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

First steps

This process leads from a blank installation to an initial response built on your documents. Allow approximately ten minutes.

1. The first-time setup wizard

On first launch, WivenLLM lets you choose the essentials:

The language model provider. The engine generates the responses. You can choose a locally run model, an inference server from your organization, or an external provider. This choice can then be changed at any time, including on a workspace-by-workspace basis. → Choosing your models

The embeddings engine. It translates your documents into vectors. The built-in engine runs locally, requires no keys, and is suitable for most uses.

Important : Changing your embedding engine later renders existing indexes unusable and requires reindexing your spaces. This choice is worth making once, and correctly.

The vector basis. The integrated local database is the default choice: nothing to install, everything stays on disk. An enterprise database is justified for large volumes or when one already exists on your premises.

Your account. Name and credentials of the first account, which is the administrator.

Your first workspace. A name is enough — for example, the name of a folder, a client, or a project.

2. Understanding the main screen

• Left, the list of your workspaces. Each one has its own documents, settings, and conversations.
• At the center, the ongoing conversation.
• Down, the input area, with file storage and the microphone.
• Bottom left, access to settings.

A workspace is the organizational unit of WivenLLM: it determines which documents are accessible, which template is relevant, and according to what guidelines. Two spaces share nothing unless you add the same document to both.

3. Import a document

Three ways:

• Drag and drop a file in the input area; ;
• open the document manager from the space bar (paperclip or folder icon) and upload from there; ;
• stick a URL to import a web page.

The import process takes place in two stages, and this is the most frequent source of confusion:

1. Extraction — The file is converted to text. It then appears in the document library.
2. Adding to space — you select the document and confirm: it is cut, vectorized and indexed.

Until step 2 is completed, the document cannot be searched.

→ Documents

4. Ask a question

Write your question in natural language. The answer will appear as it is generated, followed by a note. Sources which unfolds the extracts used: document name and exact passage.

Three tips that greatly improve the quality of the responses:

• Be explicit. «What is the cancellation period in the Dupont contract?» is better than «cancellation period?».
• Check the sources. An excerpt unrelated to the question indicates an indexing or search parameter problem.
• One conversation, one topic. Previous exchanges are part of the context sent to the model; mixing topics degrades the responses.

5. Adjust the workspace

Open the space settings (gear icon next to its name). Three settings produce the main effect:

The mode of conversation. - Cat — the template uses your documents And his general knowledge. Request — the model responds uniquely from your documents, and indicates that it does not know when they do not contain the answer.

For strictly documentary use (customer files, regulatory database), the Query mode is almost always the right choice.

The prompt system. The model's permanent instructions: its role, its tone, what it must reject, the expected format. This is the most effective lever in the entire application.

The model. Each space can use a different model than the instance.

→ Conversations

6. Going further

If that doesn't work

The response ignores my documents. Check that they are properly added to the space and not just imported. Next, check the similarity threshold: too high a threshold will exclude relevant passages.

The answer is invented. Switch space to mode Request and specify in the prompt system that any information missing from the documents must be marked as such.

Responses are slow. A local model without a GPU is inherently slow. Try a smaller model, or switch to an external provider for this space.

An import fails. The format may not be supported, or the PDF is an image without a text layer — character recognition is then applied, with a result that varies depending on the quality of the scan.