Custom LLM Provider for Foundation Models Framework

import FoundationModels
import Foundation

// MARK: - Configuration (the cache key)
struct EchoModelConfiguration: LanguageModelExecutorConfiguration {
    let prefix: String

    // Hashable conformance drives executor reuse
    func hash(into hasher: inout Hasher) {
        hasher.combine(prefix)
    }
    static func == (lhs: EchoModelConfiguration, rhs: EchoModelConfiguration) -> Bool {
        lhs.prefix == rhs.prefix
    }
}

// MARK: - Executor (does the real work)
final class EchoModelExecutor: LanguageModelExecutor {
    typealias Configuration = EchoModelConfiguration

    private let configuration: Configuration

    required init(configuration: Configuration) {
        self.configuration = configuration
    }

    func prewarm() async {
        // No expensive setup for this echo model
    }

    func respond(
        to request: LanguageModelRequest,
        channel: LanguageModelResponseChannel
    ) async throws {
        // 1. Send metadata upfront
        await channel.send(.metadataUpdate(
            LanguageModelResponseMetadata(modelID: "echo-v1", requestID: UUID().uuidString)
        ))

        // 2. Extract the last user prompt from the transcript
        let lastPrompt = request.transcript.entries.compactMap { entry -> String? in
            if case .prompt(let content) = entry { return content.text }
            return nil
        }.last ?? ""

        let reply = "\(configuration.prefix): \(lastPrompt)"

        // 3. Stream token deltas
        for word in reply.split(separator: " ") {
            await channel.send(.textDelta(String(word) + " "))
            try await Task.sleep(for: .milliseconds(40)) // simulate streaming
        }
    }
}

// MARK: - LanguageModel (describes capabilities)
struct EchoLanguageModel: LanguageModel {
    typealias Executor = EchoModelExecutor

    var capabilities: LanguageModelCapabilities {
        LanguageModelCapabilities()
    }

    func makeConfiguration() -> EchoModelConfiguration {
        EchoModelConfiguration(prefix: "Echo")
    }
}

// MARK: - Usage in an app
struct EchoViewModel {
    private let session: LanguageModelSession<EchoLanguageModel>

    init() {
        session = LanguageModelSession(model: EchoLanguageModel())
    }

    func ask(_ question: String) async throws -> String {
        let response = try await session.respond(to: question)
        return response.content
    }
}