Live demonstrations
Applied AI
Try a document assistant, correspondence translation and scientific literature search. The AI models run on Infomaniak’s Swiss infrastructure. Result panels show the models used, response time and text units processed.
Document assistant
This assistant searches the site’s pages and articles and cites the passages it uses. Its instructions ask it to acknowledge when the sources do not support an answer. The same approach can be applied to internal documents with appropriate access controls and testing.
For example
Swiss correspondence translation
Paste a message in standard German, Swiss German, French, Italian or Romansh to obtain an English translation. The demonstration also extracts requested actions, amounts and deadlines.
The pipeline identifies the document language, then extracts the requested amounts, dates and other information. Translation uses Apertus for Romansh and Swiss German, and Mistral for the other languages selected in the trials. Amounts and deadlines are copied without calculation. Each extracted value can be checked against the original document.
Scientific literature search
Enter a research topic to search OpenAlex, an open catalogue of scientific publications. The demonstration summarises the objective, method and results of recent publications, with links to the sources.
Or try
The demonstration turns the topic into English bibliographic search terms, then queries OpenAlex. Summaries use the information available in the catalogue. Weak matches are filtered out and duplicate versions are grouped before the selected results are presented. Links provide access to the original publications.
Usage conditions and limitations
Handling of requests and responses
The demonstration code does not store question or answer content on the site. Technical counters associated with a hash of the IP address limit requests. Text is sent to Infomaniak for model processing under that provider’s conditions.
A spending ceiling
Quotas limit requests per visitor, per demonstration and across the site. They help control API usage and operating expenditure.
Controls against misuse
The demonstrations check request origin and apply limits to closely spaced or numerous requests. These controls reduce unintended use; they do not guarantee that circumvention is impossible.
An assistant grounded in site content
The instructions restrict answers to retrieved site passages and require insufficient evidence to be acknowledged. Citations support verification; their presence alone does not establish accuracy.
Three examples, not a catalogue
These demonstrations are available through the browser without installation. AI processing runs on the infrastructure described above. Comparable applications can work with authorised internal documents to retrieve information, translate text or prepare a synthesis. Search evaluation checks the retrieved passages and applied access permissions. Translation and extraction are compared with the original document. Scientific literature search is checked against the cited publication references.
Discuss itA first conversation, no commitment
Trial results and model selection
The first assistant trials took place on 25 July 2026, followed by translation and extraction trials using correspondence and invoices. Each task was compared using the same documents and instructions. The results describe those limited trials with the model versions, documents and response budgets used. They are not a ranking for every use. A token is a unit of processed text; token counts do not directly correspond to word counts.
| Model | Observations |
|---|---|
| Mistral Small 4 | Selected for the assistant, amount extraction and German, French and Italian translation. Extraction trials recovered all twenty checked fields without invention, with response times below one second per document. The Romansh translations tested contained errors. |
| Apertus 70B · Swiss | Selected for the Romansh and Swiss German texts tested. Excluded from figure extraction after errors involving a price, an invoice total and an IBAN. |
| Qwen 3.5 122B | In the trials, the model consumed the response budget on reasoning without producing usable text. It was therefore not selected for these demonstrations. |
| Kimi K2.6 | Long-context trials returned an empty answer in the configuration used, so the model was not selected. |
| Ministral 3 14B | Trials produced fast, usable answers with unwanted formatting text. Retained as a possible replacement. |