> ## Documentation Index
> Fetch the complete documentation index at: https://docs.fildos.cloud/llms.txt
> Use this file to discover all available pages before exploring further.

# Ask AI

> Chat with your files through a fully on-device LLM that can mention files, run searches, and act on your data.

The **Ask AI** panel lets you research across your files in plain language.
Answers come from a model running **entirely on this device** — nothing leaves
your system.

<Frame caption="Ask AI — research across your files with an on-device LLM.">
  <img src="https://mintcdn.com/fildos/0Cun-WesrIb9SHEK/images/ask-ai.jpg?fit=max&auto=format&n=0Cun-WesrIb9SHEK&q=85&s=8863503acc0726079ff26beee534e740" alt="The Ask AI chat panel" width="1655" height="1031" data-path="images/ask-ai.jpg" />
</Frame>

## Mentions and commands

The composer understands three kinds of reference:

| Trigger    | What it does                                                                                 |
| ---------- | -------------------------------------------------------------------------------------------- |
| `@file`    | Attach a file's contents to the conversation                                                 |
| `#folder`  | Bring a folder's listing into context                                                        |
| `/command` | Run a built-in command — e.g. `/find` runs a semantic search and hands the hits to the model |

Quick-action tiles get you started: **Summarize this folder**, **Find a file**,
**Explain a file**, and **Perform a file action**.

## Acting on your files

Beyond answering questions, the assistant can *do* things — create, copy, move,
rename, list, read, and delete-to-Trash — through function calling. Every action
is:

* **Recoverable** — deletes go to the OS Trash; creations and copies never
  overwrite.
* **Visible** — each call streams to the panel as an activity chip and is saved
  with the message.
* **Yours to approve** — file-changing actions happen with your consent.

## Bring your own model

FilDOS ships a built-in catalog spanning **0.6B to 35B** parameters, grouped by
family — Llama, Qwen, Gemma, Phi, Mistral, SmolLM, DeepSeek, Granite, and more.
Every model runs on your own hardware via `node-llama-cpp` (Metal on Apple
Silicon).

* **One model stays resident**; switching disposes the old one.
* A **device probe** recommends the best-fitting model for your machine.
* A **per-model customize panel** tunes temperature, top-p, max tokens, context
  window, and custom instructions.
* You can also **add any GGUF from the internet** — paste `hf:owner/repo`,
  `owner/repo`, or a direct `.gguf` URL.

Manage everything in **Settings → Assistant**.

## Conversations are saved

Every exchange — including the files you mentioned and the sources a `/find`
pulled in — is persisted locally, so you can reopen a conversation from the
history view and continue it later.

<Note>
  The assistant respects **Hide from AI**: anything you've excluded from indexing
  is invisible to chat as well.
</Note>