An Obsidian vault explorer and builder that runs entirely in your browser — force graph, semantic search, AI chat, in-place editing, and a full ingestion pipeline that turns raw content into structured, interlinked notes. No server. No API keys required. Nothing leaves your device.

→ Try it live

Drop your Obsidian vault (as a folder or a ZIP) and get:

• Force-directed graph — every note is a node, every [[wikilink]] is an edge. Clusters emerge automatically. Switch between 2D and 3D perspective at any time.
• Ingestion pipeline — drop files, paste URLs, record voice memos, or save to _inbox/. The AI extracts entities, writes source notes, proposes linked entity pages, and updates the graph live.
• AI chat — ask questions about your vault. Context is scoped to the selected note and its 1-hop neighbours (micro-RAG). Works with a local WebGPU model (Gemma 4, Qwen, SmolLM) or any OpenAI-compatible API.
• Persistent semantic index — embeddings are saved to IndexedDB after the first run. Subsequent loads only re-embed notes that changed. A 500-note vault goes from "indexing…" to "cached ✓" in under a second.
• Wiki maintenance — gap analysis, staleness detection, and AI-powered contradiction scanning keep your knowledge base consistent over time.
• Full-text search — multi-word, frequency-ranked search across all note bodies with highlighted snippets.
• Note editor — edit any note in place with a markdown toolbar. Autosave to disk, unsaved-draft recovery, Ctrl+B/I/S shortcuts.
• Semantic audit — orphan detection, broken link report, frontmatter lint, top hubs, tag cloud.
• Web search — optional live augmentation via Tavily, Brave, or SearXNG.

Everything is a single HTML file. Open it, load a vault, start exploring.

# Serve locally (required for File System Access API and WebGPU)
python3 -m http.server 8080
# open http://localhost:8080

Or open index.html directly — ZIP drop works in all browsers without a server, though writing to disk requires folder mode.

Implements the Karpathy "LLM Wiki" pattern — a compounding knowledge base maintained by an AI — running entirely in the browser with no terminal, no CLI, no cloud.

Manual (📥 Ingest tab)

• Drag and drop files directly into the tab
• Paste a URL — page is fetched and converted to markdown
• Paste any text — clipboard captures, quotes, fragments
• Record a voice memo — Whisper WASM transcribes it

Automatic (_inbox/ folder watcher) Save files to _inbox/ from any app on your computer. VaultMind detects them within 5 seconds. Enable Auto to process without clicking.

PDF.js, Tesseract, and Whisper are all lazy-loaded on first use — zero overhead if you don't use them.

The AI adjusts its extraction strategy based on what you drop in:

1. Extract — text pulled from the file using the appropriate method above
2. Mode detect — content type mapped to an extraction strategy (see table above)
3. AI analysis — summary, entities, tags, wikilinks, and (for voice) action items
4. Source note — written to _sources/YYYY-MM-DD/ with full Obsidian frontmatter; entity wikilinks injected deterministically. Voice memos include a checkbox action items list.
5. Review — proposed entity/concept pages shown for Accept / Skip. Accept All closes the panel automatically
6. Live update — accepted pages written to vault root; graph re-renders with new nodes and edges immediately
7. Incremental re-index — new notes are embedded and added to the semantic index in the same session, no reload needed
8. Log — _log.md appended; _index.md rebuilt

SHA-256 deduplication prevents the same file being processed twice. Done items fade from the queue after 30 s; errors stay until reload.

my-vault/
├── _inbox/                  ← watched folder (drop files here from any app)
├── _sources/
│   └── 2026-04-05/
│       ├── article.md       ← source note: summary + entities + original content
│       └── voice-memo.md    ← Whisper transcript
├── _index.md                ← auto-maintained catalog
├── _log.md                  ← append-only ingest history
├── _schema.md               ← AI processing rules (user-editable)
├── Andrej Karpathy.md       ← entity page (AI-created)
├── WebGPU.md                ← concept page (AI-created)
└── ...                      ← your existing notes (untouched)

VaultMind embeds every note using all-MiniLM-L6-v2 (384-dim vectors, WASM) and stores the result in IndexedDB — so the index persists across browser sessions.

Micro-RAG — when a note is selected, context is scoped to that note and its 1-hop neighbours before expanding to the full vault. Cosine similarity threshold: 0.2 (local) / 0.25 (global).

Three on-demand operations that keep the knowledge base healthy:

Each result shows one-click actions: Create page (AI-generated), Re-synthesize (AI rewrites from all sources), Keep / Dismiss.

⬡ Generate MOC — select any connected note in the graph and click the button that appears. The AI reads the note and all its direct links, groups them into themed sections, and writes a Map of Content page to the vault root.

Click Load Model. The model downloads once and is cached in the browser. On subsequent visits VaultMind checks the cache and auto-loads your last model — no interaction needed.

Requires Chrome 113+ or any WebGPU-capable browser.

Any OpenAI-compatible endpoint. Click 🔄 to auto-discover models.

CORS for local servers: LM Studio → Local Server → Server Settings → Enable CORS. Ollama: OLLAMA_ORIGINS="*" ollama serve.

Expand System prompt inside chat Settings to customise the AI's persona and instructions. Changes persist across sessions. Reset restores the default.

Coloured with mexico-10 · NOCHE DE MUERTOS — Day of the Dead night, ofrenda lit by marigold candles. The graph reads as a memorial: cobalt notes, ofrenda-red orphans, copal-incense lime hubs, pan-de-muerto tan MOCs, marigold-candle yellow when selected. A second-brain rendered as the night you keep vigil for what came before.

Palette pulled from Rangrez, the global colour-palette library that backs all NakliTechie projects.

The entire app is a single HTML file.

VaultMind is smaller than the HTML alone on most news pages — and does local LLM inference, semantic search, OCR, and audio transcription on top of it. The only things fetched at runtime are the AI models, and those are cached in the browser after the first download.

Built by Chirag Patnaik

Built with Claude.