AI assistant and search
AI assistant and search
Who can do this: using the assistant needs the Chat right on the project. The Editor and Admin invite roles include it, Read-only does not, and an Owner has every right. The AI document menu items need AI Jobs. Search works for anyone who can open the project and shows only what that person may read. Rights are explained once in the .
What it's for
The assistant is a chat inside your project. You describe what you want in plain words. It looks things up, answers questions about your data, writes documents and, when you allow it, builds and changes things: tables, forms, dashboards, automations. Search is the quicker tool when you already know what you are looking for. This page covers both, what the assistant may change, how its approvals work, where its use is counted, and the assistant your customers can get in the portal.
Before you start
You need Chat on the project. Without it the chat button is still there, but a question is refused with a red error line under the message box.
The assistant, AI documents and meaning-based search use AI that your administrator sets up. Their use counts toward your plan's AI credits (see "What the assistant costs").
Needs your administrator: the assistant needs an AI provider for the Agent role, and AI documents need one for the Doc Generation role. Ask for "an AI provider with the Agent and Doc Generation roles". You can tell it is missing when the assistant never answers or a red line starting "Error:" appears under the message box (see ). Search by meaning needs the embedding endpoint (the Embedding Model panel in the AI tab of System). Without it, search quietly falls back to keywords only and shows no message. See .
Ask the assistant
When a document is open, a bar shows Current document and Project. Pick Current document for questions about "this page" or "this table". Pick Project for the whole project.
Type in the box and press Enter. Shift+Enter starts a new line. You can also press Send.
Watch the status line. For bigger jobs the assistant keeps a Tasks list under the conversation. Press Stop generating to cut it short.
Starter ideas from the empty panel: "How many requests were there this month?", "Find a contact by name.", and, for owners, "Create an inventory tracking table for the warehouse."
Buttons along the top of the panel:
Button | What it does |
|---|---|
Ask / Auto | Switches between asking before each change and changing without asking (next section) |
New conversation | Starts a clean conversation |
Chat history | Lists your earlier conversations in this project. Only your own appear. Choose one to reopen it, then Back to chat |
Maximize / Minimize | Makes the panel larger or smaller. You can also drag its top and left edges |
Close | Hides the panel. The round button brings it back |
What the assistant can do
Answer questions: find a table or document by name or by a phrase inside it, read rows with filters, count, total and rank them ("top 5 employees by closed tasks").
Write documents: draft a new page or add text to one.
Build and change the project, for owners only: tables and their columns, forms, dashboards, spreadsheets, automations, access rules, people and groups, working calendars, project settings and message channels. It can also install business modules.
Work with data: add, change and delete rows, and move a row along its status steps.
It acts only through the product's own functions, with your rights. A document, table or row you cannot open is invisible to it as well.
Who is chatting | What the assistant may change |
|---|---|
An | Structure, settings, access and people, as well as data and documents |
Anyone else with | Content they can edit, such as rows and documents. It will not change columns, automations, access, settings, people or channels, and says so |
Anyone with | Nothing. It can read and answer |
When the assistant needs a password or key (for example to connect a channel), it shows an Enter credentials form instead. The values go straight into encrypted settings and the assistant never sees them.
Approve what it changes
New conversations start in Ask mode, the safe one. Before the assistant changes anything, a card appears with the title "The agent wants to make a change", a one-line summary such as "Update a row: Orders", and the exact request below it in grey.
Read the summary.
Press Approve to let this one change happen, or Deny to refuse it. After a denial the assistant is told not to retry and asks how you want to proceed.
Press Always allow this to approve this kind of change without asking until you reload the page.
Reading, searching and previewing an automation never ask. Edits to an open whiteboard are made on your screen and do not wait for approval.
To stop being asked, press Ask at the top of the panel. It changes to Auto, and the assistant changes things without asking. Press it again to go back. The choice is remembered by this browser only, not by your account or the project, and a new browser starts in Ask. The full-screen setup chat on first run uses Auto, because it first proposes a plan and waits for you to confirm it.
Careful: in Auto mode and after Always allow this, the assistant can delete rows and change structure or access with no prompt. Always allow this covers every use of that kind of change, not only the one on the card (for example every row update, in any table). Use Auto on a project you can restore from a backup or whose changes you can review (see ).
When the assistant asks you something
Sometimes it stops and asks you a question, with options to pick and a box for your own answer. Choose or type, then press Send answer. Or just type the answer in the message box.
Give it text, take its answers
Select text in a document and press Add to chat in the selection bar. A small tag, "excerpt from" and the document name, joins your next message. You can also drag a selection onto the panel.
Hover over an answer and press the Add to document icon (in English in every language) to place the answer at the end of the open document (Add to bottom) or where you last had the cursor (Add to last cursor position). The icon appears only when a document is open and you can edit it.
Language and standing rules
To give it rules that always apply (your terms, how you name things), open the AI document menu in the navigator toolbar (AI document generation) and choose Set rules and definitions. The dialog Rules & Definitions (AGENT.MD) takes free text, which is added to the assistant's instructions and to AI document generation. The item is greyed out without AI Jobs.
AI documents and recipes
The AI document generation menu in the navigator toolbar has these entries. They are greyed out without AI Jobs.
Entry | What it does |
|---|---|
New AI document | Dialog AI Document: pick Single document or Compound document, type a name, pick a Recipe or leave Custom, describe what the document should contain, press Create |
Run AI recipe | Lists saved recipes. Pick one, name the document, press Generate (dialog Generate new AI doc) |
Manage AI recipes | Dialog AI Recipes: Add Recipe, fill Name and Prompt, Save. Delete removes your own recipes |
A recipe is a saved prompt you reuse. Entries tagged System are built in: you can edit their prompt but not delete them or rename them. The first three entries under Run AI recipe document source code and are of no use in a business project.
Creating the document stores the prompt and an empty page. The text is written when you start it: open the document's menu, choose Edit prompt, then Save and generate. A generation can take a while and shows a progress notice on the page.
What the assistant costs
Each answer uses AI. Its tokens are turned into AI credits at the rate your administrator sets for the provider, and the total appears in Billing & usage as AI credits, against your plan's monthly limit (see ). The monthly limit is checked only on calls through the platform's AI proxy endpoint. The chat counts its use but is not stopped by that check. Where the limit is checked, you see Plan limit reached with Upgrade plan.
If you set your own key under Agent AI provider in Agents (needs Manage Members), the assistant uses that provider too. See .
The assistant for your customers
Customers who sign in to your portal can get their own assistant. It is separate from the one above, off by default, read-only, and sees only what that customer may open. Turn it on with Portal AI under Project features in Edit, or with AI assistant in External Portals, Overview. Its answers count as AI credits too. Details, limits and the customer view are in and .
Search
You want to find | Use |
|---|---|
A project or a document by name | Quick search on the Projects page |
Text inside documents, in one project | The Search button in the project |
An item in the navigator tree | The filter box above the tree |
A row in a table | The Quick search box in the table toolbar |
Anything, by describing it | Ask the assistant |
Find a project or document by name
On the Projects page, press Quick search at the top, or press Ctrl K (⌘ K on a Mac). The button is hidden on narrow screens; the box Search projects and documents is on the page itself.
Type part of a name. Results appear as you type, in two groups: Projects and Documents, each document with its project's name.
Click a result, or press Enter to open the first one (a project first, otherwise a document).
It matches names only, not contents. It shows up to 6 projects and 15 documents, and Nothing found when nothing matches.
Find text inside documents
Inside a project, press the Search button (magnifier) in the header, or press Ctrl K (⌘ K) when your cursor is not in a text field.
Type at least two characters into "Search documents…".
Each result shows the document name, the section heading after "§", a one-line preview and a dot: green when the document is published, grey otherwise. Use the arrow keys and Enter, or click, to open it. Esc closes the list.
It reads the text of documents, section by section, and shows at most 10 documents. Archived documents are left out. With an embedding set up it ranks by meaning as well as by words. Without one it matches words only.
The icon at the right end of the box switches what is searched. If the placeholder reads "Search symbols…", press it (Find documents) to come back.
Find an item in the navigator, or a row in a table
The box above the tree ("Filter by name or #tag…") narrows the tree to matching names or tags. Clear filter brings it all back.
A table with a text column shows Quick search in its toolbar. It narrows the rows on top of your current filters and sort, and Clear search resets it. A column takes part if it is Text, Long text, Email, Phone or URL, or if you chose Include in search in its menu. Computed columns never do (see ). Tables whose rows live in an outside connector have no such box.
Visitors to your published pages get their own Search documents over published documents only (see ).
What search respects
Names and contents are filtered by each document's access, so a restricted document never shows up for someone who cannot read it.
Table search runs on the same request as the grid, so rows hidden from you stay hidden.
Project and document search show nothing from projects you cannot open.
How to check it worked
Ask "How many documents are in this project?". A real answer means the provider works.
Ask for a small change, such as adding one row to a test table. In Ask mode the approval card appears. After Approve, the row shows in the table.
Type a word you know is in a document into the project search. The document should appear with its section and preview.
Undo / what can go wrong
A red "Error:" line after a question: no AI provider for the Agent role, or you lack Chat. Check with an owner or your administrator.
The error mentions "The AI is at capacity right now. Please retry in a few seconds.": the AI is busy. Send again.
It declines a structural change: only an Owner gets those, or you lack Edit on that table.
Search finds the word but not the idea: the embedding is not set up. Ask your administrator.
A new AI document is empty: the prompt is stored but nothing has started. Use Edit prompt, then Save and generate.