Applied AI
Every demonstration queries an open-weight model served on chosen infrastructure. Every answer then shows what produced it: the excerpts read, the model queried, the compute time.
Document assistant
This assistant knows nothing beyond this site: its pages and its articles. It cites its sources and declines the rest. It’s exactly the setup an organisation applies to its own documents, except that the corpus here is public.
For example
Four-language correspondence
An email in Zurich dialect, a school register in Romansh, a reminder from Ticino, an invoice in standard German: correspondence in a Swiss company arrives in four languages. Paste one in and read the translation.
The document is identified first, then handed to the model that reads that language best: the Swiss model for Romansh and the Alemannic dialects, the fast model for German, Italian and French. Amounts and deadlines always come from the extraction step, which copies them without ever recomputing them.
Research watch
A topic in plain language, and the assistant queries OpenAlex, the open catalogue of the world’s scientific literature, then brings back recent publications reduced to what they seek, how, and what they find.
Or try
The catalogue is almost entirely in English, so a topic phrased in another language finds nothing there. The topic is therefore turned into a bibliographic query first, and the summaries come back translated. Two invisible steps for whoever is searching, which is precisely the point. Matches that are too weak are dropped rather than summarised: better to show nothing than an off-topic paper.
What frames the demonstration
Nothing is kept
The questions asked here and the answers produced are stored nowhere.
A spending ceiling
Calls are capped per visitor, per demonstration and for the site as a whole: a day’s cost is bounded in advance.
Access reserved to this site
Requests from elsewhere are refused, bursts blocked and quotas counted per address: these demonstrations can’t serve as a free gateway to a model.
A closed scope
The assistant answers only from this site’s content. Beyond that, it declines.
Three examples, not a catalogue
These three demonstrations run entirely inside a browser, and they do not cover every use. Elsewhere the public corpus becomes your procedures, your case files, your contracts or your instrument readings, and the task becomes yours: extracting, reconciling, checking, among others. What doesn’t change is the rule: your documents chunked and indexed, a model that answers only from them, and verifiable sources under every answer.
Discuss itA first conversation, no commitment
Which model for which task
The available models were tested on the same tasks: answering from the site, reading Romansh and Swiss-German dialect, picking figures off an invoice. None wins everywhere, and that’s exactly what makes the choice interesting.
| Model | Verdict |
|---|---|
| Mistral Small 4 | Kept for the assistant, for figures and for German, French and Italian: nothing invented across twenty checked fields, under a second per document. Dropped from Romansh, where it invents words. |
| Apertus 70B · Swiss | Kept for Romansh and Alemannic dialect, which it alone renders faithfully. Dropped from anything involving figures: it invented a price, then an invoice total, then an IBAN. |
| Qwen 3.5 122B | Dropped. Spends its whole compute budget reasoning before answering. |
| Kimi K2.6 | Dropped. Empty response on long context. |
| Ministral 3 14B | Solid and fast, but stray formatting. Kept as a fallback. |