Skip to content
AppArch AppArch Docs

Archito, the assistant

Archito is the robot in the bottom-right corner. It has two halves, and each one is switched on by the installation’s administrator on its own:

  • Insights - what AppArch has learned from this installation’s own data (smart learning). No AI provider is involved and nothing leaves the installation.
  • Ask - a chat that answers questions about your workspace through your organisation’s AI provider.

Archito appears when at least one half is on; with both on, its panel has an Insights and an Ask tab.

Meeting Archito

Section titled “Meeting Archito”

The first time you sign in after Archito is available, it introduces itself once and shows where it lives. What is Archito? in its panel plays the introduction again.

When Archito has something new, it pops up a short note (Show / Later). Untick Pop up when I have something new in the panel to keep it quiet; the panel itself is always one click away.

Insights

Section titled “Insights”

Insights are cards, each built from your own data and each linking straight to its subject - the merge request, or the list already narrowed to it, never a list to search through.

  • Open longer than usual - the merge request that has been open longest compared with how long similar merge requests take in that project, with the main reason and, for a large change, how long it would usually take if split in two. Links lead to the merge request and to all late merge requests.
  • Good news this week - shown only when the week really was better: merge requests merged faster, or more of them, than in the weeks before.
  • What I have learned - what the merge request model weighs most. It looks at the work, never at people: who opened a merge request is not one of its inputs.
  • What keeps failing - the Git Health and Compliance rules that fail most in this installation, how many projects fail them now, how long a fix usually takes and how often they come back, with a link to the failing projects.

Late, draft and stale

Section titled “Late, draft and stale”

A merge request is late when it has been open much longer than the estimate for it. A draft (Draft:, [Draft], WIP:) is never late - it is not ready for review. A merge request open for more than 30 days is stale: old rather than late, and never counted late. The badges, the lists, the dashboard card, Archito and the e-mail notices all use this one rule.

Workspace owners and administrators get an e-mail when a merge request runs well past its estimate, unless the workspace’s notification rules say otherwise.

Ask

Section titled “Ask”
  • Archito answers from what AppArch has measured: your projects, security findings, recent activity and Git health. A question outside that is answered with “I cannot answer that from AppArch data” rather than a guess.
  • Every answer names the provider and model that wrote it. When the administrator offers several models, you pick one per question.
  • The full-page chat is at /assistant. Clear empties the screen; the transcript is kept, because usage is counted from it.
  • When the administrator allows it, ArchPilot and ArchScan findings carry a suggest a fix button. The finding’s text goes to the provider, and the answer is a suggestion to review, not a verified fix.

For administrators

Section titled “For administrators”

Administration → Settings → Assistant sets up Ask:

  1. Read the notice: this is the one feature that sends data outside the installation - the questions and the figures used to answer them.
  2. Accept the AI processing agreement. Neither the chat nor fix suggestions can be switched on before it is accepted, and the record stays even if the assistant is switched off later.
  3. Pick the provider, enter the organisation’s API key (stored encrypted) and the models on offer; the first model is the default.
  4. Switch on Offer the assistant to users and, separately, Allow fix suggestions on findings.
  5. Set the token budgets: a per-workspace ceiling for fair sharing and an installation ceiling for the total. 0 means no limit; the day resets at midnight UTC, the week on Monday.

Administration → System → Assistant usage breaks the spend down by workspace, person, model and question type.

Administration → System → Smart learning runs Insights:

  • The Smart learning switch. Off, the expected-time badges and notices fall back to a simple estimate (how long three in four of the project’s merge requests took), and the learned model is kept for when it is switched back on.
  • The merge request model’s state - Learning (a first model needs enough merged merge requests), In use, or Simple estimate in use when the newest model was not more accurate - with what it pays attention to, what training costs and the latest attempts. Training runs once a day at 03:00 UTC.
  • Git Health and Compliance patterns - statistics rather than a model, learned in the same nightly pass.
  • Delete learning data removes what smart learning stored on this installation, for good.

Insights are computed on your server from your own data; nothing is sent anywhere for them.