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

Architecture

WivenLLM consists of three cooperating services and interchangeable components connected to the main service. This page describes the system without going into the code details.


The three services

┌──────────────┐      ┌──────────────────────────┐      ┌────────────────────────┐
│  Interface   │─────▶│         Serveur          │─────▶│ Service d'extraction   │
│ (navigateur  │◀─────│ API · orchestration IA   │◀─────│ texte des fichiers     │
│  ou appli)   │      │ base · sécurité          │      │ et des URL             │
└──────────────┘      └───────────┬──────────────┘      └────────────────────────┘
                                  │
              ┌───────────────────┼───────────────────┐
              ▼                   ▼                   ▼
        ┌───────────┐       ┌────────────┐      ┌─────────────┐
        │  Modèles  │       │ Embeddings │      │    Base     │
        │    LLM    │       │            │      │ vectorielle │
        └───────────┘       └────────────┘      └─────────────┘

Interface

The application you use: workspaces, conversations, documents, administration screens. It runs in a browser, or in the desktop application that includes it. It only communicates with the server.

Server

The heart of the product. It contains:

  • the API consumed by the interface, the web widget, the mobile application, the browser extension and third-party integrations;
  • authentication and roles;
  • AI orchestration: supplier selection, vector search, context construction, agent execution;
  • the database (spaces, conversations, users, settings);
  • background processing: indexing, memory retrieval, Second Brain consolidation, scheduled tasks.

On the workstation, the server also starts the standalone SEO Audit and Prospecting agents. The provided server configuration disables them.

Extraction service

A service dedicated to one thing only: transforming a file or URL into usable text. PDF, Word, spreadsheets, presentations, emails, ebooks, images (character recognition), audio and video (transcription), web pages, code repositories, enterprise wikis.

Separating it from the server has a practical advantage: a large document or a malformed PDF is processed in a separate process, and its failure does not interrupt your conversations.


Interchangeable bricks

Three families of components are chosen by the administrator and can be replaced without affecting the rest:

Brick Role By default
Language Model (LLM) Write the answers, manage the agents To be chosen during onboarding — local model possible
Embeddings engine Translates the text into comparable vectors Integrated engine, running locally
Vector base Stores vectors and retrieves nearby passages Integrated local database, stored on disk

Two additional optional features are included: image and video generation, and speech synthesis/recognition.

→ Supported providers


The journey of a document

  • Deposit. You upload a file to the document manager of a workspace, or you provide a URL.
  • Extraction. The extraction service identifies the format, extracts the text and metadata (title, source, date, author when they exist).
  • Available. The extracted text is saved as an available document. It is not yet searchable.
  • Cutting. When added to a space, the text is cut into fragments of configurable size, with an overlap between fragments to avoid breaking overhanging sentences.
  • Vectorization. Each fragment is transformed into a vector by the embeddings engine.
  • Indexing. Vectors are written in the vector basis, in a namespace specific to the workspace.

The document can then be searched. The same document can be added to several workspaces: it is only retrieved once, but indexed in each one.

A file dropped directly into the conversation follows a different path: it is attached to that conversation as context, without going through the space index.


The path of a question

  • Sending. You type a question in a conversation. If content control is enabled, its rules apply at that moment.
  • Orders. If the message starts with a command (/reset, /image or /img, /video, /note, or a command you have defined), it is processed before everything else.
  • Agent mode. If the message starts with @agent, or if the space is in Agent chat mode and the model knows how to call tools, control passes to the agent and the flow becomes that of the agents.
  • Research. Your question is vectorized and then compared to fragments of the space. The closest ones are retained, within the limits of the number of results and the configured similarity threshold.
  • Context construction. The server assembles: the space system prompt, pinned documents, recovered fragments, your memories (if enabled) and relevant Second Brain elements, then the latest messages of the conversation.
  • Generation. The entire set of data is sent to the model, which responds in a continuous stream—the text is displayed as it is generated. If the PII shield is activated, personal data is replaced with pseudonyms before sending, then restored in the response.
  • Restitution. The response is recorded in the history along with its sources, displayed below the message.

Where is the data stored?

Everything produced by an installation resides in a single storage directory:

  • the application's database; ;
  • the extracted documents; ;
  • vector indices, when the vector basis is local; ;
  • downloaded models, when you run models locally; ;
  • the files produced by the agents;
  • the audit log and the network output log.

Backing up an installation therefore means backing up this directory and the configuration file with its secrets, with the application stopped: copying a database while it's running may be unusable. Without these secrets, the saved vendor keys become unreadable.

→ Security and confidentiality


What comes out of the network, and when

Situation Network output
Local model, native embeddings, local vector database None for your messages and documents
External model provider The messages and context are sent to this provider.
Hosted vector database The vectors and metadata are sent to this hosting provider.
Agent skill «web search» (to be activated) The query is sent to the chosen search engine.
Importing a URL or a repository Request to the requested address
Downloading a local model Once, during the download
Integrated embeddings engine Once, on the first document import (approximately 90 MB)
Desktop application Checking for updates at launch and then every 6 hours, except in Strict Sovereign mode
Anonymous telemetry None, unless an administrator activates it.

The Privacy and Data page, at the bottom of the settings menu, displays the model provider, embedding engine, and vector base of your instance. Actual output can be viewed in Wiven Console → Sovereign Mode:

  • Surveillance, the default setting, records every external connection (host, address, port, never the content) and does not block anything;
  • Strict blocks everything that is not on the whitelist; the provided server configuration applies this setting. Remember to allow what you need, such as downloading a template or connecting to Outlook (see Installation);
  • Disabled, it neither controls nor records anything.

This control applies to outgoing connections from the server and the extraction service.


Governance, compliance and business integration

The services described above define where the processing takes place. Four additional layers are added, and these are what make the difference in operation.

Governance roles

Three role levels in multi-user mode (WivenLLM Business license): administrator, manager, user. Guests are also included if guest access is enabled, and there are two roles within an organization: organization administrator and member.

Accounts are local by default. Authentication can be done through the existing corporate directory: single sign-on (Microsoft, Google, OIDC), Active Directory, and Entra groups that decide who logs in, with what role, and where.

Each role opens a distinct perimeter of workspaces, documents, and settings. Administrators and managers see all spaces except those blocked by a wall (Wiven Console → Walls).

Traceability and nLPD

WivenLLM does not generate a processing log; it provides the necessary elements to maintain one. The audit log is time-stamped, includes user names in multi-user mode, and can be exported as a CSV or JSONL file. The sovereign mode generates a signed monthly report of network activity. The solutions are designed for compliance with the nLPD (New Data Protection Act); however, compliance also depends on the purposes of processing and the client's own records, and Wiven provides support in this area rather than guaranteeing compliance on the client's behalf.

Agent's perimeter

An agent works in the workspace where you summon them and queries the documents within that workspace. Their skills are activated for the entire instance by an administrator, not agent by agent. Some skills extend beyond the workspace: conversation search and Second Brain access your data across all your workspaces. Email and calendar skills only use the accounts of the person making the request, and sending, replying to, or deleting an email requires confirmation each time.

The client decides what is excluded from this scope: in Sovereign Surveillance or Strict mode, external connections covered by the control are logged. The table above indicates, situation by situation, what this covers at the network level.

Protections to activate

Protection Or Default state
PII Shield Wiven Console → PII Shield Disabled
Sovereign Mode Wiven Console → Sovereign Mode Monitoring (Strict in the provided server configuration)
Content control Settings → Security → Content Control Disabled
Unsourced claims Wiven Console → Response Control Disabled
Approved models Wiven Console → Approved Models Not taxed
Walls Wiven Console → Walls None
AI Transparency Wiven Console → AI Transparency Active

Business connectors

Integration is done with existing tools, without replacing them. Version 1.17.2 provides:

  • document sources: websites, Confluence, DrupalWiki, GitHub, GitLab, Gitea, Obsidian, Paperless-ngx, YouTube;
  • Agent skills: Outlook, Gmail, Google Calendar, Microsoft Teams (directory, availability, tickets), SQL databases (MySQL, PostgreSQL, SQL Server);
  • Extension points: developer API, MCP servers, agent feeds.

Connections to accounting, ERP and DMS software (Abacus, bexio, SAP…) are made by Wiven through these extension points.

→ WivenLLM security · Integrations