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

Documents

WivenLLM responds using your documents. This page explains how to enter them, what happens to them, and how to get good results.

The two stages of importing

This is the point that most often causes confusion.

Step 1 — Extraction. The file is converted to text. It joins the document library, common to the instance. At this stage, it cannot be queried in any space.

Step 2 — Adding to a space. You select the document from the library and add it to a workspace. It is then clipped, vectorized, and indexed for that workspace.

The same document can be added to several spaces: it is only extracted once, and indexed separately in each.

Drag and drop directly into a conversation automatically combines the two steps.

How to import

Files. Drag and drop into the conversation, or upload from the document manager. Multiple files at once are accepted.

Web pages. Paste a URL. The page content is extracted; advertisements and navigation menus are excluded as much as possible.

Complete site. A recursive import follows links from a starting page, down a depth and number of pages that you specify. Use sparingly: the depth quickly increases the volume.

External sources. Connectors import directly from:

From the API. Upload files, plain text, or URLs to feed WivenLLM from your own applications. → API developer

Accepted formats

Text and office automation — .txt, .md, .csv, .json, .html, technical documentation formats, .pdf, .docx, .pptx, .xlsx, OpenDocument formats.

Messaging and books — archives .mbox, digital books .epub.

Images — .png, .jpg, .webp, the text of which is extracted by character recognition.

Audio and video — main audio formats and .mp4, transcribed automatically. A meeting can also be recorded directly from the application → Meetings.

→ Detailed list

What works less well

• Scanned PDFs without text layer — Character recognition is applied, with results depending on the scan quality. A tilted or low-resolution scan produces poor text, and therefore poor results.
• complex spreadsheets — Formulas, merged cells, and very wide sheets do not linearize well. For large tabular data, exposing a database via the agent's SQL capabilities yields much better results than an indexed spreadsheet.
• Documents with very rich layout — the text is extracted; the information conveyed solely by the visual layout is lost.

What happens to the text: the segmentation

An entire document cannot fit within the context of a template. It is therefore divided into fragments, and these are the fragments that are found when questioned.

Two parameters, adjustable by an administrator in Réglages → Découpage du texte :

• Fragment size — the number of characters per fragment. Short fragments give precise results but sometimes lack context; long fragments retain context but dilute relevance.
• Recovery — the number of characters repeated from one fragment to the next, so as not to split an idea in two.

The default values are suitable for most corpora. Modify them only affects documents indexed later : existing indexes retain their structure until reindexing.

Organize the library

• Files. Organize documents by client, project, or type. Folders organize the library; they don't change how you search.
• Shift. A document can be moved between folders without losing its indexing.
• Deletion. Deleting a document from the library removes it from all removes spaces and erases the extracted text.

Automatic synchronization — beta function

For documents imported from a URL or connector, periodic synchronization can detect changes at the source and reindex the updated content.

She is active in Réglages → Fonctions bêta, Then it is adjusted document by document. Useful for dynamic documentation; useless for a static corpus.

Best practices

Import what is relevant, not everything. A well-defined corpus provides better answers than an exhaustive one. Obsolete documents provide answers just as well as others—they resurface with the same apparent authority.

Check the extracted text. After a difficult PDF import, ask a question you know the answer to. If the quoted excerpt is unreadable, the document was extracted incorrectly; converting it beforehand will yield better results.

Name your files explicitly. The document name appears in the citations. Contrat-Dupont-2026.pdf is useful; ; scan_003.pdf It is not.

Remove it before adding a new version. Two versions of the same contract indexed simultaneously produce contradictory responses, without the model knowing which one is valid.

Pin sparingly. A pinned document is injected in its entirety with each message. → Workspaces