AI-Stack

6 min read

AI-Stack

A self-hosted AI stack optimised for Intel Arc iGPU on Linux, built around Ollama + Open WebUI with automated model routing, system diagnostics tools, and a systemd-managed Docker Compose stack.

Built and documented through real-world homelab experience on Intel Arc hardware.

What’s included

ComponentPurpose
Ollama (ava-agentone/ollama-intel)LLM inference with Intel Arc iGPU acceleration via OneAPI/SYCL
Open WebUIChat interface with tool calling, pipelines, and terminal access
PipelinesServer-side plugin system for model routing and workflow automation
Open TerminalBrowser-based terminal inside Open WebUI (with sudo support)
Smart Model RouterAuto-routes queries to the best model based on content
System DiagnosticsTool for querying Ollama health, models, and VRAM across multiple machines

Hardware requirements

ComponentMinimumRecommended
CPUIntel Core Ultra (Meteor Lake)Intel Core Ultra 9 185H
RAM16 GB32 GB
GPUIntel Arc iGPUIntel Arc iGPU (any Meteor/Arrow Lake)
Storage50 GB free100 GB+ free (models are large)
OSUbuntu 22.04Ubuntu 24.04

Note: This stack uses ghcr.io/ava-agentone/ollama-intel which replaced the archived intelanalytics/ipex-llm-inference-cpp-xpu image (archived January 28, 2026).

Quick start

The source code for this stack is maintained in the [growlf/ai-stack] repository. This repository contains the production-ready Docker Compose files, systemd unit definitions, and automation scripts used to deploy the NetYeti AI Lab.

# 1. Clone the repo
git clone https://github.com/growlf/ai-stack.git
cd ai-stack
 
# 2. Configure
cp .env.example .env
nano .env   # set your username, paths, and API keys
 
# 3. Install
chmod +x install.sh scripts/check-arc-gpu.sh
./install.sh
 
# 4. Open
# http://localhost:3000

Then follow post-install.md for the Open WebUI configuration steps.

Project structure

phoenix-ai-stack/
├── install.sh                  # Main installer
├── docker-compose.yml          # Full stack definition
├── .env.example                # All configurable values
├── systemd/
│   └── ai-stack.service        # Systemd unit (auto-start on boot)
├── scripts/
│   └── check-arc-gpu.sh        # GPU pre-flight (detects card0/card1 drift)
├── pipelines/
│   └── smart_model_router.py   # Auto-routes queries to best model
├── tools/
│   └── system_diagnostics.py   # Multi-instance Ollama health + model queries
└── docs/
    ├── post-install.md          # Open WebUI configuration checklist
    ├── model-guide.md           # Model recommendations and routing table
    └── troubleshooting.md       # Common issues and fixes

Model stack

ModelUse case
qwen2.5:14bTool calling, diagnostics, sysadmin (default)
qwen2.5-coder:14bScripts, configs, code
deepseek-r1:14bComplex reasoning, root cause analysis
gemma3:12bLog analysis, summaries, documentation
nomic-embed-textEmbeddings / RAG

See model-guide.md for details.

Known Intel Arc quirks

  • The DRI card node (/dev/dri/card0 vs card1) can drift between reboots on Meteor Lake. The check-arc-gpu.sh pre-flight script detects and corrects this automatically.
  • Intel iGPU uses shared system RAM — runner.vram="0 B" in Ollama logs is expected and normal.
  • Use OLLAMA_KEEP_ALIVE=-1 to keep models resident in memory between requests.
  • renderD128 is the compute node and is stable; only the cardN display node drifts.

Multi-machine setup

The System Diagnostics tool supports querying multiple Ollama instances across your LAN. Edit OLLAMA_INSTANCES in tools/system_diagnostics.py:

OLLAMA_INSTANCES = {
    "local":   "http://ollama-arc:11434",   # this machine
    "remote1": "http://10.0.0.X:11434",     # remote machine on your LAN
}

Updating the stack

cd /path/to/phoenix-ai-stack
 
# Pull latest images
docker compose pull
 
# Restart with new images
sudo systemctl restart ai-stack.service

Critique

Strengths of the Configuration

  • Intel Arc Optimization: You are using the ava-agentone/ollama-intel image, which is the current recommended path following the archiving of the official Intel IPEX image.
  • GPU Drift Protection: The inclusion of check-arc-gpu.sh in the systemd unit’s ExecStartPre is a proactive way to handle the known issue where Arc iGPU card nodes (e.g., card0 vs card1) can change after a reboot.
  • Intelligent Routing: Your Smart Model Router pipeline efficiently matches specific tasks to the best-suited 14b models (e.g., qwen2.5-coder for scripts, deepseek-r1 for reasoning) .

Areas for Improvement and Potential Issues

  • Open Terminal Permissions: Your open-terminal service has TERMINAL_SUDO_ENABLED=true. While powerful, ensure the STACK_USER defined in your .env has a strong password, as this service provides root-level access through the WebUI.
  • Volume Consistency: The open-webui volume is marked as external: true in the compose file. This means the stack will fail to start if you haven’t manually run docker volume create open-webui first. Your install.sh handles this, but manual docker compose up calls might fail.
  • Resource Constraints: You have set shm_size: '16gb' and mem_limit: 32g for Ollama. This is excellent for 14b models, but keep in mind that since the iGPU shares system RAM, you may experience “paging out” if you try to load more than one large model at a time.
  • Remote System Connectivity: To diagnose a remote system, you’ll need to manually update the OLLAMA_INSTANCES dictionary in tools/system_diagnostics.py. There is currently no way to do this via environment variables alone.

Quick Diagnostic Checklist

  1. Connectivity: If Open WebUI can’t see Ollama, verify they are both on the ai-net bridge network.
  2. GPU Utilization: If inference is slow, run docker logs ollama-arc 2>&1 | grep -i "oneapi" to ensure the Intel GPU is actually being engaged.
  3. Pipelines: If your custom routing isn’t working, you may need to clear the Python bytecode cache by running docker exec pipelines rm -rf /app/pipelines/__pycache__.

Licence

MIT — use freely, contributions welcome.

Built with ☕ and stubbornness.

Contents:

FolderDocuments

Recent Changes / Additions

| File | Created | Updated | | --- | --- | | Recent Activity | 12:26 AM - April 16, 2026 | 12:26 AM - April 16, 2026 | | Folder | 12:26 AM - April 16, 2026 | 12:26 AM - April 16, 2026 | | Project | 2:06 PM - April 14, 2026 | 2:06 PM - April 14, 2026 | | AAR | 2:06 PM - April 14, 2026 | 2:06 PM - April 14, 2026 | | Post-Implementation Review | 2:06 PM - April 14, 2026 | 2:06 PM - April 14, 2026 | | dailynote | 2:06 PM - April 14, 2026 | 2:06 PM - April 14, 2026 | | Discovery Time | 2:06 PM - April 14, 2026 | 2:06 PM - April 14, 2026 | | Network | 2:06 PM - April 14, 2026 | 2:06 PM - April 14, 2026 | | Service | 2:06 PM - April 14, 2026 | 2:06 PM - April 14, 2026 | | Person | 2:06 PM - April 14, 2026 | 2:06 PM - April 14, 2026 | | Daily | 2:06 PM - April 14, 2026 | 2:06 PM - April 14, 2026 | | Issue | 2:06 PM - April 14, 2026 | 2:06 PM - April 14, 2026 | | index | 2:06 PM - April 14, 2026 | 2:06 PM - April 14, 2026 | | Project Proposal | 2:06 PM - April 14, 2026 | 2:06 PM - April 14, 2026 | | Base | 2:06 PM - April 14, 2026 | 2:06 PM - April 14, 2026 | | Daily2 | 2:06 PM - April 14, 2026 | 2:06 PM - April 14, 2026 | | Policy | 2:06 PM - April 14, 2026 | 2:06 PM - April 14, 2026 | | Device | 2:06 PM - April 14, 2026 | 2:06 PM - April 14, 2026 | | LearningModule | 2:06 PM - April 14, 2026 | 2:06 PM - April 14, 2026 | | IssueTest | 2:06 PM - April 14, 2026 | 2:06 PM - April 14, 2026 |