Obsidian plugin setup — Mem-style hybrid

This doc closes the human-side experience gap between Obsidian and Mem.ai by layering AI-powered plugins on top of the vault. All plugins are configured to use cloud-first.ai LiteLLM (Aspire-owned infra, $0 marginal cost per query).

⚠️ Pre-requisite: vault is cloned + opened in Obsidian, and the 3 base plugins from README.md (Git, Templater, Dataview) are already installed.

Shared LiteLLM config

Every plugin below uses the same backend. Save these somewhere convenient:

Note on embeddings: LiteLLM doesn't expose an embedding model yet (Stage 1 of KNOWLEDGE_OS_PLAN). Plugins that need embeddings either run a local model (BGE-micro fits in 100MB and is excellent), or fall back to OpenAI text-embedding-3-small. Smart Connections handles this transparently.

Plugin 1 — Smart Connections (priority, install first)

What it adds: A sidebar panel that shows semantically related notes for whatever you're currently reading. Plus a vault-wide chat that grounds its answers in retrieved notes (RAG, but you don't have to know that). Closest single-plugin match to Mem.ai's "surface what's relevant" magic.

Install:

1. Settings → Community plugins → Browse → search "Smart Connections" → Install → Enable

Configure:

1. Settings → Smart Connections
2. Embedding model: keep the default BGE-micro-v2 (local, fast, free; Smart Connections downloads it on first use). Skip the OpenAI fallback unless you specifically want it.
3. Smart Chat → Model:
4. Provider: Custom Local
5. URL: https://llm.cloud-first.ai/v1/chat/completions
6. Model name: Qwen/Qwen3.6-35B
7. API key: paste your knowledge-os key
8. Click Test Connection — should return a Pong-style success.
9. Open command palette (Cmd+P) → "Smart Connections: Open Smart View" → docks to right sidebar.

First-run wait: the BGE embedder runs over the entire vault on initial enable (~3-5 min for the current 57-file vault; minutes scale linearly).

Plugin 2 — Text Generator

What it adds: Inline AI completion while typing. Cmd+J expands a prompt or completes the current line. Mem-style "smart writing assistant" experience.

Install:

1. Settings → Community plugins → Browse → search "Text Generator" → Install → Enable

Configure:

1. Settings → Text Generator → LLM Settings
2. Provider: OpenAI (yes, even though we're not using OpenAI — the plugin's "OpenAI" provider just means OpenAI-compatible REST)
3. Endpoint URL: https://llm.cloud-first.ai/v1
4. API key: paste your knowledge-os key
5. Default model: Qwen/Qwen3.6-35B
6. Max tokens: 1024 (start conservative; raise if outputs are getting truncated)
7. Temperature: 0.5 (good for note-style content)
8. Templates folder: point at _templates/ so it picks up Templater templates
9. Click Test — should return a sample completion.

Default hotkey: Cmd+J (Mac) — generate continuation from cursor. Other useful commands (assign hotkeys via Settings → Hotkeys):

• "Text Generator: Generate text" — generic completion
• "Text Generator: Insert generated text from template" — uses one of your _templates/ files

Plugin 3 — Copilot for Obsidian

What it adds: Chat sidebar with vault context, similar to ChatGPT/Claude inside the editor. Can summarize the current note, find connections, or draft new content based on selection.

Install:

1. Settings → Community plugins → Browse → search "Copilot" → look for "Copilot for Obsidian" by logancyang → Install → Enable

Configure:

1. Settings → Copilot
2. Default chat model: Custom (drop-down)
3. Custom Model Configuration:
4. Model name: Qwen/Qwen3.6-35B
5. Base URL: https://llm.cloud-first.ai/v1
6. API key: paste your knowledge-os key
7. Embedding provider: Local (uses BGE) — same as Smart Connections, avoids the OpenAI dependency
8. QA mode: Vault QA — chat answers from your notes, not from training data
9. Click Verify — confirms the endpoint works.

Open the chat: ribbon icon (chat bubble) on the left rail, or command palette → "Copilot: Open Copilot Chat".

Plugin 4 — Khoj (optional — adds a service dependency)

What it adds: Best-in-class "talk to your vault" experience. More polished than Copilot's chat. Search across vault + remembered conversations + connected sources. Closest Obsidian comes to Mem.ai out of the box.

Trade-off: Requires running a Khoj backend (Python service). Two options:

• Khoj Cloud (paid subscription) — defeats the iZinga $0-marginal-cost goal, skip for BaaS purposes.
• Self-hosted Khoj — Docker container, free. Adds another service to maintain alongside KO engine.

For Aspire-internal use today: install only if you want to test the polish; not on the critical path.

Install (self-hosted path):

# 1. Run Khoj backend in Docker
docker run -d --name khoj \
  -p 42110:42110 \
  -v ~/Desktop/Claude/knowledge-os-vault:/data/vault:ro \
  -e KHOJ_OPENAI_API_BASE=https://llm.cloud-first.ai/v1 \
  -e KHOJ_OPENAI_API_KEY=<your knowledge-os key> \
  -e KHOJ_DEFAULT_MODEL=Qwen/Qwen3.6-35B \
  ghcr.io/khoj-ai/khoj:latest

Then: Settings → Community plugins → Browse → "Khoj" → Install → Enable. Plugin defaults to http://127.0.0.1:42110 which matches the container above.

My take: skip Khoj for Stage 1. Smart Connections + Copilot already deliver 90% of Khoj's value without the extra container. Revisit Khoj if and when iZinga has Stage-2 traffic that justifies the polish.

Recommended install order

1. Smart Connections — biggest win; covers "surface related notes" + RAG chat
2. Text Generator — inline completion; the "writing assistant" experience
3. Copilot for Obsidian — secondary chat panel; useful if you'd rather keep Smart Chat read-only and use Copilot for drafting
4. (optional) Khoj — only if 1+2+3 leaves a polish gap you can't ignore

Realistically: install #1 and #2 today (10 min total), decide on #3 and #4 later based on actual feel.

How to test it's working

After installing Smart Connections + Text Generator:

1. Open shared/2026-04-19-vault-initialized.md
2. Smart Connections check:
3. Right sidebar should populate with "Connections" panel showing 5-10 related notes from the same vault. The first time, wait for embedding to finish (status bar shows progress).
4. Try the chat: ask "what is the Knowledge OS vault?" — answer should cite specific notes.
5. Text Generator check:
6. Move cursor to end of a sentence in any note.
7. Cmd+J — should generate a continuation that references nearby content from the note.
8. If either fails, run the Test / Verify button in the plugin's settings page; the error usually points at a wrong endpoint URL or a stale API key.

Cost / observability

All chat completions through these plugins go to llm.cloud-first.ai/v1, which routes through your knowledge-os virtual key. You can:

• Watch usage live: https://llm.cloud-first.ai/ui/?page=usage
• Set a daily/monthly cap on the key (Settings → Virtual Keys → knowledge-os → Spend Limit)
• Rotate the key any time (1Password → generate new → update in all 4 plugin settings)

Embedding traffic stays local (BGE in Smart Connections + Copilot) — never leaves your machine, never hits LiteLLM, free and offline-capable.

What this doesn't replace

The plugins above give the human side the Mem-style polish. They don't replace what the engine side does:

• mcp-bridge knowledge_query — agent-callable, returns JSON with citations (T3 ✅ in aspire/infrastructure/knowledge-os)
• worker-ocr — auto-tag + auto-link on ingest (T7, in flight)
• lint worker — find contradictions and stale claims across the vault (T12)
• LightRAG semantic graph (Stage 2) — when the vault grows past ~500 pages

Plugins polish the writing experience for humans. Engine builds the agent-facing brain. Hybrid stack runs both at the same time.

Sources

• aspire/izinga/KNOWLEDGE_OS_SPEC.md — overall architecture
• aspire/izinga/AGENT_LLM_PLAN.md — model routing across cloud-first.ai
• Smart Connections: https://github.com/brianpetro/obsidian-smart-connections
• Text Generator: https://github.com/nhaouari/obsidian-textgenerator-plugin
• Copilot for Obsidian: https://github.com/logancyang/obsidian-copilot
• Khoj: https://github.com/khoj-ai/khoj