AI in Knowledge, Documents and cloud files¶
Three small modules connect the AI to the places where your company keeps its articles and files. They build on the knowledge sources described in Knowledge sources, so the same indexing, citation and access rules apply, and they install themselves when the required apps are present. AI Knowledge needs the Knowledge app, AI Documents needs the Documents app and AI Cloud Storage needs cloud storage.
Important
The pgvector extension and an embedding model are required for all knowledge sources (see Knowledge sources).
Knowledge articles as sources¶
In , create a source and choose the type Knowledge articles. Pick one or more Knowledge spaces (every article of the space is used) and/or single Articles. Include sub-articles is on by default. At least one space or article is needed. Templates and articles in the trash are not indexed, and a source covers at most 2000 articles by default.
Each article is indexed together with its position in the space (space, parent articles, title), so a search finds it by context. A citation chip in an answer opens the article.
When an article changes, only that article is indexed again, at the next indexing run (every 10 minutes) or when you click Index now. Moving an article takes its sub-articles along; deleting an article removes its excerpts at once.
The Knowledge app’s own sharing is respected for every question: a private, archived or trashed article is never cited to someone who may not see it, even before it is indexed again. Files embedded in an article are not indexed with it; add them as a file source if you need them.
Draft an article with AI¶
In the article editor, the top bar shows Draft with AI when a main AI model is configured, you are an AI user and the article is not locked or in the trash. It saves the article and opens a dialog with three steps:
Brief: describe What should the article cover? (the title is the default). If knowledge sources are available to you, you can switch on Use the knowledge sources so that the AI draws on your company’s own material. Click Propose an outline.
Outline: the AI proposes 3 to 8 sections as a list. Edit it freely, or click New outline to ask again. Then click Write the article.
Preview: read the finished article. Click Insert into the article, or Write again or go back. When the article already has text, choose When accepting between Replace the current content and Add after the current content.
The AI sees the title, the space, the current content and the names of the parent and sibling articles that you may see, and the excerpts from the sources when you chose them. It is told not to invent company facts and to put placeholders in square brackets, so review and complete the draft. The text goes into the article with your rights, is recorded in the AI change log and counts as a normal edit in the article history.
AI cover¶
When an image model is configured (see Providers and models), a picture button in the top bar (tooltip Generate a cover image with AI) opens a dialog with an editable description of the cover (by default a calm, wide banner for the title, without text). Click Generate, look at the preview, and then Use as cover, try again or Change the description. The cover appears above the article.
Document folders as sources¶
In , choose the type Document folders, pick the folders and decide with Include sub-folders whether the sub-folders count. Files attached to fields and links are not used. A source covers at most 500 documents by default, newest first, and the text of each file is read up to 25 MB. When a document is attached to a business record, its excerpts also follow the record’s access rules.
New and changed documents are indexed again incrementally. At question time a document is used only if you may read its folder and the document itself, with the same visibility as in the Documents app. Users without the Documents app get nothing from these sources.
AI filing of documents¶
The AI can propose the right folder and tags for a document, either when you ask or when a file is uploaded.
On request¶
On a document form, click Suggest a folder (you need write access to the document). The proposal is computed at once and shows the folder, the tags, the confidence and the reason. Click Apply or Dismiss. A pending proposal also appears as a banner on the document.
On upload¶
Open a folder and look at the group AI filing. Choose Suggest a folder so that every document uploaded into the folder gets a proposal that the uploader confirms, or File automatically so that proposals are applied at once when the AI is at least as sure as the Minimum confidence (0.8 by default); the other proposals wait for confirmation. The option AI may file documents here decides whether other documents can be filed into this folder. Proposals are prepared about every 10 minutes, in the name of the person who uploaded the file, with that person’s rights.
Open to see the proposals (the filter To decide is on by default), with the filters Applied, Failed and Applied automatically. A failed proposal can be retried with Try again. When a proposal is applied, the previous folder and tags are kept on it, and a note in the document chatter says who confirmed it, or that it was filed automatically.
Note
The AI reads the first part of a document’s text only when the AI read policy allows documents, or when the folder has AI filing turned on, and never when an explicit AI deny rule applies to you. Filing runs with the uploader’s AI rights, so uploads made by technical users, such as the mail gateway, usually fail with a recorded error.
Files in cloud storage¶
With AI Cloud Storage installed, files that are stored with the cloud storage provider can be indexed like local files, wherever the knowledge sources read attachments: file sources, record sources, document folders and AI filing. The server downloads the file through the signed download link of the configured provider.
Only secure (https) links are followed and redirects are not. A file larger than 25 MB by default (100 MB at most) is skipped, and each file has a time budget, so a failed or too large download skips that file only.
Cloud files are downloaded completely every time the source is indexed again; there is no local cache. If cloud storage is not configured, the source fails with a clear message.