On-device AI Assistant with Apple Intelligence

On-device AI Assistant answers document questions with Apple’s Foundation Models, which power Apple Intelligence. Starting with iOS 27, you can use Apple’s system model or provide your own Foundation Models language model.

Users can ask questions about text in the open document. This feature works best when users ask natural language questions that target specific information. AI Assistant can make mistakes, so users should verify important information.

Apple’s Foundation Models have a limited context window. In Nutrient’s on-device AI Assistant, this doesn’t limit document size. Nutrient’s on-device retrieval-augmented generation (RAG) pipeline selects relevant passages from the full document, so users can ask questions about large documents. It combines semantic and full-text search to find these passages, including in Chinese, Japanese, Korean, and Thai documents.

Requirements

Check these requirements before you enable on-device AI Assistant.

On-device AI Assistant with Apple’s system model requires a device compatible with Apple Intelligence(opens in a new tab) on iOS 26 or later. Users must enable Apple Intelligence in the Settings app on iOS and visionOS, or in System Settings on macOS.

Custom on-device AI Assistant models require iOS 27 or later. The model must conform to Foundation Models’ LanguageModel(opens in a new tab) protocol, and device and memory requirements depend on the model. For example, models hosted through MLX(opens in a new tab) need a physical device or an Apple silicon Mac with enough memory for model weights.

For production, your Nutrient iOS SDK license must include the on-device AI Assistant feature; to add it, contact support(opens in a new tab).

Enable on-device AI Assistant

Choose the integration option that matches your app interface.

struct DocumentViewerWithAIAssistant: View {
    let document: PSPDFKit.Document
    @PDFView.Scope private var scope

    var body: some View {
        PDFView(document: document)
            .toolbar {
                AIAssistantButton()
            }
            .pdfViewScope(scope)
    }
}

pdfViewController.navigationItem.setRightBarButtonItems(
    [pdfViewController.aiAssistantButtonItem],
    for: .document,
    animated: false
)

When you use Apple’s system model, Nutrient automatically hides this button if Apple Intelligence isn’t available or enabled in Settings or System Settings. With a custom model, the button’s visibility doesn’t depend on Apple Intelligence.

Use a custom model starting with iOS 27

Starting with iOS 27, AI Assistant can use any model that conforms to Foundation Models’ LanguageModel(opens in a new tab) protocol. Configure PDFConfiguration.aiAssistantSource with AIAssistantModel.custom(_:identifier:contextWindow:) to use your model instead of Apple’s system model.

A custom model can run on the device or call a server, and AI Assistant sends document text, questions, and conversation history to the model. Review the model’s data handling before you ship.

Use this configuration pattern to select a custom model:

let controller = PDFViewController(document: document) {
    $0.aiAssistantSource = .languageModel(.custom(model, identifier: "qwen3-4b-4bit:8k", contextWindow: 8_192))
}

Present the chat UI yourself

You can present the chat UI from your own interface instead of using the built-in button. Create an AIAssistantSession(document:configuration:) with an AIAssistantModelConfiguration. Then show the session with AIAssistantView or AIAssistantViewController.

The configuration selects the model and identifies stored chat history by session ID and optional user ID. Use a session ID that your app owns and stores, because a document’s uid can change when the document is moved or modified. This works with Apple’s system model on iOS 26 or later, and with a custom model on iOS 27 or later.

let configuration = AIAssistantModelConfiguration(sessionID: conversationID)
let session = try AIAssistantSession(document: document, configuration: configuration)
let aiAssistantViewController = AIAssistantViewController(session: session)

Use AI Assistant without the chat UI

To use document context from your own interface or without an interface, create an AIAssistantModelSession. The session uses the chat UI’s on-device indexing and retrieval, but your app decides what to ask and how to show the result. It uses Apple’s system model by default. On iOS 27 or later, pass a custom model to init(document:model:).

Call prepare() before requesting a response. Preparation indexes the document and can take time, so start it before the user needs an answer. Then call respond(to:) to answer a question:

let assistant = try AIAssistantModelSession(document: document)
try await assistant.prepare()
let answer = try await assistant.respond(to: "Summarize this document.")

The session generates only one response at a time. After the document is saved to its PDF file, including automatic saves of annotation changes, responses throw AIAssistantModelSessionError.documentNotPrepared. Call prepare() again to rebuild the index for the saved content. The session keeps the conversation history.

Limit context to nearby pages

By default, a response searches the whole document for relevant text. When a question is about a known page, pass RetrievalScope.nearbyPages(around:radius:) to use text from that page and the pages on each side of it. The radius defaults to one page. Page indexes in the scope are zero-based, and the requested page must contain text:

// Uses text from pages 4–6 in the viewer.
let answer = try await assistant.respond(
    to: "Explain the main argument on this page.",
    scope: .nearbyPages(around: 4)
)

Generate typed responses

To get data your app can process, define a Foundation Models Generable type and pass it to respond(to:generating:scope:). Each typed call uses a fresh model session without the chat history. Generated values can be incorrect, so validate them before using them to change the document:

import FoundationModels

@Generable
struct BookmarkSuggestion {
    var title: String
}

let suggestion = try await assistant.respond(
    to: "Suggest a bookmark title under five words for this page.",
    generating: BookmarkSuggestion.self,
    scope: .nearbyPages(around: 0, radius: 0)
)

The AI Study Companion Catalog example(opens in a new tab) uses this API to explain selected text, highlight key terms and related passages, and save answers as comments.

Privacy and security

On-device AI Assistant with Apple’s system model processes documents and questions on users’ devices; it doesn’t send that data to a server.

If you use a custom model, privacy depends on the model implementation. A custom model may process data on the device or send data to a server.

Nutrient uses two internal persistent stores on a device when you use on-device AI Assistant:

• A cache stores document text and embeddings locally to improve load time. Nutrient rebuilds the cache for the next question after each save to the PDF file, including automatic annotation saves. Nutrient also rebuilds existing caches once after you update to Nutrient iOS SDK 27.2.
• Chat history stores users’ questions and AI answers. This history typically includes content from documents.

Limitations

• On-device AI Assistant may fail to find information that exists in a document.
• On-device AI Assistant doesn’t support questions about multiple documents.
• On-device AI Assistant doesn’t support navigation to the document part where it found the answer.
• The built-in button doesn’t support multiple separate chat histories for the same document. When you use the API directly, key history by session and user ID.
• On-device AI Assistant doesn’t support password-protected documents.
• Don’t use on-device AI Assistant to summarize large documents unless the document already includes a summary.
• Use on-device AI Assistant to find specific information in large documents.
• Apple’s system model supports the languages Apple Intelligence supports, but it may favor English responses. Language support for custom models depends on their implementation.

Share your feedback(opens in a new tab) to help us prioritize AI Assistant enhancements.