Using a brain in chat

Up to 50 brains can be attached to one chat. Manage Company Brain at the end of the list takes you to the Company Brain page. If you may not use any brain, the button is missing from the message field.

1. 1

   Open the list

   Click Brain in the message field. The Brains for this chat list opens.
2. 2

   Tick brains

   Tick one or more brains. The list only shows brains you may use.
3. 3

   Ask your question

   Write your question as usual. The selection applies to this chat only.

Brains set to Active in new chats by default are already selected in every new chat. In the list they carry the tag Default. You can deselect such a brain per chat; that changes nothing for anyone else.

Projects and agents can bring brains along as well. The same rule applies on every route: you only get content from brains you may read yourself.

Under Retrieval in the same list you decide how the model uses this chat's brains. The setting applies per chat.

When a brain is active, the model gets the matching passages, the brain's Instructions and an outline of its content: the Overview and a list of the documents. Models that can use tools also search the brain themselves and read individual documents section by section.

Documents with the status Processing, Error or Replaced are not used for answers.

When a reply draws on the brain, the text carries numbers such as [1]. Below it Custodos shows, for example, 3 sources from the Company Brain. Expand the list to see the documents. A click opens the document preview at the cited passage.

Numbered entries are the passages the reply relies on. Entries without a number are what the model found during its own search. While searching, Custodos shows Searching Company Brain…, afterwards Company Brain searched.

Custodos checks on every message which brains you may read. If your access is withdrawn, the brain returns nothing from then on, including in existing chats.

A source link only opens the document for people who may read the brain. This also applies to links in shared chats. Anyone without access sees Page not found. More under Brain permissions.

• The model does not read every document in full, only the matching passages. Questions about all documents, such as “list every contract”, may be answered incompletely in large brains.
• If the information you need is only in a chart on a page that also has text, the brain does not find it, because such charts are not read.
• While the workspace is throttled, the model does not search the brain itself. The automatically attached passages still arrive. More under Credits and throttling.
• In security mode the passages from every brain are pseudonymised, including brains set to Open. Details under Data protection in the brain.