This page is machine-translated from French. Read the French original.
Frequently asked questions
Confidentiality
Is my data leaving my infrastructure? Only if you connect an external service. Conversations and data remain local as long as the template provider and embedding engine are local. In a fully local configuration, documents, indexes, conversations, and templates remain on your server. Any content output results from your configuration—external template provider, hosted vector database, web search engine, MCP server. By default, this output is logged and not blocked; strict sovereign mode blocks it. The Wiven console → Sovereign mode shows what is actually output, and the Settings → Privacy and Data screen displays the configured providers.
Does WivenLLM work without an internet connection? Yes, in a fully local configuration. A connection is only required to download a model for the first time, import a remote source, or use an external service. The desktop application also checks for updates, without transmitting any content, except in strict mode.
Is the data used to train a model? Not on the WivenLLM side: neither your documents nor your conversations are sent to the publisher, and telemetry is disabled by default. If you connect an external provider, their terms apply—check this carefully, as business plans often differ from consumer plans.
Is personal data masked before being sent to an external model? Not by default. The PII shield is disabled until you enable it. Once enabled, it pseudonymizes the data sent to the language model, but agent tools receive the real values, and embeddings are not covered. Its detection relies on rules and may miss data.
Does WivenLLM protect against prompt injection? No system based on language models is completely immune to such threats. A document, a web page, or an email can contain instructions that the model will follow. Safeguards are limited to instructions given to the model and human approval of the agent's actions.
Do my colleagues see my conversations? Regular users, no. Administrators and managers, on the other hand, can view and export the instance history — this is necessary for operation, and users must be informed of this.
How long are conversations kept? Until they are deleted: no retention period can be set for conversations or the event log. The audit log is kept indefinitely, unless you specify a retention period. AUDIT_RETENTION_MONTHS. Only guest accounts are automatically deleted upon their expiration.
Can multiple companies share an instance? This is not recommended. Organizations group people and spaces, but do not isolate businesses from one another: storage and administrator/manager roles are global. Deploy one instance per business.
Where is the data physically stored? The instance is installed where it resides: on your workstation, your server, or in the datacenter of the hosted instance. There is no central storage. External services you connect to store what they receive according to their own terms and conditions.
Costs
How much does it cost to use? In a fully local setup, there are no usage fees: the cost is that of the hardware and its operation. With an external provider, billing is proportional to the volume exchanged.
What causes the bill to increase with an external supplier? The pinned documents (injected with each message), a long history, a large number of recovered fragments, and above all the agents—each step is an additional call. There is no cache of responses: the same question is sent back to the template every time.
How to reduce costs without compromising quality? A template router: a default economy model, a powerful model for the cases your rules define. It doesn't know the price of templates: you choose them. Also, reduce the history and only pin what's strictly necessary. With Anthropic, the prompt cache can be enabled during installation.
Can I cap the spending? Not directly. You can limit the number of messages per user over 24 hours, with administrators being exempt. There's no token or cash budget: set it with your provider.
Quality of responses
The assistant is making up answers. What to do? Three actions, in this order: switch the workspace to Query mode; lower the temperature; specify in the system prompt that any information missing from the documents should be flagged as such. Wiven Console → Response Checking can also add a «To be verified» note to responses containing numbers or phrases not found in the passages.
He cannot find information that is nevertheless present. Verify that the document is added to the space and not just imported. Then lower the similarity threshold and increase the number of fragments. Finally, check that the text has been extracted correctly—a scanned PDF can be unreadable—and that the embeddings template is appropriate for the language: the default one is designed for English.
He quotes unrelated passages. The similarity threshold is too low, or you are requesting too many fragments.
The responses deteriorated suddenly. A change in the model or search parameters. If the embedding engine or vector database has changed, all indexes have been reset: documents must be re-added to each space.
He's responding in the wrong language. The assistant follows the language of your last message, even if the documents are in a different language. The interface language does not determine the language of the replies. If a small model deviates from this, try a more capable one. A prompt in the system space, such as «Always reply in French,» is not guaranteed to override this rule.
Models
Which model should I choose? Consult the hardware diagnostic: it indicates for each model in the catalog whether it is suitable for your machine. If in doubt, start small and gradually increase the size. → Choosing your models
Can I change the model later? Yes, there's no consequence: the documents remain indexed and the conversations readable. However, changing the embedding engine erases all indexes and removes the documents from the spaces, where they must be added again.
Local responses are slow. Normal without a graphics card. Get a smaller model, or switch this space to an external provider, accepting that its data will be leaked.
Can I mix local and cloud storage? Yes. The model is chosen per workspace, and a router can arbitrate message by message. The router has no confidentiality criteria: it does not guarantee that a sensitive topic will remain on the local model.
Documents
How many documents can I index? The integrated local database easily handles corpora of tens of thousands of fragments. Beyond that, an enterprise database is justified.
Can a document be used in multiple spaces? Yes: it is extracted once and indexed in each space where you add it.
What happens to a deleted document? When removed from a space, it loses its vectors in that space but remains in the library. When deleted from the library, it disappears from all spaces and its extracted text is erased.
My scanned PDFs are not being read properly. Character recognition (CR) depends on scan quality and works in English by default: specify your languages in the installation settings. A tilted or low-resolution scan will produce error text. Rescanning correctly, or using a dedicated recognition tool beforehand, is the only real solution.
Agents Verification Status
Can an agent send an email on their own? No. Sending, replying to, and deleting messages require approval each time, even if the agent is authorized to act alone, and are denied when no one can approve (scheduled task, API). Outlook is connected in read-only mode unless drafts and sending are enabled. And an unenabled skill does not exist for the agent.
Can an agent modify my files? Only if the corresponding skill is enabled, and in the authorized directory. Each elementary operation can be disabled separately.
My agent isn't using the right tools. The pattern is almost always the same: operating tools is more demanding than conversing. Assign the agent the most capable model you have available. Then verify that the required skill is enabled: this setting applies to the entire instance, not per space.
Why is an agent so slow? It makes several calls to the model, one per step. A ten-step task takes about ten times longer than a simple question.
Exploitation
What needs to be saved? The storage directory and the configuration file. Nothing else — the application will reinstall. Try a system restore.
How do I update? Save your system, install the new version while keeping the storage directory, and then restart. The necessary migrations will be applied at startup. The desktop application offers the update in Settings → Update; in strict mode, use the offline update package.
How many simultaneous users? It depends primarily on the model and hardware. A local model typically processes requests one after another; an external provider absorbs competition much more easily. A Business license also limits the number of accounts.
Can I migrate from a workstation to a server? Yes. The content fits in the storage directory and the configuration file: moving them together moves the instance.
A user cannot see a workspace. He is not a member, or a wall is blocking his access. Add him from the space settings, Members tab, and check Wiven Console → Walls.
One more question?
- The application's Help screen contains a built-in guide, available without a connection.
- The «Contact Support» link, at the bottom of the settings menu, writes to the support address defined by your administrator.
