Vizier

A local, privacy-first AI agent for Obsidian powered by Ollama. Chat with a multi-round agent that uses vault tools, write and edit notes, run semantic search, build a knowledge graph of people and entities, track claims and contradictions, summarize videos and articles, reflect on your notes, and digitize handwritten content — everything runs on your machine. No cloud accounts, no API keys, no telemetry. Completely free and open source.

What's new in 0.7: an agentic chat loop that plans and calls tools, persistent memory of your interests, a semantic vault index, an epistemic layer (claims, contradiction scanning, knowledge-gap analysis), proactive RSS intake + daily briefings, and support for Ollama cloud models (*-cloud) alongside local ones.

Features

Agentic chat

Free-form chat runs through a multi-round agent loop: the model decides which tools to call (vault search, read/write/edit notes, fetch URLs, create entities, record claims, and more), executes them, and synthesizes an answer — citing the notes it touched with [[wikilinks]]. It keeps a per-model context budget, compacts long conversations, and learns reusable skills from successful multi-step sessions. Disable Agent loop in settings to fall back to plain single-call streaming chat.

• Memory — durable facts about you (interests, positions, current reading) are extracted from conversations and surfaced in later sessions. Say remember: <fact> to store one directly. Backed by an embedding-ranked store; toggle with the Memory setting.
• Semantic vault index — notes are chunked and embedded so search finds meaning, not just keywords. Built in the background and kept current as you edit. Rebuild any time via "Rebuild vault search index."
• Active note — the agent reads the note you have open only when you ask about "this note" (via a read_active_note tool), so an open note never silently biases unrelated requests.

Writing & editing

Command	What it does
/write <topic>	Generate a structured note with AI-suggested filename, tags, and body
/edit <instruction>	Edit the active note following an AI instruction
/read	Summarize or ask a question about the active note
Free chat	Stream a conversation with any Ollama model
Add AI abstract callout	Insert an AI-generated summary callout at the top of the active note (palette only)

Web & media

Command	What it does
/summarize <url>	Summarize a YouTube video or web article
/clip <url>	Fetch, summarize, and save a URL to your Clips folder
/clip long <url>	Clip with detailed notes — ideal for long commentaries or reads
/clip learn <url>	Clip with detailed notes plus a study guide popup — ideal for lectures
/transcribe <file or url>	Transcribe an audio file or podcast URL via Whisper
/ingest <path>	Ingest a book or PDF chapter-by-chapter into your vault
/handwriting	Transcribe a photo of handwritten notes with your vision model and save as a vault note

Human Network

Build a knowledge graph of people, events, ideas, and entities inside your vault.

Command	What it does
/person <name>	Create a person note with biography and Wikipedia data
/event <title>	Create a historical event note with date, context, and Wikipedia data
/idea <concept>	Create a concept or theory note with definition and Wikipedia data
/entity <type> | <name>	Create a note for any entity type (organization, place, movement, etc.)
/link A | B	Add a bidirectional link between two entity notes
/bridge A | B	Find the shortest connection path between two entities
/timeline <topic or range>	Build a chronological timeline (e.g. /timeline Cold War or /timeline 1939..1945)

Entity notes are organized into sub-folders by type (e.g. Human Network/Entities/organization/NATO.md). Wikipedia is searched first; if no result is found, a manual-entry prompt appears.

Analysis & research

Command	What it does
/find <query>	Natural-language search across your vault
/standardize <folder>	Add missing metadata to all notes in a folder
/recluster <folder>	Cluster notes in a folder into themes
/contradict	Find contradictions in the active note — checks vault notes and model knowledge
/sources	Audit the active note for uncited factual claims
/socratic	Generate Socratic questions for the active note and capture your answers
/thesis <tag>	Build a structured thesis document from tagged notes
/research <topic>	Deep research: vault + Wikipedia + a synthesis note with sourced claims
/curriculum <topic>	Build an ordered self-study curriculum for a topic
/gaps	Analyze the vault for missing themes, entities, and readings

/contradict outputs two sections: From vault (notes in your vault that conflict with the active note's claims) and From model knowledge (well-established facts from the model's training that conflict).

Knowledge & epistemics

Vizier can attach discrete, checkable claims to notes (with confidence and provenance: primary / secondary / model / user), cite or contest them, and scan the whole vault for claims that conflict with each other. The agent exposes these as tools (add_claim, cite_claim, contest_claim, scan_contradictions); a "Scan vault claims for contradictions" command and the daily scheduler run them too. Flagged conflicts are written to the Contradictions folder for review.

Proactive (intake & briefing)

Command	What it does
/intake	Fetch your configured RSS/Atom feeds and triage each item against your interests
/briefing	Generate a daily briefing from recent intake, vault activity, and open contradictions
/runstats	Show run statistics and model usage from the trace log

When enabled, a lightweight scheduler runs intake → briefing → contradiction scan once per day while Obsidian is open. Relevant intake items and briefings land in your Vizier inbox folder — never anywhere else in the vault. Both are off by default; enable them (and add feed URLs) in settings.

Reflection

Command	What it does
/weekly	Generate a weekly reflection scaffold from notes modified this week
/monthly	Generate a monthly reflection scaffold from notes modified this month
/freewrite	Open a new timestamped blank note for free writing

All commands are also available from the Command Palette (Cmd/Ctrl+P).

Requirements

• Ollama running locally (ollama serve)
• At least one chat model pulled, e.g. ollama pull gemma3:4b (Ollama cloud models such as gemma3:27b-cloud also work — see Models)
• An embedding model for semantic search and memory, e.g. ollama pull nomic-embed-text (or all-minilm)
• Desktop only (uses Node.js fs/child_process)

Models

Vizier routes work across four roles so you can match each task to an appropriate model. Defaults fall back to your Default model if a role is unset.

Role	Used for	Notes
Default	Chat, /write, /edit, agent turns	Your main model
Utility	Tagging, dedup, memory extraction, feed triage	Pick a fast model
Research	/thesis, /research, deep synthesis	Pick your largest/strongest model
Embedding	Semantic index + memory	An embedding model (nomic-embed-text, all-minilm)

Cloud vs local: models whose name ends in -cloud / :cloud are treated as Ollama cloud models — they use their full context window and are managed by the provider. Local models honor the Local model context window (num_ctx) setting so you can size context to your hardware.

Setup

Ollama

1. Install Ollama and start it: ollama serve
2. Pull a model: ollama pull gemma3:4b (or any model you prefer)
3. Set the model name in Settings → Vizier → Default model

Handwriting transcription

/handwriting transcribes a pasted image with the vision role model chain (default qwen2.5vl:7b, which is markedly stronger at handwriting than general-purpose models) — no server, no extra downloads. Tall pages are automatically sliced into overlapping high-resolution bands so each line stays readable to the model, and the per-band transcriptions are stitched back together. Before anything is saved you review the result side-by-side with the image — [illegible] marks the words the model wasn't sure about.

Optional OCR assist: for tricky pages you can enable Settings → Vizier → Handwriting OCR assist, which additionally runs a local EasyOCR engine and feeds its raw output to the vision model as a cross-check. This requires the Vizier server (below) and a one-time ~1.5 GB install (PyTorch + models); the toggle offers the install when the server is running.

YouTube transcripts, Wikipedia lookups & OCR assist

These features require a small local Python server (vizier_server.py) that runs alongside Obsidian.

One-click setup: Open the Command Palette and run "Vizier: Setup / start Vizier server". The modal will detect Python 3, create a virtual environment, install dependencies, and start the server automatically. If the server is already running from a previous session, clicking Start will detect this and confirm without restarting.

Auto-start: After the one-time setup, the server starts automatically with Obsidian and on demand whenever a feature needs it (transcripts, Wikipedia lookups, OCR assist), and stops when Obsidian closes. This can be turned off with Settings → Vizier → Auto-start Vizier server.

Manual setup (fallback):

cd <vault>/.obsidian/plugins/vizier
python3 -m venv .venv
.venv/bin/pip install youtube-transcript-api wikipedia-api
.venv/bin/python3 vizier_server.py

The server runs on http://127.0.0.1:11435. You can change this in settings.

Article fetching

Articles are fetched via Jina AI Reader (r.jina.ai), a free public service that returns clean markdown from any URL. No API key required. The only data sent is the URL you provide. If Jina is unavailable, Vizier falls back to a raw HTML fetch.

Settings

Setting	Default	Description
Ollama URL	http://localhost:11434	Base URL of your Ollama instance
Default model	gemma3:4b	Model used for all AI tasks
Vizier server URL	http://127.0.0.1:11435	URL of the local Vizier server
Clips folder	Clips	Vault folder where /clip saves notes
AI notes folder	(empty)	Vault folder where /write saves notes (empty = vault root)
Handwritten notes folder	Handwritten Notes	Vault folder where /handwriting saves notes
Handwriting OCR assist	off	Also run a local EasyOCR engine as a second signal for /handwriting (heavy optional install)
Vision role models	qwen2.5vl:7b, gemma4:e4b, gemma3:4b	Vision-capable model chain used by /handwriting
People folder	Human Network/People	Vault folder where /person saves notes
Events folder	Human Network/Events	Vault folder where /event saves notes
Ideas folder	Human Network/Ideas	Vault folder where /idea saves notes
Entities folder	Human Network/Entities	Vault folder where /entity saves notes (sub-foldered by type)
Timeline folders	Human Network/Events, Human Network/People, Human Network/Ideas	Comma-separated folders /timeline searches
Reflections folder	Reflections	Vault folder where /weekly, /monthly, and /freewrite save notes
Books folder	Books	Vault folder where /ingest saves processed book notes
Transcripts folder	Transcripts	Vault folder where /transcribe saves audio transcripts
Theses folder	Theses	Vault folder where /thesis saves structured documents
Whisper model	base	Model size for audio transcription (tiny / base / small / medium)
Max notes per recluster	100	Maximum notes /recluster will process (most recent by modification date)
Local model context window (num_ctx)	8192	Context size for local models; cloud models use their full window. 0 = use the model's max
Model roles	per-role	Ordered fallback chains for the default / utility / research / embedding roles
Feed URLs	(empty)	RSS/Atom feeds for /intake, one per line
Inbox / Contradictions / Skills folders	Vizier/...	Where intake, briefings, contradiction flags, and learned skills are written

Feature toggles: Agent loop, Memory, Daily intake, Daily briefing, Debug logging, and per-group Command groups (declutter the slash-command picker) — all in settings. Debug logging writes a detailed conversation-flow trace to vizier_debug.log for analysis (off by default).

Building from source

git clone https://github.com/AADaoud/vizier
cd vizier
npm install --legacy-peer-deps
npm run build   # production
npm run dev     # watch mode

Release artifacts: main.js, manifest.json, styles.css.

FAQ

---

Do I need an internet connection?

Only for /summarize//clip (fetches the article/transcript via Jina), /person//event//idea//entity (Wikipedia lookup), and /transcribe with a podcast URL. All AI inference runs fully offline via Ollama.

---

What model should I use?

gemma3:4b is a good default — fast, capable, and runs on most machines. For better quality on longer documents try gemma3:12b or mistral. Any Ollama-compatible model works.

---

The Vizier server won't start — what do I check?

Make sure Python 3 is installed (python3 --version). If setup fails mid-way, delete the .venv folder inside the plugin directory and run "Vizier: Setup / start Vizier server" again. The server auto-detects if it's already running, so clicking Start when the server is up will simply confirm it's available.

---

OCR results are inaccurate — can I improve them?

Results depend on image quality. Ensure the photo is well-lit with minimal skew. After OCR, Vizier automatically runs a cleanup pass through your Ollama model to fix spacing and punctuation.

---

Port 11435 is already in use.

Change the Vizier server URL in settings to use a different port (e.g. http://127.0.0.1:11436), then restart the server.

---

Privacy

• AI inference runs locally via Ollama by default — nothing leaves your machine.
• Exception — cloud models: if you select an Ollama *-cloud model for any role, the prompts and note content sent to that role are processed by Ollama's hosted service. Use only local models to stay fully offline.
• The only other outbound requests are to r.jina.ai when fetching articles (the URL you provide is sent), and en.wikipedia.org for Human Network lookups (only the name/title you provide). YouTube transcripts are fetched locally via youtube-transcript-api.
• Runtime data (chat traces, memories, the vault index, debug logs) is stored locally in the plugin folder and is gitignored — it is never committed or uploaded.
• No analytics, no telemetry, no cloud sync.