Setting up collections

Purpose

The settings of the collection determine which language model answers, how it answers, what visitors may do and which paths lead out of the chat. Changes affect all channels of the collection: chat, search, assistants, agents.

Where

Open Edit in the collection rail. The page is divided into nine sections; on the first visit only “Basics” is open. All changes take effect after Save at the end of the page.

Changes to languages and language models affect subsequent indexing runs. Content that is already indexed is retained.

Basics

If the collection is still on the global default without stored access credentials, a notice appears here: the chat then does not answer, and no index is created.

Additional language models and processing region

The table below shows, for each purpose (chat, embeddings, voice, speech recognition), the configuration used and its region. If the main configuration does not meet the region, the attest only appears in embeds with a suitable model. If no model meets the region, dAI Pro resets it to “No restriction” on saving and reports this. A link leads to the machine-readable attest. Chapter 20 goes deeper into the topic.

Image model (copilot illustrations). The configuration with which the editorial Copilot generates illustrations (Chapter 18). Only image-capable configurations are available, today OpenAI and Google Gemini; “None” means the Copilot offers no illustrations. There is no fallback to the chat model: which provider generates images and in which processing region remains your decision. The image model itself is set in the configuration (Chapter 24).

Monthly image limit. Maximum number of generated images per calendar month for this collection, 0 = unlimited. Once the limit is reached, the Copilot reports this and generates nothing until the next month. Every image costs money at the provider; the limit keeps the costs predictable (Chapter 19).

Answer behaviour

Chapter 7 explains how these fields together determine the quality of the answers.

CAG configuration

dAI Pro knows two ways to an answer. RAG searches the index for the matching sections for each question. CAG passes the entire content to the language model, which delivers more precise answers for small collections and stays inexpensive thanks to the provider's cache.

Corporate design

Colours, fonts and style rules of your brand, plus the files dAI Pro should have of your brand. The editorial Copilot hands the style settings to the image model so that generated illustrations match the website, and uses them to design its diagrams and documents (Chapter 18); later they will also feed the chat widget. All fields are optional; without a profile the Copilot works neutral, factual and light.

Brand files. Seven further fields take files. You pick the file on your computer and dAI Pro stores it with the collection; you no longer enter an address anywhere.

Important for day-to-day use: files are saved the moment you select them. The “Save” button at the end of the form only concerns the text fields. A new file replaces the existing one of the same kind; with the reference images, a fourth image pushes out the oldest. “Remove” deletes a file permanently.

What the files are for: the diagram template places the logo at the top right of the image; documents without a letterhead carry it at the top right above the title. PDF header and PDF footer form the letterhead of the documents (section “Documents (PDF and Word)”). If a Word template is stored, the Word files are created in that template: styles, theme, headers and footers and page margins are kept, and dAI Pro only writes the content; the letterhead then applies to the PDF only. dAI Pro rejects templates with macros (DOTM, DOCM). dAI Pro places none of these files into illustrations. The two font files are embedded when diagrams and documents are rendered, so that the result appears in your brand's font even without internet access. The reference images are sent to the image model of OpenAI and Gemini as a style guide, not as content (Chapter 18).

The files are stored with the collection outside the public directory. Only someone who is signed in and has access to this collection can retrieve them. SVG files are sanitised on upload: scripts, event attributes and references to the outside are removed, the drawing itself is kept.

Documents (PDF and Word)

The files “PDF header” and “PDF footer” serve as the letterhead for the documents of the editorial Copilot (Chapter 18). dAI Pro classifies each file itself:

• Full letterhead. If the file takes up at least 60 per cent of the page height, typically a full-page PDF from Pages, Word or InDesign, it lies behind the text as a whole page. If the PDF has two pages, the second applies to the following pages.
• Strip. Narrower files sit at the top as a header and at the bottom as a footer. A PDF keeps its real width, an image spans the full page width.

Besides PDF, SVG, PNG, JPEG and WebP work. dAI Pro places a permission-protected PDF (for example "editing not allowed") under the text as a high-resolution image; the letterhead then no longer stays a vector in the PDF. A PDF that only opens with a password cannot be used; the Copilot reports this in the check report. In both cases it is better to upload an unprotected version.

Free area. dAI Pro rasterises the letterhead in cells of 2 mm and looks for the largest free area for the text, separately for the first and the following pages. The area must be at least half as wide and a good third as high as the page; a narrow column next to the logo does not count. The text keeps 5 mm away from everything printed on the letterhead and at least 15 to 18 mm from the edge of the sheet. The page margins are derived from this area.

Check letterhead sets a two-page sample with the saved settings and shows the margins determined for the first page. So after changes, save the form first; the files themselves are already saved when you select them. Without a letterhead, the check reports that standard margins apply.

Fonts. dAI Pro embeds the uploaded fonts in every PDF. If the body font is a very thin cut with a weight below 300, such as “Thin”, the document sets the body text in the heading font as well. Word does not embed fonts: the font must be installed on the computer, otherwise Word replaces it. dAI Pro reads the family name for this from the font file.

Access and visitor features

What visitors see when attaching files

Visitors have three ways to attach a file:

• Paper clip next to the input field, also with several files at once.
• Paste with Ctrl+V or Cmd+V, for example a screenshot from the clipboard. If the clipboard holds text and an image, the image is attached and the text goes into the input field.
• Drag files onto the chat window. While a file hovers over it, the window shows a dashed frame in the accent colour.

Above the input field, each file appears with a thumbnail, a PDF with a document icon, plus its name and a remove button. Up to five attachments are possible at the same time, and several files are read one after the other. The attachments belong to the next question: once its answer has arrived, they disappear from the input row again. For a further question about the same file, the visitor attaches it once more. The sent question shows them at the top of its bubble.

The chat matches uploaded images and PDFs against the knowledge base. If a file shows something from your documents, such as a wiring diagram from a brochure or an excerpt from a manual, the chat recognises the figure, names the document and page, and explains it with your knowledge instead of merely describing it in general terms. Files unrelated to the collection are answered from their own content.

The chat also reads scanned PDFs in which no text can be selected, such as scanned invoices or delivery notes. It recognises the text on the pages: with Mistral as the language model via Mistral Document AI, otherwise via the image understanding of the language model the collection uses for the chat. This takes a few seconds longer than a PDF with text. The first eight pages are read, without Mistral the first three.

Clicking a thumbnail enlarges it. The enlarged view closes via the button with the X, by clicking next to the image or with Escape.

The thumbnails are created only in the visitor's browser and are not transmitted. After the page is reloaded, the history shows only the file names. If a visitor uploads a great many files in a short time, the chat asks them to try again later.

If upload is switched off for the collection or for an embedding (Chapter 8), pasting and dragging do not work either.

Human handover

Shows the link “Talk to a human” in the chat and offers the handover when the chat could not help. Request and conversation history go by e-mail to the recipients; with “Offer a callback request”, visitors can leave a phone number and preferred time. Chapter 12 describes the process from the visitor's and the team's point of view.

Actions from the chat

Callback, ticket and appointment are confirmed actions: the chat prepares a card with all fields; nothing is sent until the visitor confirms. Here you connect CRM, ticket system, support mailbox and appointment calendar. Chapter 12 explains every connection.

Voice chat

Enables the microphone button, chooses voice and speech recognition and maintains the vocabulary for brand and product names. Chapter 11.

Change log

Every change to a collection's settings is recorded: which field changed when, from which value to which, by whom and through which channel. Open the log via the “Show change log” button at the top of the “Edit” page or via the item of the same name in the collection's navigation rail. Anyone who may edit the collection may see it.

Everything that controls processing is logged: language models and processing region, visitor functions (voice chat, upload, human handover, callback, MCP, feedback, simple language), limits and deadlines (image quota, CAG tokens), allowed websites and addresses, embedding, actions, main prompt and help texts, title, visibility and deletion. Technical timestamps and counters are not logged.

Secrets — API keys, the identity secret, tokens and passwords in the actions — never appear in the log. That such a value changed is noted; the value itself is shown only as “(changed — secret, not shown)”.

The log is your documentation of instructions in the sense of the data processing agreement (Annex A, row A18). It is never deleted automatically and only disappears together with the collection when its record is permanently deleted — for demo collections after they expire. An ordinary deletion in the backend leaves the log in place.

Frequently asked questions

Why can I no longer change the embedding LLM? The index is fixed to this model's vector dimension. For a different embedding model, create a new collection or enter it under “Additional language models”; a second index is then created.

I set the processing region, and after saving it says “No restriction”. None of the collection's language models processes in this region. Choose a configuration with a suitable region (Chapter 24) and set the region again.

Do the allowed websites also apply to the search? Yes, to all endpoints of the collection.

The field “Image model” only offers “None”. No LLM configuration of the instance is image-capable. The administration creates a configuration of type OpenAI or Google Gemini (Chapter 24); after that it appears here for selection.

The sample page still shows the old margins. Check letterhead uses the saved settings. Save the form and check again.

See also

• Chapter 7 · Good answers
• Chapter 12 · Human handover and actions
• Chapter 18 · Editorial copilot (illustrations, diagrams, documents)
• Chapter 20 · Processing region and sovereignty
• Chapter 24 · Configuring language models