Ratel
approvedby golddream-y
Graph-native AI agent for your notes. Chat with cited search, run multi-step research, keep memory and skills, edit under permission. DeepSeek, Claude, or Ollama. Local embeddings, no telemetry. - This plugin has not been manually reviewed by Obsidian staff.
Ratel
Make your Obsidian knowledge work proactively.
Ratel is a proactive graph knowledge agent. It understands your journals, goals, and linked notes, then brings you sourced reviews, timely suggestions, and knowledge issues worth acting on—inside Obsidian and under your control.
From waiting for prompts to knowing when to help
A typical vault chat starts only after you ask. Ratel is designed to let your knowledge participate at the right time:
- Heartbeat checks new journals, goals, recent changes, and knowledge relationships.
- Local rules decide whether there is anything worth interrupting you for.
- Ratel creates a sourced yesterday review, today suggestion, or knowledge-maintenance insight.
- Every result enters the insight inbox; high-value items can trigger an Obsidian Notice.
- You can inspect sources, ask a follow-up, snooze, ignore, or approve a write to your vault.
Proactive does not mean autonomous modification. Notifications may appear automatically; vault changes remain your decision.
How Ratel makes your knowledge work
Bring you information worth acting on, at the right time (in progress)
Ratel does not require a second task database. Tell it where your journals, monthly goals, and recurring work live through ordinary Markdown memory.
- Start the day with a sourced review of yesterday.
- Add today suggestions when both monthly goals and today's journal are available.
- Ask once when a familiar location is missing; remember “none” instead of asking forever.
- Keep every proactive result in a reviewable insight inbox.
- Use the status bar as a stable entry point and reserve Notices for high-value insights.
- Respect quiet hours, daily limits, snooze, ignore, and follow-up actions.
Understand relationships, not just files
Ratel combines retrieval, links, citations, maintenance, and knowledge discovery into one workflow.
- Fusion retrieval: semantic and keyword recall, enriched by links, backlinks, and properties.
- Clickable citations: numbered sources open the original note in one click.
- Knowledge entropy management (in progress): surface broken links, orphan notes, duplication, stale knowledge, and lingering tasks.
- Knowledge mining (in progress): discover themes, relationships, conflicts, and gaps across notes, then produce sourced syntheses.
Move from discovery to controlled action
Inspect sources, ask a follow-up, snooze, ignore, or let Ratel organize, link, and write notes. Every vault change continues to follow your permission settings.
What works today
Recover knowledge when you no longer remember where it lives
“What did I write about performance tuning?”
Ratel combines semantic, keyword, and note-relationship signals, then keeps clickable sources in the answer.
Keep a long-running goal across chats
“/goal finish the Q3 reading notes this week”
The goal stays with the vault, not a single conversation. Chat confirms the done criteria before creating it; the status bar and a strip above the input remind you, and a new chat can take it over.
Complete multi-step knowledge work
“Turn my product-planning notes into a background document.”
The agent can search, read, synthesize, and write. Changes follow the permission level you choose.
Understand the current Obsidian context
“Summarize this note and relate it to recently edited project notes.”
The active note, daily-note location, recent changes, and outlines are available as first-class context.
Open the note you are talking about
“Open that reading note and jump to its second chapter.”
Ask “open that note” and the agent opens it in Obsidian, jumping straight to the heading (or block) you mean.
Configure and troubleshoot by chat
“Switch my chat model.” / “Why isn’t indexing running?”
The built-in config skill reads current settings, applies whitelisted changes, and walks you through keychain setup — keys are guided into the keychain, never filled in for you.
Remember preferences and extend workflows
“Remember that I prefer conclusions before evidence.”
Memory stays as Markdown under .ratel/memory/ — pin must-keep preferences with [pinned], and the most related topic memories join each turn automatically. Add a SKILL.md, connect an MCP server, or use Subagents for more specialized workflows.
Your agent, your stack
- Model freedom: choose chat, embedding, and reranking models independently; use DeepSeek, Claude, Ollama, or compatible custom endpoints.
- Vision: attach images in chat.
- Layered memory: pinned preferences never get truncated; related topic memories join each turn automatically.
- Skills: define reusable working methods in Markdown, with optional sandboxed scripts and reference files the agent can run or read.
- MCP: connect web search and external tools with per-server and per-tool permissions.
- Subagents: delegate complex research to focused retrieval, review, and synthesis roles.
- Prompt customization: override individual prompt sections without forking the plugin.
These capabilities are the foundation. They serve proactive intelligence and graph knowledge management instead of becoming setup work for its own sake.
Privacy and safety
| Data or action | Default behavior |
|---|---|
| Vault index | Stored locally |
| Embeddings | Generated locally by default |
| Retrieved evidence | Sent only to the model endpoint you configure |
| MCP parameters | Sent only when an enabled MCP tool is invoked |
| Proactive insights | Filtered locally before the minimum necessary evidence is assembled |
| Vault changes | Follow Safe / Auto / Danger permissions; notifications cannot bypass them |
| Telemetry | None |
Once blacklist controls land, excluded content will never enter the index, candidate detection, model context, MCP parameters, logs, or notifications.
Install
Obsidian → Settings → Community plugins → Browse → search Ratel → Install → Enable.
Requires Obsidian 1.13.0+, desktop only.
Then open Settings → Ratel → Chat model, choose a DeepSeek / Ollama scene preset or custom Base, wait for the first index, and click the 🦡 ribbon or run Ratel: Ask vault.
See the User Guide for the full walkthrough.
Docs
| Doc | Contents |
|---|---|
| Product Vision | Positioning, complete product picture, and direction |
| User Guide | Setup, scenarios, slash commands, and FAQ |
| Changelog | Release history |
| Architecture | Ports, agent loop, tools, and workers |
Issues and ideas: GitHub Issues.
License
For plugin developers
Search results and similarity scores are powered by semantic analysis of your plugin's README. If your plugin isn't appearing for searches you'd expect, try updating your README to clearly describe your plugin's purpose, features, and use cases.