Documents and AI

What happens automatically

Text extraction. From PDFs, office formats and text files the text is taken directly. Images and scans go through recognition — locally, without sending the file anywhere.

Indexing. The text is cut into fragments and enters search: both exact full-text and vector search by meaning. That’s what the agent uses later.

Description and tags. A short digest and tags are prepared for a document in the background — there’s no button for it; it happens in the background job queue after parsing.

Agent memory. Durable facts, people, companies and projects are extracted from the content; they feed the agent’s wiki and the edges in the graph.

All of these need an available model. If the local one is busy and the cloud isn’t connected, they wait — the document meanwhile works completely: it opens, reads and is searchable by text.

The agent and documents

In chat the agent can:

• find a document by the meaning of your request;
• read it and answer from its content, citing the source;
• compare several documents, if you ask for that explicitly.

Answers always carry their source. Opening the link is cheaper than trusting the retelling.

Privacy

Parsing, recognition and indexing run on your device. Only what a specific cloud request needs is sent out, and the text passes sensitive-data masking first — card, account and passport numbers are replaced with placeholders.

With cloud features off, search and answers about documents run on the local model.