The Assistant now reads attachments
Attach a PDF, image, CSV, text or Word file to a conversation and the Assistant reads it with your question. Uploads are checked by their content, and attachments leave with the account.
So far the Assistant could only work with records that were already in the app. Now a user can attach a file to the conversation: a PDF, an image, a CSV or text file, or a Word document (DOCX or the older DOC). The model reads it along with the question.
What a tenant can attach depends on the model provider it runs. Anthropic models read PDFs, images and text files directly. OpenAI-compatible models read images, and other types reach them as extracted text when a matching document reader is set up. The attach dialog offers only the types the current setup can read.
Word files are converted to text on the server and need no LibreOffice installation. Tables stay readable row by row. Embedded images are skipped without being unpacked, so a picture-heavy document does not drive up memory. While a file is still being read, the Send button waits, and if reading fails the user gets a message saying so.
Uploads are checked by what they contain. A file named .pdf has to be a PDF, and a renamed file is refused. Before a Word file is opened, its internal structure is checked and archives that would unpack to an absurd size are rejected.
An attachment belongs to the person who attached it and works only in their own conversation. It counts as their personal data, so deleting their account also deletes the attachment and the stored file. Deletion follows the tenant’s retention rules as well: on a tenant with the German HGB profile, invoices, bookings and contracts stay in place after an account deletion, with the person’s details anonymized.
Attachments ship with the Assistant in the commercial Kumiko Enterprise offering. The upload checks and the retention rules are part of the open framework.