Limits and decisions

The following points are deliberate decisions, not bugs. You should know them before you roll out ChEddi.

The chat never talks to a model provider directly. Every turn goes to the AI Suite server, which holds the credentials. That also applies with your own provider keys, where the server forwards the request using your keys.

Consequence for operation: the server has to be reachable, otherwise the chat does not work even with your own keys.

Images are not yet supported as attachments. Messages currently travel to the server as plain text, so image content would not reach the model at all. Documents, on the other hand, are read, namely as text and only once ChEddi actually needs them. See Attachments.

The ability to take a change back comes from the draft workspace, not from the chat. In live mode a confirmed change takes effect immediately. Even in draft mode, the image generation tool always saves its file directly. The same applies to attachments you upload into the chat: they are in the file list right away. ChEddi does not offer a tool for uploading media.

A turn is a blocking request, so the answer appears in one piece rather than word by word. The thinking indicator during the wait stands in for streaming.

It is off unless it is explicitly enabled, and its behaviour may change between minor versions. A search sends the query out of the installation and costs credits, reading an address you named does not. Which search provider is used is a server-side decision, not a setting in this extension. Outside GDPR mode, models with their own web search use that instead. See Web research.