My Journey

A evolving with self hosted services

4 min read

this is a really solid direction — but I’m going to push you slightly so you don’t overbuild something painful to maintain on that NUC 🙂

Your J5005 + 16GB RAM is the key constraint. It’s fine for orchestration + light AI, but not for heavy local models.

So the winning strategy is:

🧠 “thin local brain + external model power”
(not “run everything locally”)

🧭 the architecture that actually fits your setup

Think in 3 layers:

1. 🧩 orchestration (brain)

→ n8n (good choice)

2. 💬 interface (your “ChatGPT replacement”)

→ web UI for LLMs

3. 🧠 memory layer

→ vector DB (only if you actually need it)

✅ 1. n8n — yes, but keep it focused

n8n is perfect for:

  • publish pipeline (Obsidian → Quartz → Cloudflare)
  • daily news ingestion
  • RSS → summarise → store
  • visitor alerts / analytics summaries
  • scheduled jobs

👉 runs very comfortably on your NUC

⚠️ but don’t try to make it everything

n8n is:

  • great at workflows ✅
  • bad as a UI ❌
  • not a knowledge system ❌

✅ 2. web LLM interface (this is your biggest win)

You’re 100% right to want to escape:

“CLI + no conversation navigation”

That’s painful long-term.

🔥 best option for you: Open WebUI

(this is basically what you’re describing)

✅ gives you:

  • chat UI like ChatGPT
  • sidebar with conversations ✅
  • supports multiple providers:
    • OpenAI
    • Ollama
    • any API (via OpenAI-compatible endpoints)
  • file uploads
  • RAG support later

💡 why it fits your setup

  • runs in Docker
  • light enough for your CPU
  • doesn’t require local models (but can if you want small ones)

🧠 how you’d use it with Mammouth

If Mammouth gives OpenAI-style API:

→ plug it into Open WebUI as a provider

👉 now you get:

  • proper UI
  • history
  • portability
  • no lock-in

✅ 3. local models (be careful here)

Your CPU:

Intel J5005 (no serious GPU)

Reality:

  • ✅ small models (3B–7B) → usable
  • ❌ anything “smart” → slow / frustrating

👉 recommendation

Install Ollama, but:

  • use it for:
    • quick drafts
    • embeddings
  • NOT as your main brain

✅ 4. vector database — only if needed

This is where people over-engineer.

You don’t need this yet unless you want:

  • semantic search across your vault
  • “chat with my notes”
  • long-term memory

✅ when it does make sense for you

Given your Obsidian vault + Quartz:

👉 this could be very powerful:

  • index your notes
  • ask:
    • “what have I written about liberalism + class?”
    • “summarise my last 10 posts”

🔥 best lightweight options

🟢 Chroma (easy)

  • simple
  • local
  • Python-based

🟢 Qdrant (better long-term)

  • fast
  • production-ready
  • still light enough

🧠 realistic stack for YOU

Here’s what I’d actually deploy:

🧱 core stack

  • n8n → workflows
  • Open WebUI → your main interface
  • Ollama → optional local models
  • Qdrant → only if you add RAG

🔗 how they connect

           Open WebUI
                |
        (Mammouth API / OpenAI)
                |
              n8n
                |
     -----------------------
     |         |           |
 Obsidian   RSS/news   Analytics
     |
  (optional)
   Qdrant

⚙️ concrete use cases you can build

📰 1. daily news briefing

n8n flow:

  • pull RSS (BBC, FT, etc.)
  • summarise via LLM
  • save to Obsidian
  • optional: tag topics

✍️ 2. writing assistant

Inside Open WebUI:

  • draft → refine → summarise
  • query your vault (later with vector DB)

🚀 3. auto-publish pipeline

n8n:

  • detect vault change
  • run Quartz build
  • push to GitHub
  • notify you

📊 4. visitor insights

  • Cloudflare analytics API
  • summarise daily:
    • top pages
    • trends
  • push summary into Obsidian

⚠️ important constraint (don’t ignore this)

Your bottleneck is:

❗ CPU, not RAM

So:

  • avoid heavy containers
  • avoid running multiple LLMs locally
  • keep services lean

🔥 if I were you (exact plan)

phase 1 (this week)

  • install Docker (if not already)
  • run:
    • n8n
    • Open WebUI
  • connect WebUI → Mammouth API

✅ immediate upgrade in UX

phase 2

  • add:
    • Ollama (test small models)
  • build:
    • RSS → summary → Obsidian flow

phase 3 (only if needed)

  • add Qdrant
  • implement “chat with my vault”

🧠 one pushback (important)

You said:

“vector database to optimise LLM usage”

👉 that’s only true if:

  • you’re querying your own data a lot

Otherwise:

it adds complexity without real benefit

✅ bottom line

  • n8n → yes ✅
  • Open WebUI → biggest upgrade ✅
  • local models → limited use ⚠️
  • vector DB → later, not now ✅

If you want, I can give you:

  • a ready-to-run Docker Compose for your exact setup
  • or a first n8n workflow (RSS → Obsidian → summary)

that would get you from “idea” → “working system” in like 30 mins 👍

Graph View