Knowledge and capabilities

The Intelligence tab is where you give an assistant things it does not know by default: your own documents, the ability to work with the page it is embedded on, and connections to your own systems. Each capability is a card — turn it on with its switch, then expand it to configure. Only one card stays open at a time.

The cards sit under two headings: Knowledge (Knowledge Base, Page Intelligence) and Engagement (Forms, Actions — plus Appointment Scheduling on self-hosted enterprise deployments).

Changes to the switches on this tab are saved with the editor's Save button. Documents, forms and actions are saved as you add them — they do not wait for Save.

Knowledge Base

Paid feature. Add credits to your account to unlock it. On a free account the switch carries a PAID badge and refuses to turn on.

Upload files so the assistant can answer from your own content instead of guessing. Turn on the Knowledge Base switch, then expand the card to manage documents.

Uploading documents

Drop files onto the upload area or click it to browse. Supported types are PDF, DOCX, TXT, MD, HTML, and each file must be 10 MB or smaller. You can select several files at once.

A file is checked against its real contents, not just its extension — a file renamed to .pdf that is not a PDF is rejected. Uploading a file that matches an existing document's name and size returns the document you already have instead of creating a duplicate.

The document table

Every uploaded file appears in a table:

Status moves through these values:

Indexing runs in the background and takes a moment. Reopen the card or reload the page to see the status change.

Only documents with the indexed status are searched. A file sitting at error is ignored entirely.

There is no in-place edit. To update a document, delete the old row and upload the new version.

How retrieval changes answers

When the knowledge base is on, each visitor message triggers a search over the indexed passages of that assistant's documents. The most relevant passages — up to five, and only those scoring above a relevance threshold — are handed to the model alongside your instructions before it writes a reply.

Three consequences worth knowing:

• Nothing relevant, nothing added. If no passage clears the threshold, the assistant answers from its instructions alone. Uploading a document does not force it to be mentioned.
• Small talk skips the search. Greetings, thanks and other pleasantries are not searched, so the assistant will not ground "hi" in your price list.
• Grounded replies are marked. When an answer used your documents, the visitor's chat window shows a Referenced knowledge base marker under that message.

Writing good source documents

Retrieval works on passages, not whole files: each document is split into overlapping chunks and the search matches a chunk at a time. Documents that read well in isolation retrieve well.

• Keep one topic per section and give each section a heading, so a chunk that gets retrieved carries enough context to stand alone.
• Keep a question and its answer close together. An FAQ where answers are three pages from their questions retrieves badly.
• Use text-based files. A scanned PDF that is really a picture of a page has no extractable text and will index as empty.
• Spell out what you mean rather than relying on a table layout or a diagram — layout is flattened to text during processing.
• Remove superseded content instead of leaving it alongside the new version. The search cannot tell which of two contradictory passages is current.

Page Intelligence

Paid feature. Add credits to your account to unlock it. Saving an assistant with Page Intelligence on is rejected on a free account.

Page Intelligence lets the assistant work with the page it is embedded on, rather than only talking about it. Turn on the Page Intelligence switch, then expand the card for three individual permissions. All three are on by default; turn off any you do not want to grant.

Turning off the master switch removes all three capabilities regardless of the individual settings.

If your site declares its own tools in code, those are controlled by the page itself and are covered in Page tools.

Forms

A form is the set of details the assistant collects in the conversation — contact details, an order, a request. The Forms card owns them outright: forms are not edited from the Actions card below, and the trigger picker there does not offer them.

The card header badges how many forms are live (2 live, or Paused when every form is switched off). Expand it for the list, then New form or a row to open the builder.

Building a form

Field types are Short text, Email, Phone, Number, Date, Long text, Dropdown (with its own option list) and Hidden value (a fixed value sent with every submission, which the visitor never sees).

Actions

Actions connect the assistant to your own systems. Each one sends an HMAC-signed HTTP request to an endpoint you control. Expand the Actions card to see the list, or New action to create one.

The two triggers

You choose the trigger when you create an action, and it cannot be changed afterwards. Create a new action if you need a different trigger.

Forms are the third trigger type in the API (form), but they are built in the Forms card above, so they neither appear in this list nor in its trigger picker.

Creating an action

The dialog shows the fields that matter for the trigger you picked:

Finish with Create action, or Update action when editing.

The actions list

Test sends a real signed request to your endpoint immediately, so point it at something you are happy to hit. The result appears as a toast with the HTTP status your endpoint returned, or the error if the request could not be completed. Event actions are tested with a test.ping payload.

Deleting asks for confirmation first — the endpoint stops receiving requests and the action cannot be recovered.

For the request format, the signature header and how to verify it, the full event list, and the delivery rules, see Actions.