Current


Access Devices / Apps	Homeserver	Cloud Services
Laptop + Obsidian + Browser - iCloud > Resilio > Homeserver sync - Gemini / Mammouth @ Browser	Monitoring Script: Triggers the Quartz build and pushes to Cloudflare when Resilio syncs	Quartz @ GitHub Cloudflare Pages - Web Analytics
iPhone/iPad + Obsidian + Browser - iCloud for Sync OR - Syncthing - Gemini / Mammouth (Web) Apps

Issues

AI exchanges are separate apps and I need to copy and paste notes across
Vector Databases are stored on the AI platform.
These are essentially my personal information in a form that is extensive and easily accessed (if the cloud provider authorises access)
If I wish to Reprompt in Mammouth then each AI will need to re-vectorize attached files / images which is wasted CPU
Prompt context needs to be hand stitched and/or via Gemini “Personal Context” feature
Mammouth includes API usage in its current subscription plan

Business Requirement

I want the ability to effectively and efficiently integrate with LLM using Obsidian on all my devices and publish notes for given folder(s) to my domain using reliable tools

Solution Option(s) and Functional Requirements

Initial

Continue to host my Vault in markdown files / folders
Continuing to use Obsidian from my access devices
Pivot to Syncthing to synchronize direct to Homeserver, bypassing Laptop Resilio bridge
Obsidian Sync has a 1GB ceiling for a vault and given its pay for this seems very constrained
Parallel run my main vault on the current solution until I build confidence with Syncthing
Continue to publish one or more folders to one or more sites using Quartz
This can be as is with Quartz building my site on my homeserver
Though this will involve me manually upgrading the Git files
Alternatively I can publish to a private vault
Introduce the ability to prompt an AI LLM using one or more files from my Vault
This builds upon existing CLI based solutions that exist and get invoked from inside Obsidian
Ability to reprompt to another AI LLM leveraging vault files and existing vectorized databases
Self host the vectorized database to reduce Mammouth API consumption
Usage will be by family of 3 based on their own individual vaults

Evaluate the ability to implement this in phases? and rate the difficulty for a retired IT Business Analyst?

Alternate

Are there alternate approaches that can help provide context from one or more vault files / folders and flexibility to interface with multiple AI LLMs seamlessly?

claude response

Solution Evaluation: Obsidian + LLM Integration

Architecture Overview

┌─────────────────────────────────────────────────────────────┐
│                    ACCESS DEVICES                           │
│  Laptop/iPhone/iPad                                         │
│  Obsidian Vault (Markdown)                                  │
│  + Obsidian Plugin Layer                                    │
└──────────────────┬──────────────────────────────────────────┘
                   │ Syncthing (Direct)
┌──────────────────▼──────────────────────────────────────────┐
│                   HOMESERVER                                │
│  ┌─────────────┐  ┌──────────────┐  ┌──────────────────┐  │
│  │  Syncthing  │  │  Vector DB   │  │  Quartz Build    │  │
│  │  (Sync Hub) │  │  (ChromaDB/  │  │  + Cloudflare    │  │
│  │             │  │   Qdrant)    │  │  Deploy          │  │
│  └─────────────┘  └──────────────┘  └──────────────────┘  │
│  ┌─────────────────────────────────────────────────────┐   │
│  │           LLM Proxy / Orchestration Layer           │   │
│  │         (LiteLLM or Open WebUI)                     │   │
│  └─────────────────────────────────────────────────────┘   │
└─────────────────────────────────────────────────────────────┘
                   │
┌──────────────────▼──────────────────────────────────────────┐
│              CLOUD / EXTERNAL                               │
│  Cloudflare Pages  │  OpenAI/Gemini/Anthropic APIs          │
└─────────────────────────────────────────────────────────────┘

Phase Analysis

Phase 1: Sync Migration (Syncthing)

Difficulty: 2/5 — Low

Task	Notes
Install Syncthing on Homeserver	Docker container, straightforward
Install Syncthing on Laptop	Replace Resilio bridge role
Install Syncthing on iPhone/iPad	Möbius Sync app (paid, ~£5) or Syncthing-iOS
Parallel run with current iCloud/Resilio	Run both during confidence period
Vault per family member	Separate Syncthing folders per person

Key Decisions:

Möbius Sync is the practical iOS option — background sync has iOS limitations
Recommend keeping iCloud as a backup sync path, not primary
One Syncthing instance on Homeserver serves all 3 family vaults with folder-level isolation

Risk: Low. Syncthing is mature, open source, and well documented

Phase 2: Quartz Publishing (Incremental Improvement)

Difficulty: 2/5 — Low

Current:  Resilio → Homeserver → Monitoring Script → Quartz Build → Cloudflare
Target:   Syncthing → Homeserver → Watch Script → Quartz Build → Cloudflare

Improvement Option	Difficulty	Notes
Keep as-is, swap Resilio trigger for Syncthing trigger	1/5	Minimal change
Auto Git pull for Quartz upgrades via cron	2/5	Small script addition
Private vault (second Quartz instance)	2/5	Second folder, second Cloudflare project
Per-family-member published site	3/5	Three Quartz builds, three Cloudflare projects

Recommendation: Swap trigger script only. Add a monthly cron job:

cd /path/to/quartz && git pull && npm install

This removes the manual upgrade burden at minimal effort

Phase 3: LLM Integration via Obsidian

Difficulty: 3/5 — Moderate

Option A: Obsidian Plugin Route (Lowest friction)

Plugin	Function	API Key Needed
Copilot	Chat with vault files, multi-LLM	Yes (your keys)
Smart Connections	Local vector embeddings + chat	Optional
Omnisearch	Fast local semantic search	No
Text Generator	Prompt templates, file context	Yes

Copilot Plugin is the strongest candidate:

Supports OpenAI, Gemini, Anthropic, local Ollama
Can index your vault locally
Conversations stay within Obsidian
No copy-paste between apps
API keys stored locally

This directly solves Issues 1, 4, and 5 from your problem statement

Option B: CLI + Obsidian (Your existing approach, enhanced)

# Example: shell command invoked from Obsidian via Shell Commands plugin
obsidian-ai-query.sh --files "Health/2026-targets.md" \
                     --model "gemini-pro" \
                     --prompt "Analyse my progress"

Shell Commands plugin calls your existing CLI tools from within Obsidian — keeps your current workflow but removes app switching

Phase 4: Self-Hosted Vector Database

Difficulty: 4/5 — Moderate-High

Why this matters for you:

Vectorize once, query many times across different LLMs
Keeps your personal health/life data off commercial platforms
Shared infrastructure for family of 3

Recommended Stack

┌─────────────────────────────────────────┐
│           Homeserver                    │
│                                         │
│  ┌─────────────┐   ┌─────────────────┐ │
│  │   Qdrant    │   │   Open WebUI    │ │
│  │  (Vector DB)│   │  (LLM Frontend) │ │
│  │  Docker     │   │  Docker         │ │
│  └──────┬──────┘   └────────┬────────┘ │
│         │                   │          │
│  ┌──────▼───────────────────▼────────┐ │
│  │         LiteLLM Proxy             │ │
│  │  Routes to OpenAI/Gemini/Anthropic│ │
│  │  Single API endpoint              │ │
│  │  Usage tracking per family member │ │
│  └───────────────────────────────────┘ │
└─────────────────────────────────────────┘

Component	Purpose	Docker?	Difficulty
Qdrant	Vector storage	Yes	2/5
LiteLLM	Multi-LLM proxy, single endpoint	Yes	3/5
Open WebUI	Browser-based chat UI, vault file upload	Yes	2/5
Ollama	Local LLM (optional, no API cost)	Yes	2/5

LiteLLM specifically addresses Issues 2, 3, 4, and 6:

One API key management point
Vectors stored on your homeserver (Qdrant)
Route the same query to multiple LLMs
Track and control Mammouth/API spend per user

Vectorisation Workflow

Obsidian Vault (Markdown)
        │
        ▼
  Python Script (runs on Homeserver, watches Syncthing folder)
  - Detects new/modified .md files
  - Chunks content
  - Generates embeddings (via local model or API)
        │
        ▼
     Qdrant
  (self-hosted vector store)
        │
        ▼
  LiteLLM Proxy
  (augments prompts with relevant vault context)
        │
        ▼
  Your chosen LLM (Gemini/Claude/GPT)

This is the hardest phase — Python scripting required. As a retired IT BA, the concepts will be familiar but the implementation requires comfort with:

Docker Compose files
Basic Python scripting or adapting existing scripts
REST API concepts

Phase 5: Family Multi-Vault

Difficulty: 3/5 — Moderate

# Syncthing folder structure
/homeserver/vaults/
  ├── vault-you/          # Your Syncthing folder ID
  ├── vault-member2/      # Separate folder, separate devices
  └── vault-member3/      # Isolated, no cross-access
 
# LiteLLM user isolation
/litellm/config.yaml
  users:
    - user: "you"
      vault_path: "/vaults/vault-you"
      budget_monthly_usd: 20
    - user: "member2"
      ...

Qdrant supports named collections — one collection per family member keeps vectors isolated

Recommended Phased Roadmap

Q1 2026          Q2 2026          Q3 2026          Q4 2026
─────────────────────────────────────────────────────────
Phase 1          Phase 2          Phase 3          Phase 4+5
Syncthing        Quartz           LLM Plugin       Vector DB
Migration        Tidy-up          (Copilot)        + LiteLLM
                                                   + Family
[2/5]            [2/5]            [3/5]            [4/5]
2-4 weeks        1-2 weeks        2-3 weeks        4-8 weeks

Start with Copilot plugin in Phase 3 before committing to Phase 4 — it may satisfy 80% of your needs at 20% of the effort

Alternate Approaches

Alt 1: Open WebUI (Fastest path to multi-LLM)

Single Docker container on Homeserver
Browser-based, works on all devices
Upload vault files per session (not auto-indexed)
Supports Ollama local + OpenAI/Gemini APIs
Difficulty: 2/5 — good stepping stone before full vector DB

Alt 2: Notion AI / Anytype (Abandon Obsidian)

Not recommended — loses your existing vault investment and publishing workflow

Alt 3: RAGFlow (All-in-one RAG platform)

Docker-based, includes vector DB + LLM proxy + UI
Designed for document knowledge bases
More opinionated than the modular stack
Difficulty: 3/5 — worth evaluating if Phase 4 feels too complex

Alt 4: n8n Workflow Automation

Self-hosted on Homeserver (Docker)
Orchestrates: file watch → vectorise → LLM query → write back to vault
Visual workflow builder — suited to BA thinking
Bridges Obsidian, LLMs, and vector DB without heavy coding
Difficulty: 3/5 — recommended complement to Phase 4

Summary Recommendation

Priority	Action	Effort	Payoff
Do first	Phase 1: Syncthing	Low	Removes Resilio dependency
Do next	Phase 3: Copilot Plugin	Low-Med	Solves LLM integration immediately
Then	Phase 2: Quartz cron	Low	Removes manual upgrade task
Plan for	Phase 4: LiteLLM + Qdrant	High	Full data sovereignty + family use
Consider	n8n as orchestration layer	Med	Reduces coding need in Phase 4

Your BA background is a genuine asset here — system decomposition, requirements, phasing — the technical gaps are Docker familiarity and basic Python, both learnable incrementally alongside each phase

My Journey

Explorer

AI enabled note taking