AI Chat

How it works

AI → AI Chat… opens the chat in a tab of the window, in the right-hand pane, that answers using your own notes as context. Instead of sending everything to the cloud, LightNote retrieves the most relevant snippets from your vault locally (via ripgrep, no embeddings and no external service), builds the question with those snippets and sends it to an API AI provider you have registered. The answer comes with the source notes as clickable citations — click to open the note. Great for "talking to your second brain" and finding what you already wrote down. Conversations are saved per folder, and you can go back to them (see below); the feature can be turned off in Settings → General → Features.

The tab and the window. The chat opens next to what you are reading, in a tab in the right-hand pane — one per window, and each window has its own conversation. In the tab, the notes are those of the window's folder, and the Current editor context follows the active tab (see below). The selectors sit at the top: the AI, the context and the Mode (what the AI may do), with the buttons to attach a file and to start a new conversation; in a narrow tab they take two rows. Before the first question the screen states what the chat will answer about and offers one-click shortcuts such as Summarize this folder. In the question box, Enter sends and Shift+Enter breaks the line; while the answer is coming, the send button turns into Stop. Closing the tab keeps the conversation in the list, like New conversation.

Move, detach and reattach. The tab can be dragged like any other — to the main pane, to the bottom one or back to the right — and when you reopen the folder it comes back to the pane where you left it. Detach to window, in Conversation actions or on the corner button, moves the same conversation — with its folder, attachments, Help section and what you allowed in it — to a separate window, and the tab closes. In the separate window, Reattach to window brings it back to the tab of the window it came from (or of another one with the same folder). A conversation is open in one place only, and it does not move while the AI is answering. The tray icon (AI → AI Chat…) opens the separate window, which has the Notes from: selector to choose the folder.

Who answered, and comparing answers. Under each answer, the assistant and the model that wrote it appear discreetly — useful when the conversation switched assistants or when the first model on the list failed. The chat remembers the last assistant chosen per window and folder. To send the same question to more than one, use Compare answers… in the chat menu or type /compare before the question: see Comparing answers.

Tools and permissions

The Mode: what the AI may do. The Mode selector, in the top bar, has three values. Notes only answers from the excerpts the keyword search retrieved, and nothing is run. Read lets the model look things up on its own — read a file, search notes, list tasks, consult this Help — before answering, without changing anything. Change with approval adds creating notes and tasks, changing properties and writing to today's note, always with your confirmation. If you never chose, it starts in Read, because reading is safe. At the end of the answer comes the list of tools used, so you can judge the answer instead of trusting it.

The model check. Read and Change with approval require an API provider (a local CLI has no way to receive tools). When you pick one, LightNote tests right away whether the model can call tools: many providers accept the request and simply ignore the tools, which without this check would turn into a made-up answer that looks researched. The result is remembered for seven days for each provider+model pair. If it fails, only Notes only stays available, and the Check again button, next to the message, reruns the test ignoring what was stored — the provider may just have been down at that moment.

Every change asks for your confirmation. The change appears in the conversation in plain words ("create the task Buy bread") with three answers: Allow, just this once; Allow in this conversation, which lets the same tool through without asking until the conversation ends (the conversation says so each time: Allowed by the conversation); and Deny, which does not end the conversation — the model is told and carries on, explaining or proposing something else. Deleting never gets that permission: every deletion asks. Reset this conversation's permissions, under Conversation actions, makes everything ask again, as does a new conversation or another folder. Overwriting arbitrary files and reaching the network are deliberately left out. When the change goes to the daily note, the request says which folder it will land in.

When the AI asks or stops. If the answer depends on something missing — which board, which of two notes — the AI can ask you: the question appears in the conversation with one option per link and, when it fits, an Other answer… field; Don't answer lets the model carry on without it. Once answered, the question turns into text, and the options of an old question do nothing. After many steps without finishing, the conversation stops and shows Continue, which gives the same job more steps; if the model repeats the same lookup three times without progress, it stops and says which one. The text of each step appears as the model writes it.

What it answers about

Notes or the open file. The Context: combo at the top of the chat chooses what the conversation is about: Vault notes (the default, described below) or Current editor — which answers about the file open in the editor instead of the vault, using the selection and the surrounding lines (or the whole file, with the Whole file checkbox). In the tab, Current editor follows the active tab of the window: switching tabs switches the file, the chat's own tab does not count, and when the active tab is not a text editor the item becomes unavailable without changing your choice. In the separate window it is the editor you opened the chat from. And in the quick assistant (Ctrl+K), the result has a Continue in chat → button that opens the chat already in the Current editor scope, seeded with that round’s question and answer — to dig deeper without starting over.

Which notes? In the tab, those of the window's folder: switching the window's folder switches the chat's, and the conversation in progress goes to the list. The separate window answers about one folder at a time, picked in the Notes from: combo, which lists the folders open right now, followed by the recent ones; Choose folder… points to any other — that is how you ask one project's notes while you work in another. Once picked by hand, that folder stays until you change it again. While a question is in progress — including while it waits for your confirmation — the AI, context and folder selectors are locked: the answer on its way belongs to the model and the folder you had when you asked.

How to ask. Retrieval is by subject, not by instruction: LightNote takes the words from your question and looks for notes that contain them. Name a topic that appears in the notes ("what did I note about Postgres?") instead of a generic request ("summarize this") — the latter has no content word to search for, so nothing is retrieved and the chat warns that the answer did not use your notes. Only .md/.markdown notes are read; in a folder without any, the chat says so and does not spend an AI call.

Better search, summary and rewrite. Retrieval uses a local index with BM25 ranking that ignores accents ("configuração" finds "configuracao") and also works in languages without spaces between words (Japanese, Chinese, Korean) — ripgrep is the fallback while the index is still being built. The Summarize this folder button — and Summarize with AI… in the context menu of any folder in the explorer — reads the notes in batches and returns a document. And in Settings → General → Features, the Rewrite the question with AI option expands your question into keywords before searching (it costs one API call per question).

Key and conversations

Just one key. On first run, LightNote already registers OpenAI, Google Gemini and Anthropic (Claude) under Settings → AI Assistants — you only need to paste your API key into the provider you use. If no provider has a key, the window shows a Configure now button that opens that screen directly.

Conversations are kept. The clock button at the top opens the list of previous conversations for that folder: click one to go back to it. New conversation archives the current one instead of deleting it, and right-clicking an entry deletes it for good (with confirmation). Conversations are stored per folder — switching the notes folder switches the list. Reopening the folder brings the chat tab back with the conversation that was open.

What to do with the answer

What to do with an answer. Under each answer there is Copy (puts the Markdown text on the clipboard, formatting intact) and Capture to note (sends the answer to today's note, along with links to the cited notes). The Conversation actions menu holds Copy conversation, Save conversation as note… (writes a .md in the queried folder), Regenerate last answer and Edit last question.

Attaching a file. When the search does not find what you had in mind, use the + button to attach a file: it always goes into the question, outside the search ranking. Notes, code, text and PDF all work (PDF text is extracted); a binary file is refused right away instead of becoming noise in the context. The attachment shows up as a pill (click it to detach) and is dropped when you start a new conversation. With providers that support it, the answer also appears gradually, as the model writes.

Assistant memory. In the Conversation actions menu (the ☰ button at the top), Assistant memory… opens — creating it the first time — a Memory.md note at the root of the selected folder. Whatever it holds goes into every question asked about that folder, ahead of the excerpts found: your preferences, the decisions you have already made, what you ruled out and why. It is an ordinary .md — edit it whenever you like, version it with Git, delete it to turn memory off. With write tools enabled, the assistant can also propose storing a fact there, and like every write it waits for your approval.

Summarizing a folder. It works for the whole workspace or for any subfolder: right-click it in the explorer and choose Summarize with AI…. Before spending anything, LightNote tells you how many notes will be read and how many AI calls that will make — each one billed by your provider — and lets you pick the shape: Overview, flowing text, or Consolidate knowledge, with fixed sections (decisions, open items, ideas, contradictions and open questions; an empty section is omitted rather than filled with invention). If the folder holds more notes than the cost cap allows, the answer says how many were left out. At the end, Save as note writes the summary inside the folder it summarized, with the date and the note count in the header, and never overwrites an earlier one.

Commands, Help and links

The assistant knows LightNote. Besides your notes, it knows it is running inside LightNote: it gets the index of this Help, a summary of the app's Markdown (wikilinks, callouts, diagrams, formulas, calendars) and, with every question, the folder being searched, the active tab and the date. Ask "how do I create a task board?" the same way you would ask about a note.

Commands with /. Typing / at the start of the empty field opens the command list, which filters as you type; Enter picks and Esc closes. /help followed by a topic (/help tables) attaches to the conversation the section of this Help whose title matches the topic, with or without accents; without a topic, you get the section for the screen you opened the chat from. With a question instead of a topic (/help how do I create a table), the list shows the sections that may answer it: the one you pick is attached and the question goes along.

The attached section. It shows up as a Help: … pill next to the attachments, goes into every question before the note excerpts, and leaves with a click on it or with a new conversation. On the empty screen, How does the chat work? does this with this section and asks right away.

The other commands. /summarize summarizes the folder and /new starts a new conversation. The prompts in the Prompts folder appear in the same list: picking one puts its text in the field, so you can review it before sending.

Notes with [[ or @. Typing [[ or @ in the field opens the list of the folder's notes, which filters as you type, with or without accents. Enter picks one, and the note comes in as an attachment — the same pill as the + button, reread on every question and outside the search ranking —; what you typed to find it leaves the field. Esc, Space or a click elsewhere close the list without attaching anything.

Help and Markdown as tools. In the Read and Change with approval modes, the model looks up this Help on its own (help_lookup) and the reference of the syntax LightNote renders (markdown_reference) — that is how it gets the menu path and the fence of a diagram right — and, when something is missing, it asks you (ask_user). The first two are also in the MCP server, for AI CLIs; the question is chat-only, because it needs someone in front of the screen.

Links to screens. When the answer needs a screen of the app, it carries a link instead of describing the menu: lightnote://settings/ai opens Settings on the AI assistants section, lightnote://help/… a section of this Help, lightnote://wizard/clone the clone wizard, lightnote://open?file=… a file in the folder and lightnote://new?type=lnt creates a new file. The same link works inside a note.

Nothing opens by itself. Your click is your consent. A link to outside the folder, to a feature that is turned off or to a target that does not exist opens nothing, and the status line says why.

Setting up the app and creating connections

Changing a setting. In the Change with approval Mode, ask as you would ask a person: "switch to the dark theme", "make the editor font 14", "turn off the minimap". The request shows up as before → after ("Theme: Light → Dark"), with a link to the section of the Settings where the same option lives, and it only takes effect after Allow; then the conversation says Applied with the new value.

Every change asks again. Allow in this conversation does not apply to settings. Interpreters and AI CLIs can be added, changed or removed the same way, always naming the program, never a folder. A shell (pwsh, bash) is accepted only as an interpreter that opens the file, as in pwsh -File {file}, never with a command written on the line itself.

What stays out. Only a closed list of options can be reached: theme, language, editor font, how Markdown opens, features, daily notes and the AI options, among others. API keys, provider addresses, program paths and terminals never change through the chat: the assistant answers with a link to the right section. This is on purpose — the text of your notes goes into the conversation, and a note that came from elsewhere must not be able to point your AI at another address.

Creating an HTTP test or a connection. Ask "create a test for GET https://…", "save the SSH connection to deploy@server" or "connect to the sales database": after Allow, the assistant creates the .lnh, .lns or .lnc in the same format the app writes, and the conversation shows Created with the link that opens the file. It never overwrites a file, only creates inside the folder, and does not create the type of a feature turned off in the Features section of the Settings.

A detail is missing? It asks. When a detail that changes the result is missing — the database, the server, the authentication type —, it asks with the options before creating.

No password goes through the chat. Database and SSH connections are created without a password: it is asked when you connect, and AI access to the database connection stays off until you allow it. In an HTTP test, the token or password becomes an empty variable ({token}) that you fill in the test or in http-client.private.env.json; a secret written out in full is refused. The three creators are also on the MCP server, for AI CLIs; changing settings is chat-only.