AI
Nether’s AI uses a provider you choose, with your own key: any OpenAI-compatible endpoint, such as OpenAI itself or a model on your own computer through Ollama, vLLM or LM Studio, or Anthropic. Nether ships no provider of its own and sends nothing anywhere until you ask. Until a provider is set up, nothing of AI shows in the app.
In the desktop app, providers are your computer’s, and the app sends to them itself. In the web app, they are kept on your Nether server, yours or ones its admin sets for everyone, and your requests go through it.
Setting up a provider
Section titled “Setting up a provider”Open the settings from the Settings menu and choose the AI tab, then Add a provider:
- Name: what you call it, such as “Ollama”.
- API: OpenAI-compatible for OpenAI, Ollama, vLLM, LM Studio and most gateways, or Anthropic.
- Address: the API’s address. For OpenAI,
https://api.openai.com/v1; for Ollama on this computer,http://localhost:11434/v1; for Anthropic,https://api.anthropic.com. - Model: the model, as the provider names it, such as
llama3.2orclaude-sonnet-5-5. Nether lists none: type the one you use. - Key: the provider’s key. It is kept in your system’s keychain (the Keychain on macOS, the Credential Manager on Windows, the Secret Service on Linux), never in a file or a vault. Leave it empty for a local model that needs none.
The first provider you set up is the default; Make default makes another one so. Edit changes a provider, and keeps its key unless you type another or choose Remove the key. Providers are this computer’s: another computer of yours has its own.
In the web app
Section titled “In the web app”The web app’s AI tab lists the providers your server keeps: those its admins set for everyone, marked Everyone’s, and your own. Add your own with Add a provider, as above: its key is kept on the server, sealed with the server’s secret key, and never shown again. Admins choose under For whether a new provider is theirs or everyone’s.
Your requests go to the server, which adds the provider’s key and sends them on, and writing assist says so: “Sends the text through nether.example.org to Team model, llama3.2 at ollama:11434.” The server:
- reaches your own providers at public addresses over
https://only; - reaches the providers its admin sets for everyone on its own network too,
where the admin allows it with
NETHER_AI_PRIVATE; - keeps no keys until its admin sets
NETHER_SECRET_KEY, though a provider that needs none still works; - sends no AI for a workspace whose owners turned it off, or for an end-to-end encrypted one, whose text the server must never read: the desktop app offers AI there, with a provider you choose.
The admin’s settings are in the self-hosting guide.
Choosing one for a vault
Section titled “Choosing one for a vault”AI in this vault chooses what the vault uses: the default provider, another one, or Off in this vault.
A vault that syncs with an end-to-end encrypted workspace uses no provider until you choose one for it here, since its text is kept from everyone but its members. A provider on this computer, such as Ollama, keeps the text on it.
Turning AI off for a workspace
Section titled “Turning AI off for a workspace”A workspace’s owners and admins can turn AI off for everyone in it: in the AI tab of a vault that syncs with it, clear Members’ apps may offer AI in the workspace. Members’ apps then show nothing of AI in its vaults, and ask the server again before every request, so the change holds from the next one. Others see that AI is off, and why.
The switch tells Nether’s apps not to send the workspace’s text to AI; it cannot stop someone from copying notes elsewhere.
When your server approves the endpoints
Section titled “When your server approves the endpoints”Your server’s admin may lock AI to the endpoints the organization
approves, as its own model at https://llm.example.com/v1. Then, for every
workspace on that server, Nether’s apps send AI requests only to a
provider at one of those addresses, or under one:
- the AI tab says so: “nether.example.org allows AI only through https://llm.example.com/v1.”;
- a provider elsewhere is marked Not allowed by the server, and a vault that syncs with it does not use it; with the default elsewhere, the first provider that is allowed stands in;
- search by meaning waits while its provider is elsewhere;
- the apps ask the server again before every request, so a change holds from the next one;
- in the web app, the server refuses to keep or use a provider elsewhere.
A vault that syncs with no server is your own: no server locks it. The lock keeps Nether’s apps to the approved endpoints; it cannot stop another program from reading the vault’s files. Admins set it in the admin console.
Writing assist
Section titled “Writing assist”Select some text in a note, or put the caret in a paragraph, and press Ctrl+J (Cmd+J on a Mac), or choose Writing assist in the Note menu. It opens beside the note, with the text it works on highlighted, and says where the text will go, such as “Sends the text to Ollama, llama3.2 on this computer.” Then choose what to do:
- Rewrite: say it better, keeping its meaning. Type how in How to rewrite, if you like, such as “shorter”.
- Fix grammar: fix spelling, grammar and punctuation only.
- Summarize: a few sentences, put after the text.
- Translate: into the language in Into, the app’s by default.
- Continue writing: a paragraph or two more, put after the text.
The reply appears as it is written. Stop ends it early. Check it, then choose Replace to put it in the text’s place, Insert below to put it after, Try again, or Discard (Escape). Nothing goes into the note until you choose, and one undo, Ctrl+Z, takes it out again. If the text was deleted while the reply came, nothing is put in.
Only the text you selected is sent, as the note’s Markdown, with the instruction. Links, tags and formatting in the reply become the note’s own. Nether records nothing of what is sent or received.
Writing assist works on notes shown as blocks, not in source mode.
Asking your notes
Section titled “Asking your notes”With a provider, the sidebar has an Ask tab, and the View menu has Ask your notes. Type a question and press Enter; Shift and Enter makes a new line. The tab says first where it goes, such as “Asks Ollama, llama3.2 on this computer”.
Nether finds the notes your question is about with the vault’s own search, by its words, and by meaning too once search by meaning is on for the vault. It sends the question with at most eight passages of those notes, each under its heading, and the questions and answers before it. If no note has your words, nothing is sent: try others.
The answer cites its passages by number, as [2]. Each citation is a link: choose it to open the note at that passage. A citation that names no passage Nether sent shows dotted and is no link, as the model wrote it unchecked. Read from under each answer lists every passage that was sent, those cited in bold.
The questions stay while the vault is open. Save as note keeps them as a note, named by the first question, each answer’s citations links to their notes’ headings, which Obsidian reads too. Clear starts again.
In the web app, asking goes through your server as writing assist does, and an end-to-end encrypted workspace has it in the desktop app only.
Search by meaning
Section titled “Search by meaning”Search by meaning finds notes by what they are about, not only by their words: “keeping tomatoes from splitting” finds a note on watering that never says “split”. It is off until you turn it on for a vault.
In the AI tab, under Search by meaning, choose an
OpenAI-compatible provider and its embedding model, as the provider names
it, such as text-embedding-3-small for OpenAI or nomic-embed-text for
Ollama, then Turn on. Anthropic offers no embedding models.
Nether then sends the vault’s notes to that model, a passage at a time, the text under one heading, and keeps what comes back in the vault’s index on this computer. That sends the whole vault once, then each passage that changes; a paid provider charges for what it is sent. A provider on this computer, such as Ollama, keeps the text on it. The tab shows how far it has come.
In the sidebar’s Search, choose By meaning, type what you are
looking for, and press Enter. Each note found shows its passage closest in
meaning, and opens at that passage’s heading. Filters such as tag: and
path: work as they do by words.
Another model embeds the vault anew. Turn off stops, and keeps what was embedded, for turning it on again. While AI is off in the vault, or its workspace’s owners and admins turned it off, search by meaning is off too. In a vault of an end-to-end encrypted workspace, choosing a provider here gives it the vault’s text, as choosing one for AI in this vault does.
Building with AI
Section titled “Building with AI”With a provider, the File menu has Build with AI. Choose what to build, describe it, and choose Build:
- A database: its properties and views; tick With a few sample rows for a few rows to start from.
- A template: a note in the templates folder, to start others from.
- A project plan: a folder of notes, the first its outline, and a database of its tasks.
The dialog says first where the description goes. The model answers in a form Nether checks as it checks a database you make by hand: when the answer does not read, as a board with no property to group by, Nether says why and makes nothing, and you can try again or describe it another way. Relations and rollups to other databases are added once the database is made.
What it would make shows before anything is written: every note, at its path, a database with its properties and views, a template or a note with its first lines. Change the Name and the paths follow; untick rows or notes to leave them out. Names the vault has already get a number, as imports do. Make it writes them and opens what was built; Cancel makes nothing.