Choosing what the AI can read
Everything you can see in the file tree is in reach of the AI by default — notes, PDFs, books, Office files, and images and scans once they have been read. This page covers how to draw the line, and how to see what the AI can actually reach.
Three levels of "not this"
| You want | Do this | Effect |
|---|---|---|
| The AI to leave a file or folder alone | Right-click it → Hide from AI | Still in the tree and in search; AI retrieval skips it |
| A folder gone from MindMax altogether | Settings → Knowledge base → Exclude | Gone from the tree, search and AI. Nothing on disk is touched |
| One conversation to use only certain documents | @ them, then These N only above the input | That conversation reads only those — no knowledge base, no web |
A hidden file is marked in the tree, so a document that never comes up has a visible reason — and a folder hidden from AI hides everything inside it. If you add a hidden file to a conversation yourself, the AI reads it that time: hiding is about what it finds on its own, not about what you hand it.
Excluding is stored in a .mindmaxignore file in the folder and can be undone from the same
page. Folders that look like they hold no source material — build output, archives — are
suggested for exclusion.
What is left out automatically
Chat history and the assistant's own working files are never searched as sources, so it does not find its own earlier answers and cite them back to you as evidence. What it hands you as finished work — a translation, a draft, a collected note — is searchable, because that is material you may want to quote.
Seeing what the AI can reach
Settings → Knowledge base opens on a count: The AI can read N of your M files. Under it, anything that needs you:
- Being read in — PDFs, books and Office files are read in the background the first time. This finishes on its own. If it is stuck on one very large file, Skip the rest; skipped files are read in next time the knowledge base opens.
- Couldn't be read — the file took too long or may be damaged. It is not retried automatically until the file changes or MindMax restarts; press Retry, or open it to check.
- Images not yet read — an image becomes searchable once it has been described. Read these N does it in one go. A description, once made, is shared by the image viewer, the canvas and search, so nothing is read twice.
- Scanned documents — no text layer until you recognise them, one document at a time.
- Not searchable — audio, video and files that are not documents.
Helping it pick the right file
Every file can carry a one-line AI summary that helps the assistant choose between files that look alike. New and edited files get one the first time they are needed; Write the remaining N fills in the rest in advance. This is optional — the full text is searchable either way.
How each part of the app uses the folder
The defaults differ on purpose, and the same settings page spells them out:
| Where | Default |
|---|---|
| Chat | Searches the whole knowledge base. Documents you add are always read in full. |
| Writing | Looks at what you already have before drafting, then tops up per section. |
| Canvas | Off. Turn it on with the ◇ switch when you want it — diverging works better without your old conclusions up front. |
| AI in the editor | Generating instructions search; rewriting instructions do not. |
| Translation | The source text only, plus translation memory for terminology. |
What leaves your machine
Your files stay where they are. To answer, the passages the assistant reads are sent to the model you use — through MindMax, or straight to the provider of your own API key. If nothing should leave the machine, point MindMax at a local model such as Ollama or LM Studio.