Goal: build the full stack (LM Studio + Open WebUI + ComfyUI) from scratch on Nobara (Fedora-based), x86-64. All components run locally - no cloud access.

1. Download the AppImage into ``/home/rootai/Downloads``.

2. Make it executable:

  ``chmod +x ~/Downloads/lm-studio-*.AppImage''
  

3. Nobara/Fedora compatibility (FUSE): if the AppImage does not start:

  ``./lm-studio-*.AppImage --appimage-extract``
  ``./lm-studio-*/AppRun''

4. Load the model: LM Studio → Models → download Qwen3 27B (GGUF) (~16 GB).

  Models and settings live in ``~/.lmstudio`` - the AppImage itself is replaceable, the data stays.

5. Link the lms CLI globally (symlink to the AppImage):

  ``sudo ln -sf /home/rootai/Downloads/lm-studio-*.AppImage /usr/local/bin/lms''

  Note: after every AppImage update with a new filename, recreate the symlink (see troubleshooting).

Test:

  ``lms ls''

  # start the server (LM Studio -> Developer -> Start Server, port 1234)
  lms load qwen/qwen3.8-27b

  # API test
  curl -s http://127.0.0.1:1234/v1/models

Expected: JSON list containing ``qwen3.8-27b``.

  # pull the image
  podman pull ghcr.io/open-webui/open-webui

  # start the container (name: open-webui, port 3000, own data volume)
  podman run -d --name open-webui -p 3000:8080 -v open-webui-data:/app/backend/data --restart unless-stopped ghcr.io/open-webui/open-webui

Setup in the browser (``http://<server>:3000``):

1. LLM connection: Settings → Connections → OpenAI API

• Base URL: ``http://host.docker.internal:1234/v1''
• API key: any value (LM Studio does not require one)
• Test: model ``qwen3.8-27b'' must appear

2. Embedding model: Settings → Models → set ``nomic-embed-text-v1.5'' as embedding model 3. Create user: ``artz'' with read access to the knowledge base 4. Important: disable the built-in tool Ask User - otherwise the model asks follow-up questions

Test:

  podman ps | grep open-webui
  curl -s http://127.0.0.1:3000

Start (background, log to ~/comfyui.log):

  nohup ~/comfyui-env/bin/python ~/comfyui/main.py --listen --port 8188 >> ~/comfyui.log 2>&1 &

Test:

  curl -s http://127.0.0.1:8188/system_stats

Connect in Open WebUI: Settings → Connections → ComfyUI

• URL: ``http://localhost:8188''
• Model: Z-Image

Log on problems: ``tail -f ~/comfyui.log''

Script: ``/home/rootai/start-ai.sh''

  #!/bin/bash
  # wait for system services and the LM Studio server
  sleep 15
  # automatically load the Qwen model into LM Studio
  lms load qwen/qwen3.8-27b
  # give it a few seconds to fully initialize
  sleep 10
  # restart Open WebUI
  podman restart open-webui

Start order matters: LM Studio → load model → Open WebUI.

  # 1. LM Studio API
  curl -s http://127.0.0.1:1234/v1/models

  # 2. Open WebUI reachable
  curl -sI http://127.0.0.1:3000 | head -1

  # 3. ComfyUI reachable
  curl -s http://127.0.0.1:8188/system_stats | head -c 100

1. Browser → ``http://<server>:3000'' → login ``artz'' 2. Ask a free question to ``qwen3.8-27b'' → answer without RAG citations (RAG = Part 2)

Goal: add the 5 billing documents as the knowledge base Zahnarztabrechnung in Open WebUI and create the model Zahn-Abrechnungs-Assistent.

RAG does not work without an embedding model.

1. Open WebUI → Settings → Models 2. Embedding model: ``nomic-embed-text-v1.5'' (built-in, no Ollama/LM Studio needed) 3. Verify: the models list shows the embedding model

Fallback if German retrieval is poor: install ``bge-m3'' via Ollama and set it as the embedding model.

Pre-check for each PDF:

• text can be selected/copied → OK
• images only → OCR problem (BEB)

Symptom without OCR: BEB citations missing or nonsense, other PDFs work fine.

Fix: 1. Open WebUI → Admin → Settings → Documents 2. turn ON the option PDF Extract Images (OCR) 3. re-upload / re-index the BEB file (delete the old document first, otherwise the stale empty index stays)

1. Open WebUI → Knowledge → Create, Name: ``Zahnarztabrechnung'' 2. Add documents: upload all 5 PDFs 3. wait until every document shows status processed (no errors) 4. Access: user ``artz'' → Read access on the knowledge base

Knowledge-base test (no model prompt yet):

• new chat session, attach the Zahnarztabrechnung KB
• test question: “What statutory co-payment (Festzuschuss) applies to a crown?”
• expected: answer with source citations from the FZ-Kompendium

1. Open WebUI → Settings → Models → Create 2. Name: ``Zahn-Abrechnungs-Assistent'' 3. Base model: ``qwen3.8-27b'' (LM Studio) 4. System prompt:

  You are an expert in German dental and dental-technician billing
  (GOÄ/GOZ, FZ-Kompendium, BEL/BEB, SGB V). Answer strictly on the basis
  of the documents provided. Give concrete position numbers, prices, the
  edition (e.g. BEB 2024, BEL 2025) and the source. If a topic is not
  covered in the documents, say so clearly. This is not legal advice.

5. Important: disable the built-in tool Ask User for this model - otherwise it asks follow-up questions (“Which tooth number?”) instead of answering from the documents

1. Browser → ``http://<server>:3000'' → login ``artz'' 2. Select model Zahn-Abrechnungs-Assistent 3. Sample test questions (German):

• “Which GOZ position applies for a composite filling on an incisor?” → expect GOZ 2010 with source
• “What does a full-ceramic crown cost per BEB?” → expect BEB position + price + edition 2024
• “What is the statutory co-payment for inlays?” → expect FZ-Kompendium 2025

4. Checklist per answer:

• source citations visible (document + section)
• concrete position number + price
• edition stated (BEB 2024 / BEL 2025 / FZ 2025)
• for missing info: a clear statement instead of a hallucination