OrcAI Assistant

The OrcAI Assistant is OrcAI's built-in helper and not a bot. Nobody owns, edits or shares it, and an admin of your organisation chooses its model. Every member can use it, whatever their role.

Opening the AssistantCopy Anchor Link

Open OrcAI Assistant in the sidebar, or go to /app/assistant. The page shows a composer for a new conversation and your earlier conversations in the active organisation, most recently active first.

Sending the first message starts a conversation and opens it at /app/assistant/<id>. The composer shows the Assistant model as read-only text where a chat shows its model picker. Only an admin can change it.

Questions about OrcAICopy Anchor Link

Ask the Assistant how something in OrcAI works and it reads these docs before it answers. An answer cites the docs page each statement came from and links the app page where you act, such as Bots (/app/library/bots). A docs citation shows the page's title, and opening it shows the title with an Open page action. Copying an answer writes a docs citation as the cited text followed by the page's title and address in parentheses. Docs pages and app pages open in a new tab without asking, so the conversation stays open. The work log of an answer lists the docs pages the Assistant read, each linked. Like every citation, a docs citation promises that the Assistant read the page in this conversation, not that the page supports the sentence.

The Assistant says so when the docs do not cover something, and it does not guess permissions, limits or settings. It knows who you are (your name, email and user id), your role in the active organisation and what that role may do, so it tailors answers to you. It also knows today's date in your browser's time zone, or in UTC when the browser sends none. When a task needs a role you do not have, it says who can do it. What you claim about your role in the conversation changes nothing.

The Assistant answers in the language you write in. The docs are in English, so the names it gives for buttons and pages are the English ones.

If the instance has no docs site configured, the Assistant says that the docs are unavailable instead of answering from memory. Docs citations in earlier answers then show and copy the page's title only.

Working in OrcAICopy Anchor Link

The Assistant can also do things for you in the active organisation. It acts as you, with your own permissions, so it can do exactly what you could do on the pages yourself and nothing more. When something is not allowed, it tells you that the action was denied and why.

What it offers depends on your organisation role. Every role can ask it to find bots, behaviours, repositories and assets you can read, to show how a bot is set up, and to list the groups you can see.

Members, managers and admins can also have it set up bots: create and edit bots, behaviours and repositories, link a behaviour, repositories and features to a bot, find the organisation's chat models and recommend or require one for a bot in the active organisation (it reports how many of the bot's audience a quota pool covers), find what the Catalogue offers, publish a bot, ask you to upload files, edit asset metadata, generate asset metadata with the utility model, retry a failed asset, and delete bots, blocks and assets. It also handles sharing: it lists who can access a bot, block or asset, finds people and groups, and grants or removes access on bots, blocks and assets you manage. Listing in the Catalogue, which opens a preview to everyone on the instance, stays in the access dialog, and the Assistant does not change it.

Managers and admins can also have it add bots, behaviours and repositories from the Catalogue to the organisation (a bot with its behaviour and repositories too, when asked) and remove them again, and set the model of a bot the organisation added. It shares what it added for use only, like the access dialog. They can also have it do people work: create, rename and delete groups and change their members, list organisation members, change roles, remove members, and invite people or withdraw invitations. Inviting skips members and people already invited, as on the Members page, and the Assistant says whom it skipped. Only an Admin can give or change the Admin role, through the Assistant as on the Members page.

Working on bots follows the organisation role only, so a viewer who holds Can edit on a bot edits it in the bot editor, not through the Assistant.

Each action appears in the answer's work log as one line, such as "Delete bot “Citation check”" or "Invite 3 people as members". A failed action shows as failed with the reason.

Bigger edits, such as a bot's rich description, a behaviour's content or a repository's retrieval settings, belong in the editor. The Assistant links the bot editor or the block's editor for them.

The Assistant can answer with structured elements such as checklists, reviews and matrices, as described under Generative UI.

PlansCopy Anchor Link

Before the Assistant creates bots, behaviours, repositories or groups, it proposes a plan: a short title and the steps it means to take. When something is unclear, such as a new bot's purpose, audience, behaviour, repositories or features, it first asks you questions, with its suggestion as the first option, and asks again when your answers raise new questions. Once nothing that matters is open, it proposes the full plan. Changes to things that already exist, such as renaming a bot, and questions need no plan.

The plan sits above the input, with a tally such as "2 of 5 done". Each step shows whether it is pending, in progress, done or skipped, with its specifics below its name, and the Assistant ticks the steps off as it works. Select the plan's header to collapse or expand it, and it collapses by itself once every step is done or skipped.

A proposed plan waits for your answer:

• Approve plan lets the Assistant start. From then on it can create things for the rest of the conversation.
• Revise asks what should change. The Assistant adjusts the plan and proposes it again.
• Writing a new message instead also counts as asking for changes, and the Assistant does not create anything until you approve a plan.

Each plan update also appears in the answer's work log as one line, such as "Plan “Set up the citation bot”: 2 of 5 done". An approved plan does not replace confirmation cards, so destructive and access-changing actions still ask first.

Confirming actionsCopy Anchor Link

Actions that destroy something, take access away, widen access, change a role or send something out pause the answer and show a confirmation card first:

• Deleting a bot, a block or an asset (Delete)
• Removing someone's access to a bot, block or asset, removing people from a group, removing members from the organisation, and removing something added from the Catalogue (Remove)
• Withdrawing an invitation (Withdraw)
• Granting access to a bot, block or asset (Grant)
• Generating metadata for assets (Generate)
• Adding people to a group, and adding something from the Catalogue (Add)
• Changing a member's role (Change)
• Inviting people (Invite)

The card names the action and offers its button and Cancel. Choosing the button runs the action and the Assistant continues its answer. Choosing Cancel leaves everything as it was and marks the action Denied. Writing a new message instead of answering also leaves everything as it was and marks the action Skipped. When an answer shows several cards, the Assistant continues once every card is answered.

Creating and editing, linking blocks and features to a bot, attaching assets to a repository, retrying an asset and publishing a bot run without a card. Creating needs an approved plan first.

If you answered every card but the page reloaded before the Assistant continued, the answer shows a Continue button that picks it up again. If the page reloads while some cards are still unanswered, the answers you already gave are lost, so answer those cards again.

UploadsCopy Anchor Link

When you want to add files, the Assistant shows an Upload files field in the conversation, with a line saying what the files are for. Upload the files as on Add assets (/app/library/assets/add). Each file becomes an asset as soon as its upload finishes, and the field lists it. Upload more adds further files, and Remove deletes an asset after you confirm. Continue gives the Assistant the files in the list, and it carries on, for example by attaching them to a repository. The field needs the right to create assets, which members, managers and admins have.

Cancel tells the Assistant that you added nothing, and the files you already uploaded stay in the Library. During an upload you are asked first, and cancelling stops the upload. Writing a new message instead closes the field and marks it Skipped. The list holds up to 100 files and is stored with the conversation, so after a reload the field shows it again. If storing it fails, the field says so and Save again retries. Files still uploading when the page reloads are not added, and files deleted elsewhere drop out of the list when you choose Continue.

In the work log the field shows as "Uploaded 2 files" with the file titles, or as "Upload cancelled". Metadata next to a title opens the asset's metadata, so you can add it after the Assistant has carried on.

QuestionsCopy Anchor Link

When the Assistant needs your input, for example before it plans, it asks up to five questions in the conversation, as in a bot chat. Answer some, all or none and select Submit, or Skip to go on without answering. Writing a new message instead skips the questions, and the Assistant reads your message as your reply. A quick yes or no may still come as a plain question in its text. It can also run a quiz when you ask for one.

Asking from other pagesCopy Anchor Link

The bot editor (/app/library/bots/<bot id>/setup) and Groups (/app/groups) have an Ask the Assistant button. It opens the Assistant with a prefilled first message, such as "Help me set up the bot “Citation check”", which you can change before sending. The button shows only while the Assistant is available in the active organisation.

One conversation per taskCopy Anchor Link

Start a new conversation for each task. A long conversation is summarised like a long chat, and the Assistant always keeps the current plan and the tool groups it loaded in view. A conversation belongs to you and to the organisation that was active when you started it, and it is listed only while that organisation is active.

From the history, or from the menu under the conversation's title in the header, you can rename and delete a conversation, and the title menu also downloads it. Unlike chats, an Assistant conversation cannot be branched, shared or edited and resent, and it has no model selector, settings, blocks or source pickers. Copy, download, rating and Stop work as in chat, and the usage indicator shows the context and the pool that covers the Assistant model.

Assistant conversations live on their own page and do not appear in the chat list or the chat sidebar, except in the chat search results while their organisation is active.

When the Assistant is unavailableCopy Anchor Link

The Assistant answers with the model an admin chose on the Organisation page. Until a model is chosen, the Assistant page explains that the Assistant is unavailable. Admins get a link to the Organisation page, and everyone else is asked to contact an admin. Existing conversations keep their transcripts and show the same notice instead of the composer.

When an admin switches to another model, existing conversations use it from their next message on.