Turn VaultMind from a read-only explorer into a vault builder. Drop raw content in; the AI compiles it into structured, interlinked Obsidian notes written directly to disk — the Karpathy "LLM Wiki" pattern, in the browser.

• Drop zone — drag files into the 📥 tab. Supported: .md .txt .pdf .html .csv .json, images (OCR via Tesseract), audio (transcription via Whisper).
• 🔗 URL — paste a page URL; VaultMind fetches and converts it. CORS-blocked? Paste the text instead.
• 📋 Paste — paste any text (article snippet, quote, thought) directly into the queue.
• 🎤 Record — capture a voice memo in the browser; Whisper transcribes it automatically.
• _inbox/ watcher — save files to the _inbox/ folder from any app on your computer. VaultMind detects them within 5 s and queues them. Enable Auto to process without clicking.
• Processing pipeline — extract text → AI analysis (summary, entities, tags, links) → source note saved to _sources/YYYY-MM-DD/ → entity/concept pages proposed for review.
• Review overlay — accept or skip each proposed note individually, or hit Accept All. Accepted pages are written to vault root and appear in the graph immediately.
• Deduplication — SHA-256 hash checked on every file. Duplicate sources are silently skipped.
• Done items fade out after 30 s. Errors stay until you reload.

Folder mode required for _inbox/ watching and writing notes to disk. ZIP mode still works for manual drops — notes are held in memory for the session.

The Audit sidebar tab gives a structural health report of your vault, plus three AI-powered wiki maintenance operations.

• Orphan Notes — notes with no inbound or outbound links at all.
• No Inbound Links — notes that link out but nothing links to them.
• Broken Links — [[wikilinks]] that don't resolve to any note.
• Most Linked — top hub notes by inbound link count.
• Tag Cloud — click any tag to filter the graph to only those notes.
• Frontmatter Lint — flags frontmatter fields present in most notes but missing in a few.

Wiki Maintenance — three on-demand operations:

• 🔎 Analyze gaps — no AI needed. Finds entities mentioned in 2+ source notes that have no page, broken links referenced from 2+ notes, and entity pages with only 1 source. Each gap shows a Create page button (AI-generated) or Skip.
• 🕓 Check staleness — no AI needed. Compares each entity page's last_updated timestamp against newer source notes that mention it. Stale notes get a Re-synthesize button — the AI rewrites the page body incorporating all newer sources.
• 🔍 Scan contradictions — AI-powered. Compares your recent source notes against existing entity pages and surfaces factual conflicts (different dates, numbers, claims). Each conflict can be kept, flagged, or dismissed.

Ask questions about your vault. Context is pulled automatically from the selected note and its neighbours.

• Local (WebGPU) — runs a small LLM entirely in your browser. No data leaves the device. Requires Chrome 113+ or equivalent.
• Auto-load — after your first successful model load, VaultMind remembers your choice. On the next vault open it checks the browser cache and loads the model automatically if it's already downloaded.
• First load is slow — the model downloads once (~1–4 GB) and is cached. Every subsequent session loads in seconds.
• API mode — connects to any OpenAI-compatible endpoint: OpenRouter, Groq, LM Studio, Ollama, and more. No API key needed for local servers.
• System prompt — expand System prompt inside Settings to customise the AI's persona and instructions. Changes persist across sessions. Reset restores the default.
• Click a node first — chat scopes context to that note and its 1-hop neighbours. The context pill in the header shows what's in scope.
• 🌐 Web search — toggle to augment answers with live search results (Tavily, Brave, or SearXNG).

LM Studio / Ollama? Enable CORS: LM Studio → Local Server → Server Settings → Enable CORS. Ollama: restart with OLLAMA_ORIGINS="*" ollama serve.

VaultMind uses Retrieval-Augmented Generation (RAG) to ground AI answers in your actual notes — not the model's training data.

• How it works — notes are split into 800-character chunks and embedded with all-MiniLM-L6-v2 running in the browser via WASM. Each chunk gets a 384-dim semantic vector.
• Persistent index — embeddings are saved to IndexedDB after the first run. On subsequent loads, only notes whose content changed are re-embedded. Everything else loads from cache instantly.
• Micro-RAG on node click — selecting a node pre-warms a context window scoped to that note and its direct neighbours only.
• Full-vault fallback — if no relevant local chunks exist, search expands to the entire vault.
• Cosine similarity — chunks ranked by semantic closeness to your question. Only above-threshold chunks are included.
• Live updates — accepting new notes via the ingest pipeline triggers an incremental re-index in the same session.
• re-index link — appears next to the RAG status line. Click it to wipe the cache and embed everything from scratch if needed.

Tip: click a note before asking for the most focused answer. The status line shows cached ✓ when the full index was loaded from disk — no wait needed.

Click any note to open the preview panel, then press ✏️ to edit it in place.

• Toolbar — B / I wrap selection; H1–H3 prefix the line; `c` inline code; ``` code block; 🔗 link; • / 1. / ☐ lists; — rule; " quote.
• Ctrl / ⌘ + B / I / S — bold, italic, save.
• Autosave — tick the checkbox; saves 1.5 s after each keystroke.
• Save — writes back to the .md file on disk (folder mode only).
• Unsaved draft — switching notes without saving keeps your edits in memory (amber ● unsaved draft badge). Return to continue, or Discard.

The search box in the Notes sidebar searches full note bodies, not just titles.

• Multi-word — each word scored independently; notes mentioning all terms rank higher.
• Title hits weighted 10× body hits — a title match always floats to the top.
• Snippets — short extract around the first match, with terms highlighted.
• No query — list reverts to link-degree order (most connected first).
• Tag filter — click any tag in the Audit tab to filter the graph; the active tag is shown in the filter bar above the canvas.
• Search is instant and in-memory — no network, no indexing step.