When a pull request merges, Autopilot works out what it means for your help center in three steps, and each step has its own model. Owners and admins pick those models on every plan. On Business, owners can also tune how sure Autopilot must be before it suggests a change, how much it looks up and how much it thinks.
Autopilot's three steps
Autopilot decides: reads the merged pull request, looks through your code and your help center, and decides which articles to fix and which new ones you need. Drafts count as articles too. Each suggestion comes with how sure Autopilot is.
Autopilot double-checks: when it suggests a new article, Autopilot decides again from scratch what the pull request means for your help center, and that answer is the one used. Fewer duplicates of articles you already have get written.
Autopilot writes: writes each fix and new article, looking things up in your code as it goes, and the pull request's summary, its changelog line and its updates on posts. Verify uses this model too.
What then waits for your team, and what goes out by itself, is up to Settings → Autopilot: see Set up Autopilot.
Pick each step's model
Open Settings → Models. The page is called AI models.
Under Autopilot decides, Autopilot double-checks and Autopilot writes, choose a card. Since each step looks things up as it goes, only models that can use tools are offered for them.
Fast: the default, Tuti's model for that step. If Tuti changes its default, the step follows.
Balanced and Best, where offered: larger models that can do better, and cost more.
Off, for Autopilot double-checks only: "No double-check: cheaper, more duplicates".
Custom model: any other model id.
Choose Save. "AI models saved" confirms it.
The other rows on the page are for your other AI features: see Choose AI models.
What each step costs
Each card shows roughly what one use costs in AI credits:
Autopilot decides and Autopilot writes: per merged pull request.
Autopilot double-checks: per double-check, which only runs when Autopilot suggests a new article.
The estimates are for a typical pull request. A large one costs more, and so do more lookups or more Thinking (see below). What Autopilot really spent, its three steps and Verify together, shows under Autopilot in Usage, at the bottom of Settings → Billing.
When to pick a bigger model:
Autopilot misses articles a pull request makes wrong, or suggests fixes you don't need: a bigger model for Autopilot decides.
Its new articles repeat ones you already have: keep Autopilot double-checks on, or give it a bigger model. If you rarely get new articles, Off saves credits.
You rewrite a lot of its fixes and drafts before approving them: a bigger model for Autopilot writes.
Open Advanced tuning
Open Settings → Models, then Advanced tuning, under Autopilot's rows.
It's on the Business plan. On other plans the page says "Advanced tuning needs the Business plan.", with See Billing.
Owners change it. Admins can open it and read it: "Only owners change these." Members can't open it.
Each value starts as Tuti's default, and what you change is for this project only. Pull requests analysed after you save use the new values; updates Autopilot already suggested stay as they are.
If you're not sure, leave a value at its default. Change one value at a time, then see how the next few pull requests go, and how How it's doing in Settings → Autopilot moves, before you change another.
What each value does
Value | Default | Range |
|---|---|---|
Fix an article from | 50% | 0% to 75% |
Write a new article from | 30% | 0% to 80% |
Double-check a new article from | 20% | 0% to 100% |
Lookups before deciding | 12 | 1 to 50 |
Lookups before writing | 10 | 1 to 50 |
Thinking | Off | Off, Low, Medium or High |
Change shown up front | 20,000 | 1,000 to 500,000 |
Fix an article from: how sure Autopilot must be that a pull request makes an article wrong or incomplete before it suggests a fix. Lower it if articles go out of date without a fix; you'll dismiss more fixes you don't need. Raise it if you dismiss many of its fixes; more out-of-date articles slip by.
Write a new article from: how sure Autopilot must be that you need a new article before it writes one. Lower it if features ship without an article; you'll get more duplicates. Raise it if it drafts articles you don't need; more gaps are left.
Double-check a new article from: a new article Autopilot is this sure of or more is double-checked. At 20%, below Write a new article from, every new article it would write is double-checked. Lower it to double-check more, which costs more. Raise it to save credits, and more duplicates get through. It does nothing while Autopilot double-checks is Off.
Lookups before deciding: how many times Autopilot may look things up in your code and help center before it must decide. Raise it if you have a large product or many articles and it misses some; it costs more and takes longer. Lower it to save credits, and it may miss articles.
Lookups before writing: the same, for writing each fix and new article. Raise it if its text gets details of your product wrong; it costs more. Lower it to save credits, with more to fix by hand.
Thinking: how much Autopilot thinks before it decides, and before it double-checks. Low, Medium or High decide better on hard changes, and cost more and are slower. Off is cheapest and fastest.
Change shown up front: how much of the pull request's change Autopilot reads at the start, in characters. Files past it are only listed. For your docs, Autopilot looks up the rest by itself, but which posts the pull request completes or moves forward, and its changelog line, are decided from this much only. Raise it if your pull requests are large and Autopilot misses their posts or changelog lines; each pull request costs more. Lowering it saves credits on large pull requests.
Save, and go back to the defaults
Choose Save to save every value at once. "Advanced tuning saved" confirms it, and the page says who changed it last, such as "Last changed by Maya 2 hours ago".
A value you changed says what it was changed from, such as "Changed from 50%", with Reset beside it. Reset puts Tuti's default back in its field, and Reset all does it for every value you changed. Neither saves by itself: choose Save. A value that's back on the default follows Tuti's default from then on.
See which pull requests used your values
When a pull request was analysed with any value you changed, its page in Autopilot → Activity says "Ran with this project's tuning", and names the values, such as "Ran with this project's tuning: Fix an article from, Thinking". Owners and admins can click it to open Advanced tuning.
After a move off Business
Your values are kept, but not used: pull requests are analysed with Tuti's defaults. Advanced tuning says "Advanced tuning needs the Business plan. This project's values are kept but not used; back on Business, they're used again." Your picks under Settings → Models keep working on every plan.