AI Setup & Privacy

1. Install and start your preferred OpenAI-compatible LM Studio or Ollama server and make a model available.
2. In RelicNotes, open Settings → AI & Search → AI Chat.
3. Choose Local server, then enter the endpoint and model information used by your local provider.
4. Test the connection and save a working configuration. AI Chat remains in its setup state until a model is configured.

Local AI can work offline when the configured provider and model are available on your computer. Your local server software may have its own logs, resource requirements, and privacy behavior.

1. Create or use your own OpenRouter account and API key, then review the provider’s model prices and data policies.
2. Open Settings → AI & Search → AI Chat and choose OpenRouter (cloud).
3. Enter your API key, choose a model, and test the connection. Review the cloud privacy warning and one-time consent.
4. Keep auto-apply AI write operations off if you want to confirm every proposed note change.

1. Run Pet: Chat with Pet (AI Chat) from a command palette. The pet branding is an optional persona, not a separate AI feature.
2. Choose Add note to pin exact notes as context, or enable auto-find relevant notes for semantic retrieval with fuzzy fallback.
3. Ask a question. With OpenRouter, the prompt can include your message, relevant chat history, and selected or automatically retrieved note content.
4. To change content, explicitly ask for an edit. Review the resulting Confirm/Reject card before applying a write operation.

AI conversations are stored in the local SQLite database. Graph Mode lets you branch a conversation, and a branch can be exported to a note as a transcript or AI summary.

• Open a note and run Note: Convert to Flashcards with AI, or use the relevant note/block generation action. Selected source content is sent through the configured OpenRouter service; preview and edit cards before accepting them.
• Optional AI judging for typed answers sends the question, expected answer, and your answer to OpenRouter for evaluation.
• Disable AI judging if you prefer local/simple validation. Existing cards and ordinary review work offline; AI generation and judging require the relevant online service.

Open Settings → AI Models to download and manage supported models, inspect loaded state, open or change the model folder, and unload models.

• Model downloads require internet; supported on-device inference remains local afterward.
• Changing the model folder moves existing model subfolders so they do not need to be downloaded again just because the location changed.
• Enable Unload AI models when idle or choose Unload now to free RAM. Unloading does not delete model files, but the next operation may need to reload the model.

• Local notes and on-device processing stay on your local path unless you explicitly share content or choose a cloud operation.
• OpenRouter sends the content required for the requested AI operation to OpenRouter and the selected model provider.
• YouTube embeds and title lookups, external link previews, model downloads, account/community features, and update checks can contact external services.
• A masked password block can still be included if you deliberately send its containing note to a cloud provider or publish it. Remove secrets first.

See the Privacy Policy for the full explanation of local data, online services, and third-party processing.

• For local AI, confirm that the server is running, the model is available, and the endpoint and model settings still match.
• For OpenRouter, check internet access, API-key validity, model availability, and your provider balance or limits.
• For on-device models, confirm that the download completed, the model folder is accessible, and enough free RAM is available.
• Use Stop to end an active request and keep any partial reply. Requests can time out after 120 seconds without an initial response or 60 seconds when streaming stalls.
• If raw edit-note or note-op text appears instead of a valid change, the model may have formatted the operation incorrectly. Retry with a clearer request or a more capable model.

Continue with Shortcuts & Troubleshooting or contact Support if the problem persists.