Frequently asked questions

The questions that come up most. If yours isn't here, tell us — we'd rather hear it once than have ten people stub their toes.

• +−Which LLM does OpenMind use?

  By default Anthropic's Claude Sonnet 5, picked for the extraction quality on conversational German + English. You can switch to OpenAI via Settings → LLM provider (the choice is saved in your browser, not per project); Ollama and LM Studio (local) need a self-hosted instance. Bring your own API key in Settings → LLM provider; the hosted instance needs one before it will send a message.
• +−Where does my data live?

  On the hosted demo: a Supabase Postgres database in the EU (Stockholm, eu-north-1), row-level-security-locked to your account; an encrypted weekly backup of that database is kept for 14 days on GitHub’s storage in the US. On self-host: wherever you put it (your laptop, your server, your company's network). See /help/privacy for the full breakdown — what we store, who can read it, how to export everything, how to delete your account.
• +−Can I bring my own API key?

  Yes — and on the hosted service you need one: nothing is sent to a model without it. Settings → LLM provider → Your API Keys (Anthropic or OpenAI). The key is encrypted before it's stored on the server and never sent back to your browser; afterwards you only see its first characters. Your chat replies, extraction and Ask answers then run on it, billed directly by the provider, with no markup. Embeddings run on your own key as well, and that takes an OpenAI key (Anthropic has no embedding model): without one, your nodes are saved without embeddings, search finds them by name only and Ask starts from the nodes that contain your question’s words.
• +−What attachments can I drop into the chat?

  TXT, MD, PDF, DOCX are extracted in-process. Audio and video (MP4, MOV, MKV, WebM, MP3, WAV) are queued to a worker that runs faster-whisper for transcription — available on self-host; the hosted demo rejects audio and video uploads with HTTP 422, because it runs without the transcription worker. Drag-drop or paperclip into the composer.
• +−What does a node's confidence mean?

  The extractor gives each node a confidence when it creates it. To see it, search for the node by name in the mindmap pane and click the result to open its detail panel. Anything below 0.6 is handed to Ask-KG tagged [uncertain] — usually a sign the model wasn't sure how to canonicalise an entity (an ambiguous pronoun, a tentative decision, a half-described risk). Mentioning the node again in chat doesn't change its confidence; if a node is wrong, delete it from the search results and describe it again.
• +−Can I share a project with someone else?

  Yes — kebab on a project card → Transfer ownership. Currently the model is single-owner; the new owner gets full control and you stay as an editor. A multi-member share without ownership transfer is on the roadmap (workstream-ui-completeness).
• +−How do exports work?

  Every project has an "Export ▾" dropdown in the workspace header with six formats: Markdown, Mermaid, OPML, FreeMind, PNG, SVG. /help/exports breaks down when to use which. Mermaid is canonical — PNG and SVG render from it, so whenever an image export is allowed, what you see in Mermaid is what you get as an image. It isn’t allowed for a map of 1,500 nodes or more, nor for one nested several hundred levels deep.
• +−Is there an API I can call?

  Self-hosted deployments include one: `/api/public/v1/...`, read-only (projects, nodes, edges, exports, and Ask), gated by personal access tokens minted in Settings → API keys. The hosted service does not offer it at the moment.
• +−Can I run OpenMind on my own hardware?

  Yes — the full docker-compose stack ships at the repo root: `docker compose up -d --build` on your own machine gives you a working instance (web + API + Postgres + render sidecar; `ATTACHMENTS_VIDEO_ENABLED=true` with `--profile worker` adds audio and video). In our test the running stack used about 1 GB of memory, and 0.2 GB more while it rendered an image; building the images the first time needs more than that. Pair it with a local model of your choice through Ollama or LM Studio, and no conversation, document or model call leaves your network; the model needs memory of its own, and the graph is only as good as the model that extracts it. The repository is private for now: request an invite, say you want to self-host, and access is granted by hand; the self-host guide then walks you through the five steps.
• +−I clicked Delete account — is it really gone?

  First the files uploaded to your projects and your profile picture are removed from file storage; if one cannot be removed, your account is not deleted and you can try again. Then your account and everything that cascades from it in the database are deleted immediately: your profile, your stored keys, and every project you own with its conversations, messages, nodes, edges, attachment records and embeddings. There is no soft-delete grace period. Not removed: the messages you sent in projects other people own, and the files attached to them, stay in those projects (a file you uploaded there but never sent is removed), your email stays on any waitlist or invitation record, the token counts recorded for your chat replies and extractions stay with their link to you, the project and the message cleared, and on the hosted service the encrypted weekly backup keeps a copy for up to 14 days. Type-the-email-to-confirm gates the action; once submitted, it's irreversible. Use Settings → Account → Download my data first if you want a backup.
• +−If I paste a whole article, will OpenMind read all of it?

  Yes, up to 32,000 characters per message. Long pastes (over ~2 000 tokens) are split into segments behind the scenes; the extractor walks every segment in sequence, threading earlier-segment nodes into the next so the same entity isn't duplicated across them. The success badge under your message shows 'from N segments' once it's done so you can confirm everything was processed.
• +−Can I edit or delete a node directly?

  All of these start in the mindmap pane’s search panel. Search finds nodes by name; with your own OpenAI key on the hosted service it also lists nodes related by meaning (an Anthropic key alone doesn’t cover that part). Delete: yes — search for the node, hover the result row, click the trash icon. Edges and attributes touching the node disappear automatically (FK cascade). Re-mention the corrected concept in chat and the LLM creates a fresh node. Recategorize (change the kind): yes — multi-select rows in the search panel and pick a new kind from the toolbar. The label, summary, and edges stay intact; only the bin changes, so the chat history and graph stay consistent. Merge duplicates: yes — select the rows and use Merge; they fold into one node and their edges move with it. Edit (label / summary / confidence): no end-user affordance, intentionally — it would drift the graph from the chat history. See ADR 0010 for the trade-offs.
• +−What are embeddings?

  Number vectors computed from each node’s text by an embedding model. They power the by-meaning half of the mindmap pane’s search, choose which nodes Ask starts from, and are what suggested connections compare. On the hosted service they are made on your own OpenAI key (Anthropic has no embedding model): during extraction, for each search by meaning and each Ask question, and by the “Embed” button, which adds them to every node of the project that has none, including, in a shared project, the nodes other members created, whose text is then sent on your key. Without an OpenAI key your nodes are saved without embeddings, search finds nodes by name only and Ask starts from the nodes whose name or summary contain your question’s words instead. Once you add a key, “Embed” fills in the nodes that have none.
• +−What are suggested connections?

  The mindmap pane can suggest edges between nodes whose embeddings are similar but that aren’t linked yet. It is a similarity measure, not a trained model, and nothing learns from your choices: accept adds the edge, reject keeps that pair from being suggested again. Two suggested nodes with the same label are more likely a duplicate — merge them from the search panel instead.

Still stuck? Reply to your invite mail or use the waitlist form — a human reads both.