Ask Compass

Ask Compass

Route: /ask · Sidebar: Ask Compass · ⌘K: "Ask Compass"

A retrieval-augmented (RAG) assistant that answers questions grounded in your own knowledge base. It finds the relevant note snippets, hands them to a model, and cites the source notes in its answer — so you can trust where the answer came from.

How to use it

• Type a question in the box — "Ask anything about your notes… (⌘↵ to send)". Press ⌘↵ to send.
• The answer streams back with a Sources list of the notes it drew from.
• Each answer footnotes which provider · model produced it.

Agent mode

Next to the input, an Agent toggle switches Ask Compass from plain retrieval-and-answer into a bounded, Anthropic-only tool-use loop: instead of just reading pre-fetched note snippets, the model calls tools to read your agenda and finance summaries directly, and can propose changes — which land in the Claude Inbox for you to review and approve, the same approval mechanism external Claude/MCP proposals use.

• Agent mode is only available when Anthropic is your active provider (Settings → AI assist) — the toggle is disabled otherwise.
• Vault entries and task/finance detail stay excluded even in Agent mode; only summaries are exposed to the tool loop.
• Plain Ask (Agent off) stays read-only: it composes local semantic search results into a prompt and returns a single grounded answer with no tool calls and no proposals.

Bring your own key (local-first)

Ask Compass is opt-in and BYO-key. Configure it from this page or from Settings → AI assist:

• Local Ollama is preferred — fully offline, no key, default endpoint http://localhost:11434.
• Anthropic or OpenAI — paste your own API key. The key is encrypted at .vault/assistant.enc and never crosses the IPC boundary again after you set it; the renderer only ever sees a masked tail. Requests are made from the main process (so CSP doesn't block them).

You pick the active provider and model.

What it can and can't see

• Can: your Knowledge Base notes (via local semantic search — top-K cosine over the embeddings index — with a keyword-scan fallback if no index exists yet).
• Cannot: your Vault. Vault secrets are categorically excluded from every AI surface.

How retrieval works

assistant:ask composes the local semantic search (Phase 5.9 embeddings) to pull the most relevant snippets, builds a grounded prompt, and issues a single non-streaming request to your active provider — this is the plain (Agent off) path.

assistant:agent is the second, agentic path behind the Agent toggle above: a bounded Anthropic tool-use loop where the model reads your data via tools (rather than a single pre-built prompt) and can propose writes that route through the Claude Inbox for approval.

Ask Compass vs. Claude & MCP

• Ask Compass (plain) = Claude/OpenAI/Ollama embedded inside Compass, reading your notes, read-only.
• Ask Compass (Agent mode) = the same embedded surface, but Anthropic-only and tool-using — it can also propose changes, routed through the Claude Inbox for approval.
• Claude & MCP = external Claude (Claude Code today) reading Compass over a read-only connector and proposing changes you approve.

The clean line used to be "Ask Compass never writes, only external Claude/MCP proposes changes." With Agent mode shipped, that's no longer strictly true — the distinction now is plain Ask (always read-only) vs. Agent mode or external Claude/MCP (both can propose writes, both land in the same Inbox for your approval).

Related

• Knowledge Base · Settings · Security & Privacy