Reference.
Resources, IPs, paths, models, glossary. The lookup index for everything else.
Hosts & networks
Hardware
| Contabo VPS | 8 vCPU · 24GB · Ubuntu 24.04 |
| Davas (Windows) | i7-13620H · RTX 4050 · 32GB |
| Homelab | HP ProDesk · i5 7th · 16GB · 256GB SSD + 4TB |
| Hypervisor | Proxmox VE |
Networking
| Contabo | 173.212.206.198 |
| Contabo · Tailscale | 100.101.161.24 |
| Davas · Tailscale | 100.84.112.38 |
| Homelab · Tailscale | 100.97.203.77 |
| Langfuse | 100.97.203.77:3500 |
Cloudflare-tunnelled endpoints
| Service | URL | Host |
|---|---|---|
| Runner UI | control.davidcockson.com | Contabo |
| Davas Ollama | davas-ollama.davidcockson.com | Davas |
| Plex | plex.davidcockson.com | Homelab Prod |
| Gitea | git.davidcockson.com | Homelab Prod |
| Uptime Kuma | status.davidcockson.com | Homelab Prod |
| Langfuse | langfuse.davidcockson.com | Homelab Prod |
Routing table
| Model | Provider | Use |
|---|---|---|
| qwen2.5:14b | Ollama (Contabo) | Always-on default |
| gemma4:27b | Ollama (Davas) | Project mode · health-check first |
| llama-3.3-70b-versatile | Groq free | Fast iteration · reasoning |
| models/gemini-2.5-flash | Gemini free | Long context · planning |
| qwen/qwen3-32b | Groq free | Coding tasks |
| claude-sonnet-4-6 | Anthropic (paid) | Use sparingly · max_tokens=8192 |
Where things live
Contabo
- /opt/runner/ — compose + .env
- /root/obsidian-vault/ — Syncthing target
- /root/obsidian-vault/_queue/
- /root/obsidian-vault/_active/
- /root/obsidian-vault/_completed/
- /root/obsidian-vault/_failed/
- /root/obsidian-vault/runner-outputs/
Homelab + IaC repo
- /opt/docker/prod-services/docker-compose.yml
- /opt/docker/dev-learning/docker-compose.yml
- /etc/nginx/
- homelab-iac/ansible/inventory/hosts.yml
- terraform state · S3 backend bucket
- s3://homelab-backup-fast
- s3://homelab-backup-slow
Limits worth knowing
Groq
~14,400 / day
Fast chat, iteration
Gemini 2.5 Flash
1,500 / day
Long context, planning
AWS S3
$120 credits
Backups, both tiers
HuggingFace
Not in use
Free tier too unreliable
Terms in 12 words or fewer
- MCP
- Model Context Protocol — standard way for an LLM to call tools.
- LangGraph
- Graph runtime for LLM workflows — nodes, edges, state.
- Qdrant
- Vector database. Stores embeddings, returns nearest neighbours.
- Tailscale
- Mesh VPN. Every host gets a stable 100.x.x.x address.
- Restic
- Encrypted, deduplicating backup tool. S3 backend supported.
- SearXNG
- Self-hosted metasearch. Aggregates Google, Bing, etc.
- Langfuse
- LLM observability — every
@observe'd call traced. - Davas
- Windows laptop running Gemma4 over Ollama for project mode.
- MTP
- Multi-token prediction. Davas-specific Gemma4 inference variant.
- SSE
- Server-Sent Events — one-way streaming over HTTP.
Copy at the end of every session
# Session Note — [DATE] ## Current Phase Runner: Phase ___ / Homelab: Phase ___ ## Completed This Session - ## In Progress (resume here) - ## Blockers - ## Next Actions 1. 2. 3. ## Files Changed -