AI assistant

Noted's AI features work with a provider you choose and configure. Nothing is sent anywhere until you set one up, and by default requests to cloud providers are masked for personal data first (see PII masking).

Providers

Configure your provider in Settings → AI.

Cloud services

• OpenAI, Anthropic (Claude), Google Gemini and OpenRouter
• Groq, Mistral AI, DeepSeek, xAI (Grok), Together AI, Fireworks AI, Cerebras, Hugging Face and Perplexity

On this computer

• LM Studio (default http://localhost:1234/v1) and Ollama (default http://localhost:11434)
• Unsloth Studio (http://localhost:8888/v1), llama.cpp server (http://localhost:8080/v1), vLLM (http://localhost:8000/v1) and Jan (http://localhost:1337/v1)

Other

• OpenAI-compatible (your own address): any service or server that speaks the OpenAI chat protocol, at the address you give.

Every service in the lists above except the first four, LM Studio and Ollama is an address on the OpenAI protocol: choosing one fills in its address and you can change it (a different port, a server on another machine). Unsloth Studio asks for the key you create in its Settings → API; put it in API key.

The list names no model, because models are replaced faster than any list: after you enter your key press Detect models and pick from what the service offers today, or type a model name. A model is never assumed: with none chosen, a cloud service says so instead of using a name that may have been retired. For a server on this computer the models are listed as soon as you choose it, and leaving the model blank picks the first one (LM Studio and Ollama). Keys are stored with the macOS Keychain via safeStorage when available.

A provider counts as configured when a cloud provider has an API key, or a local provider has a model available. Until then, the assistant shows a setup prompt instead of running.

No streaming

Responses are returned whole rather than token-by-token. Requests time out after 60 seconds and retry with backoff on transient errors; failures surface as clear, localized messages (bad key, rate limit, model not found, provider unreachable, and so on).

Inline suggestions (ghost text)

As you write, Noted can propose a faint continuation after your cursor. Press Tab to accept it, Esc to dismiss it, or just keep typing to clear it.

Choose the behavior in Settings → Editor → AI suggestions:

• Off
• Manual — trigger a suggestion on demand with ⌘L (the default)
• Auto while typing — suggestions appear on their own after a brief pause

Slash commands

Type / at the start of a block (or after a space) to open the AI command menu:

Command	What it does
Continue	Continues the text for a few sentences
Expand	Expands and deepens the surrounding text
Summarize	Summarizes the whole note into a few bullets
Improve	Improves clarity and flow, keeping your meaning
Bullets	Rewrites the text as a bullet list
Translate	Translates between Italian and English (other languages to English)

Navigate with the arrow keys, run with Enter or Tab, and press Stop to abort a running command.

AI actions bar

Enable the actions bar with Settings → Editor → Show AI bar (off by default). It sits under the toolbar with two groups:

• Transform — Continue (appends), and Expand, Shorten, Refine (rewrite the selected text).
• Analysis — Summarize, Review, Devil's advocate, and Q&A, each appended below a divider.

Rewrite actions require a text selection, so they can never silently overwrite a whole note; a badge shows when a selection is active. Any running action turns into a Stop button.

Reviewing a rewrite

A rewrite (Expand, Shorten, Refine, Translate, Tone, Bullets, or a custom instruction on a selection) does not replace anything at once. When the model answers, a review opens: the proposal compared with your selection, line by line (a paragraph is a line), with the words that changed marked. Each run of changed lines is one change that you Keep or Drop on its own; Keep all and Drop all do it for every change. Apply puts in place of the selection your text with the kept changes made; Discard leaves the note exactly as it was. Nothing in the note moves while you read.

If the selected text changed while the model was working (you kept typing in it), nothing is replaced and you are told, rather than putting the rewrite over something else. Applying is one undo step.

Two limits. The selection is replaced as a whole, as rewrites always were, so inline formatting inside it (bold, links) is not kept even for the parts you keep: select just the paragraphs you want rewritten. And the grain is the line: a long paragraph that the model changed in a few words is one change, with the words marked. The review can be turned off with Settings → AI → Review AI edits before applying (the rewrite then replaces the selection at once, as before).

Chat

Open the AI Assistant tab in the right panel to chat with your notes as context. The chat sends the active note (trimmed to a configurable size) plus the most relevant notes from your vault, retrieved automatically. It keeps the last ten turns and can be cleared at any time.

Answers appear as the model writes them, for every provider (OpenAI, Anthropic, Gemini, OpenRouter, LM Studio, Ollama and OpenAI-compatible servers). Stop cancels the request at the provider, which stops generating, and keeps the part already written. If a provider fails half way, the part that arrived stays and the error is shown after it. A stream that sends nothing for a minute is cut.

Retrieval (RAG)

When you send a question, the assistant retrieves the most relevant sections from your whole vault rather than sending everything — however many notes you have, and however old the one you need is. A section is the text under one heading (long ones are cut at paragraph ends, at about 1,500 characters), so a long note contributes the part that matters instead of its first lines, and the assistant sees where each piece comes from (Note › Heading › Subheading).

Two rankings are merged:

• Words. The app's search index (BM25, kept current as notes change, including edits made outside Noted) picks the best notes, and their sections are ranked against your question. Fast, fully local, and always on.
• Meaning (optional, Settings → Integrations → Embeddings, Beta). Every section of the vault has a vector made by OpenAI, LM Studio or Ollama; the question is embedded too, and the most similar sections are found even when they share no word with it ("car maintenance" finds a section about the oil change).

The two lists are combined by reciprocal rank fusion, which needs no tuning between them: a section that both like comes first. If the question cannot be embedded (provider down, no vectors yet) the answer uses the words alone.

Sources and citations

What the assistant is shown is numbered: the open note is source 1 and the sections found for your question follow. It is asked to put the number in brackets after each statement that comes from a source, like [2]. In the chat those markers become links, and the sources an answer cites are listed under it (Note › Heading). Click a marker or a source: the note opens at that section and the passage is marked for a few seconds (the mark is not a selection, so nothing you type can replace it, and it does not change the file). If the note has changed and the passage is gone, you land on the heading; if the heading is gone too, the note just opens.

Numbers the model invents (a [9] with only four sources) are removed from the answer, and citations are only as good as the model: a source it cites supports the answer only as far as you check. Earlier answers are sent back without their markers, since the numbers belonged to that turn's sources.

Vault only

The book button in the chat header (and Settings → AI → Vault only) makes the assistant answer from your notes and nothing else:

• the model is told to answer only from the numbered sources, and to say so, in a sentence, when they do not contain the answer;
• sections that merely resemble the question are not offered unless they hold up (they contain at least half of the question's significant words, or are close in meaning; the similarity cut-off of 0.3 is a heuristic, and models differ);
• when there is no open note and nothing in the vault bears on the question, the model is not asked at all: you get "I couldn't find anything about this in your notes";
• an answer that cites no source is marked "No source cited: this answer is not backed by your notes".

This reduces answers from the model's memory, it does not prove an answer is right: a model can cite a source for something it does not say.

The index

Vectors are kept in .noted/embeddings/ inside the vault, one file per model, and survive restarts. A section is identified by its text, so only what changed is embedded again: editing one paragraph costs one section, moving a note to another folder costs nothing, renaming a note re-embeds its sections (the title is part of what is embedded). Sections are sent to the provider in batches, in the background, a few seconds after notes change; Settings → Integrations shows how far it has got and can rebuild it from scratch. For OpenAI, personal data is masked before text is sent (see PII masking). The .noted/ folder is never synced by Git.

The index of one model may use up to 512 MB of memory (about 100,000 sections of 1,536 dimensions); a vault beyond that keeps its first sections embedded and is searched by words for the rest, and the settings say so.

Retrieval is tunable in Settings → AI: how many notes to send (Top-K, default 3; the chat sends twice as many sections), how many candidate notes the word ranking looks into (5–100, default 30), and how much of the active note to include (default 8000 characters). A debug toggle shows, for each section, its place in each ranking.

PII masking

Masking is on by default. Before any request goes to a cloud provider, Noted replaces detected personal data — emails, phone numbers, card numbers, and similar patterns — with typed placeholders such as [EMAIL_1]. Requests to local providers (LM Studio, Ollama and any OpenAI-compatible server at localhost) are sent verbatim, since they never leave your machine.

The placeholders are turned back into the original values on your machine, in the answer, as it arrives: what the model says about [EMAIL_1] reads as the address, in the chat and in the text that an action or a slash command writes into your note. A placeholder means the same value for the whole conversation.

You can toggle it under Settings → Editor → PII masking. The chat panel shows a shield indicator and how many items were masked.

A hint, not a guarantee

PII masking is a best-effort filter over common patterns, not a security boundary. Review sensitive content before sending it to a cloud provider.

Smart tag suggestions

Enable smart tags (Settings → AI, off by default) to have Noted suggest a few tags after a substantial edit. Suggestions appear as a dismissible chip picker; accepted tags are appended to the note.

Next steps

• Git integration — version your vault and open pull requests.