Try Onyx yourself

Three live AI-search demos over three different kinds of material — a website, a published report and one fishery's catch data. Ask them anything, see the citations, and watch them decline what they can't ground.

This is the chat widget we deploy for clients. Three assistants are running below, each over a different body of material — because what the corpus looks like changes what the assistant is good at. Ask one a question and watch what comes back, and just as importantly what doesn't.

What to expect

Ask in plain language, the way you'd ask someone at a counter. You'll get an answer with the source cited underneath it, so you can check the claim against the document rather than taking the model's word for it.

Then ask one something it has nothing to say about — the weather, say. It won't reach out to the open web for you; it only ever looks inside the material it was given, and it will tell you when the answer isn't in there.

The chat opens in a panel over this page. Close it and reopen it and your conversation is still there.

Two practical notes. There's no sign-up — the demos take anonymous questions. And the machine sleeps when nobody is using it, so the first question of the day can take a few seconds longer than the rest.

Three to try

This website

The first is indexed over this site — services, case studies and the blog. Nothing was prepared for it: it reads the same pages you are reading now.

What it shows: ordinary web content with nothing done to it beforehand. No structured data, no curation. This is where most organisations start.

Try asking:

  • What is Inka?
  • What does PretaGov do for government clients?
  • Which projects involved a design system?
  • How does PretaGov approach accessibility?

The World Happiness Report

The second is over editions of the World Happiness Report, including their data tables. A published report is a harder corpus than a website, because the answer people actually want is usually a number in a table rather than a sentence in a paragraph.

What it shows: TableRAG. Ordinary retrieval handles a table badly — it finds the page the table sits on and guesses at the number. Our TableRAG pathway computes the answer from the table itself: computed answers, not retrieval guesses.

Try asking:

  • Which country ranked highest in the latest report?
  • What was Australia's score, and where did it rank?
  • How has the United Kingdom moved between editions?
  • Which countries score highest on social support?

One certified fishery

The third is deliberately narrow: a single MSC-certified fishery — Alaska flatfish — and its catch-data tables. One fishery, not the whole certification programme.

What it shows: TableRAG again, on a much narrower corpus. Catch figures are spread across tables on several pages, so the assistant has to read across rows rather than quote a sentence — and it still tells you which page each figure came from.

Try asking:

  • What was the reported catch by species?
  • Which species had the largest reported catch?
  • What is this fishery certified for?
  • What gear types are used?

Want to see this over your own content? We run fixed-cost pilots on the material you already publish — usually three weeks from engagement to something you can click.