newest release each launch. For a stable, version-pinned setup, pin (open-webui==0.6.x) per the update policy in docs/11-backup-updates.md and upgrade deliberately, not implicitly.