Grounded answers

Prompter is not “trained on” the Cratis documentation - no model ever memorizes it. Instead it uses retrieval-augmented generation (RAG): for every question it looks the answer up in the documentation and has a language model phrase what it found. That distinction drives everything you experience as a user.

Answers are looked up, not remembered

When documentation changes, Prompter’s knowledge updates within minutes - the changed pages are re-indexed and the very next question retrieves the new content. There is no retraining, no release, no lag between what cratis.io says and what Prompter says.

Citations are the contract

The language model is instructed to answer only from the retrieved passages and to reference which ones it used - those pages become the Sources: links on every answer. If an answer ever surprises you, follow the links: the answer should be recognizable from the source. That’s your verification path, and it’s why answers without sources don’t exist.

Refusals are a feature

Before answering, Prompter scores how well the best retrieved passages match your question. Below a threshold, it refuses - “I couldn’t find anything in the Cratis documentation that answers this with confidence” - rather than letting the model improvise. A wrong-but-confident code snippet costs you more than an honest “ask a human,” so honesty wins by design.

Refusals also have a second life: they are logged (see Privacy for what exactly is stored) and reviewed as documentation gaps. If Prompter keeps refusing a reasonable question, that’s a page someone should write - and the community’s 👎 reactions work the same way.

What “current” means precisely

Prompter’s knowledge is the published documentation - nothing more. A feature that shipped without documentation is invisible to it, exactly as it is to a human reader of cratis.io. The knowledge base refreshes automatically when documentation deploys, with a nightly sweep as a safety net.