Skip to content

FAQ

Three possibilities: you weren’t in one of its surfaces (mention it, use /ask, the ask channel, or a help-forum thread - it ignores everything else by design); you hit the per-user rate limit (a few questions per ten minutes - it tells you when that happens); or it refused because the documentation doesn’t cover your question - which is deliberate.

React 👎 and say so in the channel - a human picks it up, and the verdict feeds the quality measurements. Every answer links its sources, so you can check where the wrong claim came from; if the documentation itself is wrong, that’s a valuable bug report.

Within minutes of documentation changes - the knowledge base re-indexes automatically when cratis.io deploys. It only knows published documentation: undocumented features are invisible to it.

Claude (Anthropic) writes the answers from retrieved documentation passages; Voyage AI computes the embeddings used for retrieval. Neither trains on your questions - see Privacy.

No. It only processes messages on its answer surfaces, and it stores a one-way hash of your identity - never your username. Privacy lists exactly what is kept and for how long.

It’s MIT-licensed on GitHub and points at any docs site that publishes a sitemap with markdown mirrors - see Run Prompter locally. It’s built for the Cratis community first; PRs are welcome, but there are no support promises.