You can attach documents to a conversation so the AI takes them into account. The absolute rule: the native file is never transmitted, even when the provider’s API accepts attachments. Trust6 reads the document locally, extracts its text, masks it like everything else, and sends only that.

Attaching a file

  1. Click the paperclip in the composer and pick your file.
  2. A card appears with the file’s name, size and extraction status.
  3. Once extracted, the document becomes the conversation’s context: it accompanies every send while attached. Removing it is one click (and it is logged).

Supported formats

  • Text: .txt, .md, .csv
  • Word: .docx
  • PDF containing text (a PDF scanned as images, with no text layer, cannot be read: the app will say so rather than guess)

What actually leaves

At send time, the extracted text is analyzed by the local engine, masked, then inserted into a neutral, numbered block (“ATTACHED DOCUMENT 1”). Two details that matter:

  • The file name is never sent: “Tremblay-contract-final.docx” would be information in itself.
  • Tokens match those in your messages: if “Jean Tremblay” is PERSON_001 in your conversation, he is PERSON_001 in the attached document too. Mappings are stable per vault.

You can verify this yourself after a send: open “View network payload” and read the document block, masked, exactly as the provider received it.

If the file cannot be read

A corrupted or unsupported file gets an “Unusable” card, and sending is blocked while that card is attached. Content the app cannot read never leaves as-is: that is the fail-closed principle. Remove the card, and sending works again.

See also