Files and knowledge

This page separates three things that get confused constantly in daily work: attaching a file to a message, making a whole document collection searchable, and searching your own conversations. Two of them are there after installation, one is not.

Three things, one word

What you want	What it is called	State as shipped
The colleague should read this one file now	Attaching a file (context)	on
A collection of many documents should be permanently searchable	File search, also known as RAG, short for retrieval augmented generation, meaning answers built from looked up passages	off, set up by the admin
I am looking for a passage in my old chats	Full text search across conversations	on

File handling comes from LibreChat, as does the RAG interface behind it.

Attaching a file to a message

This is the normal case. You attach a document to your message, the content moves into the context of this conversation, and you ask your question about it. Attaching works through the menu in the input field or by dragging onto the chat window.

This makes sense for individual documents: reviewing a contract, summarizing minutes, having a table analyzed. The content stays in this conversation and does not become a searchable collection.

Formats

The usual office and text formats are processed, among them PDF, Word (.docx), Excel (.xlsx), PowerPoint (.pptx), text, Markdown, CSV, JSON and XML. Images in JPEG, PNG, GIF and HEIC work too, but only on a model card that understands images. On the text only card an image has no effect, see Chat and models.

The exact list of permitted file types is configurable. Where, is covered further down in the section for admins.

Limits

How many files you may attach per message and how large a file may be is set by the admin. The default is generous and is rarely the real limit. What matters more in practice is how much text a model can read in one request. A document of several hundred pages fills the context, crowds out the rest of the conversation and makes the answer worse. File search exists for collections like that.

Scanned documents need a text layer

Text recognition is not part of the delivery. A PDF consisting only of images, for example a scanned contract or a photographed receipt, contains no readable text. This applies in the web chat exactly as it does in the Outlook add-in.

The usual route: convert the document into a searchable text PDF beforehand with the tool of your choice (many scanners and document management systems do this anyway) or copy the text out and paste it into the chat.

File search across a document collection

File search is something different from attaching. Uploaded documents are split into sections and stored in a way that is searched not by keyword but by similarity of meaning. In response to a question the colleague then fetches exactly the matching sections instead of putting the whole document into the context.

The advantage shows with many long documents: a handful of manuals, a collection of policies, a project archive. Instead of “read these 400 pages” it becomes “find the three passages that answer my question”.

Switched off as shipped

File search is off out of the box, because it requires additional building blocks and a provider for the embeddings. That way only what actually works is visible after installation. How to switch it on is explained below.

Full text search across your own chats

The search in the sidebar searches your conversations, not your documents. It finds the passage where something was discussed. This search is always on, it needs no embeddings provider.

For admins

Switching on file search

1. Store a provider for the embeddings. Without one, documents cannot be prepared for search by meaning. If you want to keep the EU positioning, choose a provider that processes EU resident.

2. Provide the additional building blocks. File search brings its own service for preparing and searching the documents; it is addressed internally only.

3. Enable the interface. Only then does the switch appear in the chat.

4. Check the role permission. In the role editor under “Werkzeuge” (tools) you find the permission “Datei-Suche (RAG)” (file search) with the note “Hochgeladene Dokumente durchsuchen” (search uploaded documents). Optionally alongside it “Datei-Zitate” (file citations) for source references from files. If you want to grant file search to individual roles only, maintain the permission exclusively here.

5. Check afterwards that the search really returns hits from the collection in a fresh chat.

Most common failure pattern

File search stays ineffective because the additional building blocks are not running at all. The switch is visible, but every request comes up empty. So before enabling it, check that the search service is running.

Changing limits and permitted file types

In the panel under “Konfiguration → Dateien & Speicher” (configuration, files and storage). There you find the upper limits for uploads, the limit for profile pictures and the list of permitted file types.

One sensible intervention during setup is to lower the upper limit per file to a value that is realistic for your daily work. That keeps accidentally uploaded videos and archives out of storage.

What happens when an account is deleted

When a person is deleted through the admin panel, their files are removed as well and the search indexes are cleaned up. The dialog explicitly offers “Daten exportieren (JSON)” (export data as JSON) beforehand, because the disclosure required by the GDPR has to happen before deletion.