Embed the question on the device

xcode-select --install
xcodebuild -version
# File > New > Project > App, Interface: SwiftUI
# Deployment: iOS 27 or macOS 27
# Add FoundationModels and CoreAI to the target.

An embedding is a list of numbers that keeps meaning. Close questions sit close in that space. You need one embedding function and you must keep using it. Mixing two embedders breaks the cache.

Honest API path

Apple does not document one forever name for "embed this string" across every SDK drop. Check these in the Xcode 27 docs, in this order:

1. A Foundation Models embedding or representation API on the session or model, if present.
2. A Core AI utility embedding from the official coreai-models catalog, loaded as an .aimodel.
3. A CoreAILanguageModel only if that type is how your SDK exposes a custom embedder. Many embedders are not language models. Do not force chat APIs onto a vector model.

If none of those exist in your SDK, say so in the project README and stop. Do not call a hosted embed API "for convenience." That leaves the device.

import CoreAI
import FoundationModels

// Illustrative. Confirm the embedder type in the installed SDK.
struct QuestionEmbedder {
    let embed: @Sendable (String) async throws -> [Float]

    func vector(for question: String) async throws -> [Float] {
        let trimmed = question.trimmingCharacters(in: .whitespacesAndNewlines)
        return try await embed(trimmed)
    }
}

Prepare the embedder once at launch, the same way you prepare a Core AI function in the clothing course. First load is slower. Later questions should reuse the prepared model.

WWDC26 Core AI (session 326) shows custom models behind a session. Use that path for a catalog embedder. Next, compare two vectors.