Documents & Knowledge

A central document library with folders, visibility scopes and preview – plus per-chat document knowledge (RAG) for grounded, cited answers.

The Documents area is where you store your files centrally, organize them into folders and subfolders, and decide who may see them. Most importantly, these documents become a knowledge base for the chat: you add selected files per chat, and the AI answers your questions from their content – with source citations. Kasimir automatically picks the right processing strategy (whole document in context, or targeted excerpt retrieval) depending on size and model.

What this area does

• Upload and manage files – PDFs, Office and OpenDocument files, text, CSV, Markdown, HTML, XML, JSON and images.

• Organize in folders – arbitrarily nested folders with emoji, color and description.

• Control visibility – Private (only you) or Company-wide (everyone in the company).

• Preview – PDFs and images right in the browser; Office files are converted to a PDF preview server-side.

• Use as chat knowledge – add documents per chat on purpose; the AI answers from them and cites the passages.

Uploading documents

1. Open the Documents area from the sidebar.

2. Click Upload at the top right and pick a file from your device.

3. The file uploads and is processed immediately (text is extracted). The status moves from Processing to Ready – or to Failed if no text could be obtained.

Supported file types

Visibility on upload

When uploading outside a folder you can choose between Private (the default – only you see the file) and Company-wide (everyone in the company).

When you upload inside a folder, the new file automatically inherits the folder's visibility – no separate choice needed. The active tab also sets the default: on the Shared tab uploads become company-wide, otherwise private.

Folders & organization

Tabs, search and view

Above the file list you'll find:

• Tabs All · Shared · Private – filter folders and files by visibility.

• Search – filters the current level by file and folder names.

• View – toggle between Grid (with thumbnail) and List (with details and an actions menu).

Creating a folder

1. Click Folder.

2. Give it a name, optionally an icon (emoji) and a description.

3. Choose the visibility (Private or Shared).

4. When a folder is open, new subfolders and uploads automatically land inside it (shown in the breadcrumb at the top).

Managing files

In List view, the three-dot menu per file offers these actions:

• Change visibility – switch between Private and Share company-wide. Only the uploader or an admin may do this.

• Move to – into another folder or back to General (no folder).

• Delete – with a confirmation prompt.

Preview

Clicking a file (status Ready or Failed) opens the preview:

• PDFs and images are embedded and shown directly.

• Office / OpenDocument files are converted to a PDF preview server-side and shown the same way.

• All other formats offer a download of the original.

The preview also shows file size, type and a Download button for the original file.

Recently used (from chats)

At the very bottom – only at the root level and outside the Shared tab – you'll see Recently used (from chats). These are files you attached in chats (strictly private, visible only to you). Per file you can:

• use Save to library to keep the file permanently (private) in your documents,

• or delete it permanently.

Using documents as chat knowledge

This is the heart of the feature: documents become the AI's source of knowledge.

Step by step

1. Open a chat. In the composer bar click the Documents button.

2. The picker is grouped:

   • Private – your own private documents.

   • Project – the current project's documents (only in a project chat).

   • Company-wide – company documents, grouped by folder.

3. Tick the documents you want. The button's counter shows how many are selected. Only documents with status Ready appear here.

4. Ask your question. The AI answers from the selected content.

The selection is remembered per conversation.

Citations

When the AI falls back to targeted excerpt retrieval, it marks used passages inline with [1], [2], …. Below the answer you'll find the Sources panel with filename, page/section and a text excerpt. That makes every statement traceable – important for safety-critical or standards-related topics.

Assistants

In an Assistant chat, knowledge works differently: the knowledge assigned to the assistant is used automatically (linked company-wide documents and the assistant's own uploads) – independent of the manual picker. That is the assistant's own knowledge base.

How the AI processes your documents (strategies)

Depending on the size of the added documents and the context window of the chosen model, Kasimir automatically selects one of several strategies:

• The default threshold is 50,000 tokens.

• The threshold is additionally capped by the model's context window (minus a safety reserve of ~14,000 tokens for the system prompt, history, tools and answer). This way even a small-window model never overflows.

• When switching to excerpt retrieval, documents not yet indexed are split and embedded on demand. While that happens you see a brief status note in the chat.

Setting the strategy per project

In project chats the strategy can be fine-tuned. Inside the project, under the document strategy tab, you'll find:

• Three tiles Auto · Full · RAG (Hybrid + a manual threshold live under Advanced).

• A capacity meter that compares your documents' token usage against the active model's usable budget (including a rough page estimate, ~600 tokens/page).

• A breakdown of "in project" vs. "company-wide" plus the number of already indexed documents.

• A Reindex button that embeds documents still pending.

Visibility at a glance

Limits & good to know

• Maximum file size: 25 MB. Larger files are rejected.

• Pure images as company knowledge only work via the image description (text is derived from them); pixels without text yield little.

• Status Failed usually means no text could be extracted (e.g. a scanned PDF without an OCR layer). The file can still be opened/downloaded but is not usable as a knowledge base.

• Changing visibility is restricted to the uploader or admins.

• Token values are estimates; the actual strategy may vary slightly by model.

• Search on the overview page filters file/folder names only (not full text) – content search happens in the chat via excerpt retrieval.