Chat and models

This page describes what your team finds in the chat: history, search, model cards, memories and the behavior of very long conversations. At the end there is a section for administrators with the switches behind all of it.

The chat interface comes from LibreChat. What we add is the curation of the model cards, the stored system prompt and the permission handling on top.

Four terms up front

Term	In one sentence
Token	The unit language models compute in: a piece of text of roughly four characters, so about half a word to a whole word. Everything you write and everything the colleague answers is counted in tokens, and usage, budgets and context limits all hang on that.
Model card	A ready made package of provider, model and presets that you pick under a comprehensible name instead of naming a raw model.
System prompt	The working instruction handed to the model ahead of your question, which sets language, tone and behavior. It belongs to the card and is not visible in the chat.
Context	Everything the model can read at once for one request: your question, the history so far, attached files and tool results. The room for it is limited and is measured in tokens.

Everyday use

You type a question and the answer appears progressively, word by word. Every conversation lands in the history in the left column and is given a title automatically, derived by a model from the first messages. You can overwrite that title at any time.

Feature	What it does
History	All conversations stay stored until you delete them or the retention period takes effect
Automatic titles	The title is created on the first exchange, without you doing anything
Bookmarks	Mark important conversations and find them again faster
Parallel answers	Run several chats side by side and compare the answers
Private chat	A conversation that is not stored
Fork	Open a new branch from any message without changing the original
Share	Publish a conversation as a link

“Privater Chat” (private chat) is called “Temporäre Chats” (temporary chats) in the role editor of the admin panel, where it carries the note “Unterhaltungen ohne Speicherung” (conversations without storage). How long such a conversation is kept at all is set by the admin.

Forking is the right move when a conversation should continue in two directions. You keep the history so far and try a different variant in the branch, without losing the original conversation.

Sharing creates a link to a snapshot of the conversation. Whoever has the link can read along. The permission for this is a separate role permission and can be withdrawn.

Searching old chats

The sidebar offers a search across the full text of all your own conversations. So the search also finds a phrase in the middle of an answer from three weeks ago.

Three searches that are not the same thing

The full text search across your chats is something different from file search across a document collection, and different again from web search. Which one is responsible for what is covered in Files and knowledge and in Web search.

What a model card is

A model card is a ready made preset. It bundles three things:

1. a specific model at a specific provider,
2. the stored system prompt, meaning the working instruction given to the model,
3. the preselection of tools that are armed on this card.

The system prompt is inserted server side and never reaches the browser. The shipped cards carry the same prompt; it sets the language, the tone and the careful handling of facts.

New chats start on the default card. You are allowed to switch cards, the switch is not locked.

Which cards you find

Which cards appear in the picker is decided by the admin. Two distinctions matter in daily work:

• Text, or text and images. Not every card understands images. Screenshots and photos belong on a card marked as image capable; on a text only card an image has no effect.
• EU or non EU. Cards whose requests leave the EU carry the marking ⚠ Nicht-EU (non EU) in their name.

The marking is deliberate

The sign ⚠ and the addition “Nicht-EU” (non EU) are part of the card name so that it is visible at the moment of choosing that this request leaves the EU. The marking must not be removed as long as the card points at a provider outside the EU. Anyone who wants a purely European instance does not offer such cards in the first place.

Memories: the long term memory

The AI colleague remembers a few things about you across conversations and pulls them into every new answer. This happens by itself, you do not have to say “remember this”.

Deliberately little is stored, and only in four categories:

Category	Content
Preferences	How you want answers: language, format, level of detail, tone, recurring corrections
Work context	Role, team, responsibilities, systems in use
Projects	Ongoing projects, customers and topics with a short status
Notes	Other stable facts that make future answers better

Explicitly not stored are passwords and keys, health, religious and political information, salary data, sensitive facts about third parties and one off details with no reuse value.

The store is deliberately kept small. When it fills up, the colleague merges related entries instead of appending indefinitely.

How to turn it off. In the settings under “Datenschutz & Sicherheit” (privacy and security), section “Erinnerungen” (memories), you find the switch “Gespeicherte Erinnerungen verwenden” (use stored memories). Off means your memories no longer flow into the answers. In the same area you can view your entries, edit them individually and delete them.

Very long conversations

A model can only read a limited amount of text per request. When a conversation gets very long, a separate run summarizes the older messages automatically and from then on sends that summary to the model instead of the full history.

• The point at which this kicks in depends on the card you chose: cards with a smaller context window compact earlier.

Your visible history stays complete throughout. Only what goes to the model is summarized.

There is no manual command for this

A command such as /compact with which you trigger compaction yourself does not exist. The process runs automatically and only automatically. If you want a deliberate fresh start, open a new chat or fork at the appropriate point.

For admins

Model cards per role and group

Which cards a person sees is decided by the admin panel. The cascade is group before role before base.

• “Verwaltung → Rollen” (administration, roles): one switch per card, per role. Plus the switch “Weitere Modelle (Roh-Untermenü)” (further models, raw submenu) with the note “Aus = diese Rolle sieht NUR die oben aktivierten Karten” (off means this role sees ONLY the cards enabled above).
• “Verwaltung → Gruppen” (administration, groups): per group either “Erben” (inherit) or “Eigene Auswahl (übersteuert die Rollen-Regel)” (own selection, overrides the role rule).

This is the clean way to grant management additional cards, for example, and show everyone else only the EU cards.

Hiding the model picker entirely

For a deliberately simple setup, set “Modell-Auswahl anzeigen” (show model selection) to off in the panel under “Konfiguration → Funktionen” (configuration, features tab). Then everyone works on the default card and no longer sees any model selection.

Parameter panel

The panel for temperature and similar fine settings is off as shipped. That is deliberate: team members without an AI background would otherwise change values whose effect they cannot judge. Turning it back on is done per role through the configuration cascade in the panel; in practice the role ADMIN gets that override and all other roles inherit the disabled behavior.

A sensible order when setting up

1. Maintain prices and context sizes under “Konfiguration → Preise” (configuration, prices) so that usage metering is correct. An empty price field does not mean free, LibreChat then calculates with a built in flat rate.

2. Define the cards per role and group.

3. Only then invite your team.