AI models and the knowledge base
Both halves of the AI — the model that answers and the index it can consult — live on your own device. The website is a remote control for them, not a place they run.
Choosing a model
The Recommended tab lists models that suit the selected device: the device compares an up-to-date list of models with its own processor, memory, graphics card and free disk space, and shows the best fits for chat, code and images, with where each would run (GPU, GPU + RAM or CPU) and how fast. Pick and install downloads the top chat model and switches the device to it; Install next to any model does the same for its role. Every download is checked against a fingerprint before it is used. Downloads run on the device, not in your browser, so you can close the tab and come back; a chip shows the percentage while it runs, and it can be cancelled.
Models already on the device are listed first, then the rest from the best fit down; the best fit carries a star, and that is the one Pick and install takes (if it is already downloaded, the device just switches to it). Each row says whether the model runs on the graphics card, on the graphics card and RAM, or on the processor, and roughly how many tokens per second to expect. Open a row for its size, context, memory use, license, popularity, last update and a link to the model's page.
The HuggingFace and Ollama tabs stay available for picking a model yourself. A device running Filemote older than 1.12.24 shows a plain Catalog tab instead of recommendations.
Models are GGUF files run through llama.cpp. A rough rule for picking one:
- A 1–2 B model runs comfortably on an ordinary laptop and is fine for summarising and short answers.
- 7–8 B wants a machine with real memory to spare, and answers noticeably better.
- Anything larger is for a workstation with a GPU.
When a model is loaded the panel says so and the chat starts answering. Deleting a model frees the disk it occupies; the settings keep pointing at it, so pick another one afterwards.
Specialist slots
Beyond the main model, four optional slots exist: document, vision, code and summary. Filling one means questions of that kind are routed to that model instead of the main one; leaving it empty means the main model handles everything.
This is worth doing when one kind of work matters to you — a vision model to read scanned documents, say — and not worth doing otherwise, since every loaded model holds memory for as long as it is loaded.
What the knowledge base is
Indexing a folder turns its documents into chunks the AI can look things up in. Then, with Use knowledge base in chat on, relevant chunks are passed to the model along with your question, so it can answer from your material rather than from what it happens to remember.
The panel shows how many chunks and documents are indexed. Index now runs a pass over your chosen folders; Clear base throws the whole index away, after confirmation.
Indexing reads a lot of formats — documents, spreadsheets, PDFs, images through OCR, audio and video through transcription. Big folders take a while, and the first run is the slow one.
The three privacy tiers, precisely
This is the part worth reading carefully, because the tiers are not the same thing as "private".
| Tier | What goes in it | Used to answer questions? | Included in a dataset export? |
|---|---|---|---|
| 0 | Ordinary documents you indexed | Yes | Yes |
| 1 | Browser history | Yes | No |
| 2 | E-mail (.eml, .msg) | No | No |
So e-mail can be indexed and searched by you, but is never fed to the model as context — being in the knowledge base is not the same as being in an answer. Browser history is used in answers but never leaves in an export.
Both of the sensitive tiers are off until you turn them on and tick the consent box, and the data behind them stays on the device in every case. Nothing indexed is uploaded anywhere: the index lives beside the AI service, and the website only ever sees counts.
Scheduled re-indexing
A folder that changes daily is worth re-indexing on a schedule, which the desktop app
configures with days of the week and a time of day. The fields behind it are KbScheduleDays
and KbScheduleTime in
the agent settings; an older KbScheduleIntervalHours
in a settings file is read but no longer controls anything.
Mirroring a public page
Everything above indexes files already on your machine. A mirrored page is the same idea for something on the web: you give an address, the page is downloaded once, and from then on it is an ordinary document in your knowledge base.
The point of doing it this way is what happens afterwards. Answers that use a mirrored page make no request at all - the text is already local - so they work with web search switched off, and nothing about your question leaves the machine.
Two things are worth knowing before you add one. A mirrored page is stored as a public document: it takes part in answers and is included in dataset export, which matters if you are tempted to mirror a page that sits behind your own login. And a page that could not be downloaded stays in the list with the reason it failed - that list is the only place where you can see why a page you added is not answering anything.
Refreshing re-downloads a page and re-indexes it only if the text actually changed. You can refresh by hand at any time, or let it happen during the knowledge base's scheduled run - that is a separate switch, off by default, because a schedule you set up for your own files should not start reaching the network just because you added a page later.
When the panel says the AI server is not running
The knowledge base is managed by the AI service, so with the service stopped the panel says so plainly. It may still show you what is on disk, marked as coming from disk — counts you can read, controls you cannot use until the service is up.
Starting it is one button here. If it refuses, the AI does not load covers why.
Doing the same from the desktop
Everything on this page exists in the Windows configurator as well, with more detail — see the AI Hub. The two are views onto the same service: a model downloaded from the browser is the same file the desktop app then loads.
Updated 2026-10-03