Vault Brain

Fully-local, multimodal AI for Obsidian. Talk to your vault, snap images into notes, search across everything semantically, and write with AI — all powered by Gemma 4 on Ollama. No cloud. No API keys. Zero data leaves your machine.

Every AI feature runs against a local Ollama endpoint on 127.0.0.1. There is exactly one network call site in the entire codebase (src/core/ollama-provider.ts) and it only ever talks to your configured localhost endpoint. See Privacy & network audit.

Contents: Features · Requirements · Setup · Commands · Settings · Privacy · Development · License

Features

🎙️ Voice

• Voice memo → structured note — right-click an audio file (or process your most recent) → verbatim transcript + ≤5-bullet summary + extracted - [ ] tasks, written to your daily note (or a new note). Serbian & English. Your original audio is never modified.
• In-app recording — a mic ribbon button opens a floating control bar with Pause / Resume / Stop / Cancel and a live timer. Stop → it transcribes and opens the note at the new memo. (Transcribe-and-discard — no audio clutter.)
• Microphone selection — choose your input device in settings.
• Meeting mode — right-click audio → "Process as meeting (who said what)" → diarized transcript + decisions + action items.
• Auto-watch folder — drop audio into a configured folder and it's transcribed automatically.

🖼️ Vision

• Image → Markdown — cursor on an embedded image → "Extract text from image below cursor" → clean Markdown (tables included) inserted below it.

💬 Chat, search & edit

A right-sidebar panel with three modes:

• This note + links — chat over the active note and its 1-hop links. Add more with the + Context picker (any note or folder).
• Whole vault (RAG) — semantic search across your whole vault with [[note]] citations (local nomic-embed-text, file-based index, auto-built & incremental).
• Edit this note — give an instruction ("replace X with Y", "make headings consistent"); preview a red/green diff, then Apply (Cmd-Z undoable).

Answers render as Markdown with clickable [[links]], a copy button, and modern chat bubbles.

🔗 Discovery

• Related notes sidebar — a live panel showing the most semantically-related notes to whatever you're viewing (reuses the vault index).

✍️ Writing

• Selection actions (right-click → Vault Brain): Summarize · Improve · Format · Translate SR↔EN · Fix grammar.
• Custom prompts — define your own Name :: instruction actions in settings; they appear in the same submenu.
• Continue writing — "Continue writing at cursor" streams an AI continuation in place.
• Suggest tags — AI suggests tags and writes them to the note's frontmatter.
• Link mentions — wraps plain-text mentions of your other note titles in [[ ]] (deterministic, no AI).

🔌 Status

• Connection indicator (🟢/🟡/🔴) with one-click fixes, and a 🧠 activity indicator showing live status of every running operation + a recent-activity log.

Requirements

• Obsidian 1.5.0+ on desktop (macOS / Windows / Linux). Desktop-only — local inference isn't available on mobile.
• Ollama 0.30.5+ installed and running.
• gemma4:12b (multimodal — audio + vision + text) and, for vault search, nomic-embed-text.
• ~16 GB RAM recommended (the 12B model is ~7.6 GB on disk). Apple Silicon or comparable.

Setup

1. Install Ollama (ollama.com) and make sure it's running (ollama serve, or open the app).
2. Pull the models:

   ollama pull gemma4:12b
   ollama pull nomic-embed-text   # for whole-vault search
   

3. Install the plugin (until it's in the community directory): copy main.js, manifest.json, and styles.css into <your-vault>/.obsidian/plugins/vault-brain/, then enable Vault Brain under Settings → Community plugins.
4. Open Settings → Vault Brain — defaults work out of the box; the status bar should show 🟢 Vault Brain.

About the model: gemma4:12b is the recommended build — it's fully multimodal (audio + vision + text), so voice, image OCR, and chat all run on one model. It needs Ollama 0.30.5+ (older versions return a 412 when pulling it). Want something lighter? The 8B gemma4:latest is also fully multimodal and faster (just less capable) — pick it in Settings → Vault Brain → Model. Avoid gemma4:12b-mlx: that MLX build is text-only (no voice or vision).

Remote server (advanced)

Vault Brain can also talk to Ollama running on another machine you own — a home server or a DGX-class box — so a light laptop can use a heavy model:

• Bare Ollama on your LAN — set Ollama host to http://192.168.x.y (keep the port). Recent Ollama allows Obsidian's origin by default; if you hit a CORS error, start the server with OLLAMA_ORIGINS="app://obsidian.md".
• Open WebUI (fronts Ollama with API keys) — set Ollama host to https://your-host/ollama (Open WebUI's Ollama proxy path; the port setting is ignored when the host has a path) and paste an API key from Open WebUI → Settings → Account into the API token setting. The token is sent as Authorization: Bearer … on every request.

Privacy: pointing the plugin at a remote server sends your note content to that machine — only use hardware you trust. The token is stored unencrypted in .obsidian/plugins/vault-brain/data.json. The single auditable network egress is unchanged.

Commands

Voice memo → note · Start/Stop voice recording · Process as meeting · Extract text from image · Open Q&A panel · Open related notes · Rebuild vault index · Selection: Summarize/Improve/Format/Translate/Fix grammar (+ your custom prompts) · Continue writing at cursor · Suggest tags for this note · Link mentions of existing notes · Test connection.

Settings

Privacy & network audit

Local-only by design. Verify it yourself:

1. Read the code. One network call site — src/core/ollama-provider.ts:

   grep -rln fetch src/      # → only src/core/ollama-provider.ts
   

2. Watch the traffic. With Little Snitch / lsof -i / a proxy, exercise every feature — the only connections are to 127.0.0.1:11434.
3. Pull the plug. Turn Wi-Fi off and record/transcribe a memo — it still works end-to-end.

Development

npm install
npm run dev      # esbuild watch → main.js
npm test         # node:test unit + live Ollama contract tests (skip if offline)
npm run build    # type-check + production bundle

Architecture: a thin, obsidian-free src/core/ (24 unit-tested modules — transport, context building, prompts, parsing, similarity, diff) behind an LlmProvider interface, with 9 feature orchestrators in src/features/. 87 tests. Design specs and build plans live in docs/.

License

MIT — see LICENSE.