This page is machine-translated from French. Read the French original.

Conversations

This page describes the settings that determine how the assistant responds, and the concrete effect of each.

Conversation modes

The mode is selected in the space settings, tab Conversation.

Cat (by default)

The template uses your documents And his general knowledge. If he finds nothing relevant in space, he still answers based on what he knows.

For : General assistance, writing, brainstorming, questions combining documents and general knowledge.

Request

The model responds exclusively based on the documents in the space. In the absence of a relevant passage, it returns a rejection message instead of responding.

For : Customer files, regulatory framework, technical documentation — any use where a made-up answer would be worse than no answer.

THE rejection message is customizable. By default it indicates that no relevant information was found; you can replace it with wording adapted to your users ("This information is not in the file. Contact the manager.").

Automatic

Each message is assigned to an agent, who alone decides whether or not to use tools.

Reserved for spaces dedicated to automation. This mode is slower and more unpredictable than the other two, and assumes a model capable of correctly controlling tools. For occasional use of agents, prefer the prefix @agent in a chat-mode space.

The prompt system

This is the application's most powerful lever: the permanent instructions transmitted to the model before each message.

A good prompt system specifies:

• the role — «You assist the employees of a Swiss trust company.» ;
• the reference frame — what rules, what law, what method apply; ;
• the expected behavior in the face of uncertainty — «If the information is not in the documents, state it explicitly. Do not deduct amounts.» ;
• the format — length, structure, response language; ;
• the prohibitions — something the assistant should never do.

Example :

Tu assistes les collaborateurs d'une fiduciaire.
Tu réponds uniquement à partir des documents du dossier client.
Si l'information n'y figure pas, indique-le clairement sans supposer.
Cite systématiquement le document et la date de la pièce utilisée.
Ne calcule jamais un montant qui n'apparaît pas explicitement dans une pièce.
Réponds en français, de façon concise et factuelle.

A default system prompt can be defined at the instance level (Réglages → Prompt système par défaut): it applies to new spaces without overwriting those that have their own.

The prompt system accepts variables — date, time, username, space name, plus any others you define. → Commands and variables

Temperature

Controls the degree of unpredictability of the responses.

For documentary purposes, a low value is almost always preferable: twice the same question should give the same answer.

History length

The number of previous messages returned to the model for each question. Default value: 20.

• Higher — the assistant follows a long discussion better, but each message costs more context, therefore more time and, with an external provider, more money.
• Lower — faster and cheaper responses, at the cost of less conversational follow-up.
• Zero — Each message is processed individually.

Helpful reminder: open a new thread is often preferable to reducing the history. → Workspaces

Document search parameters

Space settings, tab Vector base.

Number of fragments found

How many passes are retrieved and provided to the model?. Default: 4.

• Increasing coverage improves broad issues, at the cost of consumed context and potential noise.
• Diminish focuses the response on the most relevant passages.

For long documents and cross-cutting questions, 6 to 10 often yields a better result. Beyond that, the context becomes saturated.

Similarity threshold

The minimum degree of proximity that a fragment must reach in order to be retained. Default: 0.25.

Symptom and remedy: «"He can't find it even though it's written" → lower the threshold. "He quotes unrelated passages" → raise it.

Search mode

Some vector databases allow you to combine semantic search with keyword search. Semantic search retrieves different wordings; keyword search retrieves exact references (article numbers, product codes). Combining the two is useful for technical corpora.

What exactly the model sees

For each question, the assembled context contains, in this order:

1. THE prompt system of space, resolved variables; ;
2. THE pinned documents, in its entirety; ;
3. your memoirs relevant and, where applicable, the elements of Second Brain ;
4. THE fragments found through research; ;
5. THE latest messages thread; ;
6. your question.

All of this must fit within the model's context window. When it's full, the oldest elements are discarded — hence the importance of not pinning everything and not letting a thread drag on forever.

Sources and verification

Each answer built from your documents displays the extracts used. These allow you to:

• check that the answer depends on the correct passage; ;
• diagnose a search problem — off-topic extracts indicate a threshold or a sectioning that needs to be reviewed; ;
• go back to the source for complete control.

An answer without a displayed source was not built on your documents: either the search found nothing, or the model responded from its general knowledge (Chat mode).

Adjusting a space: three profiles