Foundation Models 프레임워크에 LLM 제공자 적용하기

2:00 - Choose a language model

import FoundationModels
import MLXFoundationModels

// On-device Apple Foundation Model
let model = SystemLanguageModel()

// Private Cloud Compute model
// let model = PrivateCloudComputeLanguageModel()

// Custom Core AI model
// let model = try await CoreAILanguageModel(resourcesAt: modelURL)

// Open-source MLX model from HuggingFace
// let model = MLXLanguageModel(modelID: "mlx-community/my-model")

let session = LanguageModelSession(model: model)
let response = try await session.respond(to: "...")
print(response.content)

3:46 - Configure Package.swift for your model package

// Package.swift

let package = Package(
    name: "MyModel",
    platforms: [
        .macOS(.v27), .iOS(.v27), .visionOS(.v27), .watchOS(.v27)
    ],
    products: [
        .library(name: "MyModel", targets: ["MyModel"])
    ],
    dependencies: [
        .package(url: "...", .upToNextMinor(from: "1.0.0"))
    ],
    targets: [
        .target(name: "MyModelRuntime"),
        // public: LanguageModel conformance
        .target(name: "MyModel", dependencies: ["MyModelRuntime"]),
        .testTarget(name: "MyModelTests", dependencies: ["MyModel"])
    ]
)

4:56 - LanguageModel and LanguageModelExecutor protocols

// LanguageModel protocol

public protocol LanguageModel: Sendable {
    var capabilities: LanguageModelCapabilities { get }
    var executorConfiguration: Executor.Configuration { get }
}

// LanguageModelExecutor protocol

public protocol LanguageModelExecutor: Sendable {
    init(configuration: Configuration) throws
    func prewarm(model: Model, transcript: Transcript)
    func respond(
        to request: LanguageModelExecutorGenerationRequest,
        model: Model,
        streamingInto channel: LanguageModelExecutorGenerationChannel
    ) async throws
}

6:25 - Implement LanguageModel and Executor conformances

// LanguageModel conformance
public struct MyLanguageModel: LanguageModel {
    typealias Executor = MyLanguageModelExecutor

    public var capabilities: LanguageModelCapabilities {
        LanguageModelCapabilities(capabilities: [
            .toolCalling, .guidedGeneration, .reasoning
        ])
    }

    public var executorConfiguration: Executor.Configuration {
        Executor.Configuration(/* ... */)
    }
}

// Executor conformance
public struct MyLanguageModelExecutor: LanguageModelExecutor {
    public typealias Model = MyLanguageModel

    public struct Configuration: Hashable, Sendable { /* ... */ }

    public init(configuration: Configuration) throws { /* ... */ }

    public func respond(
        to request: LanguageModelExecutorGenerationRequest,
        model: MyLanguageModel,
        streamingInto channel: LanguageModelExecutorGenerationChannel
    ) async throws { /* ... */ }
}

7:28 - Manage model resources with prewarm and respond

// One approach to managing resources

struct MyLanguageModelExecutor: LanguageModelExecutor {

    private mutating func loadModelIfNeeded() throws -> LoadedWeights {
        let weights = try loadedModel ?? loadWeights()
        loadedModel = weights
        return weights
    }

    func prewarm(transcript: Transcript) {
        loadedModel = try? loadModelIfNeeded()
    }

    func respond( ... ) async throws {
        let weights = try loadModelIfNeeded()
        // ...generate with 'weights'...
    }
}

9:00 - Map Transcript entries to model messages

// Transcript entries

let transcript = Transcript(entries: [
    .instructions( ... ),  // "You are a helpful assistant"

    .prompt( ... ),        // "What's the weather in Pittsburgh?"
    .toolCalls( ... ),     // getWeather(location: "Pittsburgh")
    .toolOutput( ... ),    // 65°F, sunny
    .response( ... ),      // "It's 65°F and sunny in Pittsburgh"

    .prompt( ... ),        // "What's the address of Apple Park?"
    .response( ... ),      // "One Apple Park Way, Cupertino, CA 95014"
])

10:42 - Read generation and context options from the request

// Parse generation and context options

func respond(
    to request: LanguageModelExecutorGenerationRequest,
    model: MyLanguageModel,
    streamingInto channel: LanguageModelExecutorGenerationChannel
) async throws {
    let reasoningLevel = request.contextOptions.reasoningLevel
    let temperature = request.generationOptions.temperature
    let maxTokens = request.generationOptions.maximumResponseTokens
}

11:47 - Stream tokens and metadata through the channel

// Streaming text tokens

func respond( ... ) async throws {
    // 1. Report metadata
    await channel.send(.response(action: .updateMetadata([
        "modelID": "my-model-2026-06-08",
        "requestID": request.id.uuidString
    ])))
    // 2. Report prompt token usage before generating
    await channel.send(.response(action: .updateUsage(
        input: .init(totalTokenCount: promptTokens, cachedTokenCount: cachedTokens),
        output: .init(totalTokenCount: 0, reasoningTokenCount: 0)
    )))
    // 3. Stream text deltas as the model generates
    for try await token in tokens {
        await channel.send(.response(action: .appendText(token)))
    }
}

13:33 - Honor the developer's intent or throw

// Honor the developer's intention where possible

// The developer set sampling: .greedy, but our service only takes temperature
if request.generationOptions.sampling?.kind == .greedy {
    serviceRequest.temperature = 0
}

// Otherwise, throw an error

// The token budget is too small to satisfy the schema
if let schema = request.schema,
   let budget = request.generationOptions.maximumResponseTokens,
   budget < minimumTokens(for: schema) {
    throw LanguageModelError.unsupportedCapability(
        .init(
            capability: .guidedGeneration,
            debugDescription: "Token budget too small to satisfy this schema."
        )
    )
}

13:57 - Built-in errors that any model can throw

// Built-in errors that any model can throw

public enum LanguageModelError: LocalizedError, CustomDebugStringConvertible {
    // Transcript grew past the model's context window. Trim entries and retry.
    case contextSizeExceeded(     )
    // Too many requests in a short window. Space them out or reduce load.
    case rateLimited(     )
    // Model declined to answer. Fall back to a message of your choosing.
    case refusal(     )
    // Safety guardrails tripped on the prompt or the response.
    case guardrailViolation(     )
    // Model lacks a feature you used, such as guided generation or tools.
    case unsupportedCapability(     )
    // Prompt contains content the model can't process (bad files, unknown formats).
    case unsupportedTranscriptContent(     )
    // A generation guide (e.g., a regex pattern) isn't supported by this model.
    case unsupportedGenerationGuide(     )
    // Prompt asked for output in a language or locale the model doesn't support.
    case unsupportedLanguageOrLocale(     )
    // Request timed out before the model produced a response.
    case timeout(     )
}

14:14 - Handle errors from your model executor

// Custom errors

public enum MyModelError: Error, LocalizedError {
    // User hit monthly token limit. Prompt upgrade or wait for reset.
    case exceededSubscriptionTierLimit
    // Model variant isn't enabled on this account.
    case modelNotProvisioned
    // Billing or policy review locked this account.
    case accountSuspended

    public var errorDescription: String? {
        switch self {
        case .exceededSubscriptionTierLimit:
            String(localized: "Your plan limit has been reached.")
        // ...
        }
    }
}

16:08 - Attach custom metadata to responses

// Attach service-specific performance metadata

let elapsed = Date().timeIntervalSince(startTime)
let tokensPerSecond = Double(tokenCount) / elapsed
let timeToFirstToken = firstTokenTime?.timeIntervalSince(startTime) ?? 0

await channel.send(.metadataUpdate([
    "tokensPerSecond": tokensPerSecond,
    "timeToFirstToken": timeToFirstToken
]))

17:05 - Define and use custom Transcript segments

// Define a custom segment
public struct AudioSegment: Transcript.CustomSegment {
    public var id: String
    public var content: URL
}

// Pass it in a prompt
let recording = AudioSegment(id: UUID().uuidString, content: URL(filePath: "/path/to/recording.m4a"))
let response = try await session.respond {
    "Where was Frank Lloyd Wright's original architecture school located?"
    recording
}

// Emit a custom segment from the executor
for try await event in stream {
    switch event {
    case .audioFileGenerated(let file):
        await channel.send(.response(action: .updateCustomSegment(
            AudioSegment(id: file.id, content: file.url)
        )))
    }
}

18:09 - Implement server-side tools in your model

// Configure server-side tools
public struct MyLanguageModel: LanguageModel {
    public struct ServerTool: Sendable {
        public static let webSearch: ServerTool = ...
    }
    public init(serverTools: [ServerTool] = []) { }
}

// Surface tool results through the channel
let client = MyServerClient(serverTools: model.serverTools)
let response = try await client.send(prompt: .init(request))
for try await chunk in response {
    switch chunk {
    case .webSearch(let webSearch):
        await channel.send(.response(action: .updateCustomSegment(
            WebSearchSegment(url: webSearch.url, content: webSearch.html)
        )))
    case .textDelta(let textDelta):
        await channel.send(.response(action: .appendText(
            textDelta.text, tokenCount: textDelta.tokenCount
        )))
    }
}