Documentation

In-app help

Every page has a help button in the masthead. It opens one popover with three ways to get unstuck: search the documentation and the page itself, ask a question, or take a guided tour.

Search covers three local sources:

  • documentation sections and their headings
  • glossary definitions for the jargon a page uses
  • the named controls on each tool

The controls on the page you are on are listed before you type, and a matching control is labelled as a page result rather than a documentation result.

It does not index the model or GPU catalog. Those are live facts owned by the API, and copying them into a search index would let the answer go stale. Read them where they are shown: the GPU explorer for hardware, or the tool you are already using for the models and systems it offers.

A control result opens the page it lives on. When that control has a documentation section, a Docs link beside it opens the explanation directly.

Ask

The question box explains the documentation: what a page is for, what a control changes, and what a term means. It reads the same sections as search, so it answers from the documentation this site ships with rather than from the live API.

It does not read the model or GPU catalog, current pricing offers, or the results of a calculation. For a specific model, GPU, or configuration, open the page that shows it. It also never runs sizing: a sizing question ends with a link to the right tool, carrying the values you named, so you press Calculate (or Compare) and the result appears in the full page. A value only travels when the page's link contract accepts it, and the link says which fields it sets and which named values it left out, with the reason: a value the contract refuses, a question that offers two alternatives, an input the destination page does not have, or an input the selected goal does not use. The model and GPU system come from your settings unless the question names its own, in which case the question wins. Anything else the question did not name keeps the value the page already had. The answer cites the documentation sections it used. If a number cannot be traced to a source or to your question, the panel shows the sources without the prose rather than presenting an unverified value.

Ask is optional and can be switched off by the deployment. When it is off, the button, search, and tours still work.

Tours

Tours are manual. Nothing opens one automatically, and no "seen" flag is stored, so a tour starts only when you choose Take a tour. The tour follows the page's current layout, skips a step whose target is missing, and closes on Escape, Skip tour, or Got it on the last step. It never changes your inputs.

Documentation

The documentation index is at /docs, and every page has a stable address such as /docs/settings#sharing-a-view. A deep link carries the simulator, backend, and backend version that produced the view; opening one never changes your saved settings. A link writes only the values that differ from the page's starting state, so a partial link keeps your own values for whatever it leaves out. See Settings and simulators for what happens when the link's simulator differs from yours and for the fields a link always carries.

Configuration

Self-hosted deployments opt in to the question box with:

Variable Meaning
ASSISTANT_ENABLED Set to true to show the question box
ASSISTANT_MODEL Model id served by the inference endpoint
ASSISTANT_INFERENCE_URL OpenAI-compatible endpoint that answers questions
TRITON_INFERENCE_URL FIL classifier endpoint; a legacy fallback for the above

Without ASSISTANT_ENABLED, ASSISTANT_MODEL, and an inference endpoint the assistant reports that it is unavailable. Search, tours, and the documentation are build-time and local, so they need no configuration.

The assistant and the FIL classifiers are separate services. The FIL endpoint serves /v2/models/<name>/infer and cannot answer a chat question, so ASSISTANT_INFERENCE_URL points at the assistant's own server. TRITON_INFERENCE_URL is only read when the assistant endpoint is unset, which keeps a deployment that predates the split working as it did before; setting both always prefers the assistant endpoint.

The assistant also offers the model read-only catalog, pricing, and memory lookups as tools. An endpoint that does not support tool calling ignores them, and the answer then comes from the documentation alone; either way, a number in the answer is checked against a documentation section or a tool result before the prose is shown.