AI features
Every AI feature is off until an administrator configures a model provider for it. Where a feature touches the editor, you can also turn it off for yourself alone, apart from that instance setting.
An administrator connects one or more providers first, under Settings → AI, then assigns each one a role: chat, embeddings, or vision. A feature only runs when both are true — its own switch is on, and the role it needs has a provider assigned. On a freshly installed server no provider exists yet, so every feature behaves as off no matter what its switch says.
Once a role has a provider, most switches need no further action, because they default on already. The exception is Similar-issues LLM rerank, which defaults off, because it costs one model round trip every time someone opens the similar-issues list.
| Setting | What it does | Default |
|---|---|---|
| Editor AI actions | The editor's AI writing actions, with diff review. | On |
| Semantic search | Meaning-based search in search, similar issues, and the palette Ask mode. | On |
| Issue summarize | The Summarize action on issues. | On |
| Natural language → SLQ | The Ask bar that turns plain language into an SLQ filter. | On |
| Similar-issues LLM rerank | Rescores duplicate candidates and explains each match. | Off |
Two more switches on that settings page belong to other features. The mail intake and attachment storage pages document LLM mail routing and LLM storage routing instead.
Separately from all of that, open Settings → Profile → AI assistance to turn off the editor's AI actions for your account alone. This does not touch anyone else's account. It also does not touch find-similar or summarize, which have no personal opt-out.
Click Summarize in an issue's AI panel. You can also open the sparkle menu on its description and click Summarize issue. Radd builds a digest first: the fields, the description, recent comments, recent history, and — when the project logs time — a time digest. It sends that digest to the chat model. The result is markdown, shown to you, and never stored.
Click Find similar in the same panel, or Find similar issues in the sparkle menu on a description or comment. Radd ranks candidates by meaning when semantic search is on, and falls back to a text match otherwise. Turn on Similar-issues LLM rerank to have the chat model re-score the list and explain each match. This costs one model round trip per open, so it is off by default.
On a view, open the query bar. Switch it to Ask — an empty bar opens in Ask mode by default, and a keyboard shortcut toggles it. Type a plain-language question, such as "open bugs assigned to me." Radd returns an SLQ filter and explains what it built. The server checks that the query compiles before it hands the filter back, and retries once on its own if the first attempt fails.
Select text in the editor. Then open the AI menu. With nothing selected, the same menu acts on the whole document instead. Choose a built-in action, an administrator's preset (below), or type your own instruction. The built-ins are Refine writing, Format as markdown, Summarize, and Fix grammar only. The result streams into the editor as it arrives.
Open the search palette. Start typing. A row reading "Ask: … — search by meaning" appears at the bottom of the results. Click it, or switch directly to Ask mode, to search by meaning instead of by matching words. A search for "emails that never send" can then surface an issue titled "SMTP retries stop after one failure."
An editor transform never writes straight into the document. It streams into a preview band at the top of the editor first. Once the result arrives, it turns into one or more change pairs inside the text, each with its own Accept and Reject buttons. The band shows how many changes remain and offers Accept all and Reject all. The editor accepts no new typing until you decide every change.
When you accept a change, the editor updates the document in front of you. You still need to click Save on the field to keep it — exactly as you would after typing the words yourself.
On an issue page, a summarize or find-similar answer opens in a panel beside the reading column. This is wider than the narrow rail you clicked it from. On a wiki page, and on a comment, the same actions open in a popover instead.
An administrator can publish named prompts under Settings → AI → Editor presets. A published preset appears in every user's editor AI menu beside the built-in actions, labeled with its name. Only the name reaches your browser. The prompt text itself stays on the server.
You are entitled to know what a feature sends before you use it. Every feature reads through your own account first. It can only ever send what you could already see — nobody else's private worklog note reaches a summary you could not open yourself.
- Editor actions send the document or your selection, plus your instruction, to the chat model.
- Summarize sends the fields, description, a budgeted slice of recent comments and history, and a time digest when the project logs time.
- Find similar and semantic search send the issue's title and description, or your search text, so the model can match it by meaning. The optional rerank also sends candidate titles to the chat model.
- Ask a question sends the words you typed, plus the field, type, and label names Radd needs to build a valid query.
Which company, if any, receives this data depends on the provider your administrator configured. Check Settings → AI → Providers to see which one. A self-hosted model keeps the data on your own server. A hosted provider receives it over the network.
Mirrored from project.radd-hq.com on 2026-09-12. Documentation is written there; this copy is regenerated by scripts/publish_wiki.py and hand edits do not survive it.