Foundation Models — On-Device AI for Apple Platforms

When to Use This Skill

Use when:

• Implementing on-device AI features with Foundation Models
• Adding text summarization, classification, or extraction capabilities
• Creating structured output from LLM responses
• Building tool-calling patterns for external data integration
• Streaming generated content for better UX
• Debugging Foundation Models issues (context overflow, slow generation, wrong output)
• Deciding between Foundation Models vs server LLMs (ChatGPT, Claude, etc.)

Related Skills

• Use axiom-foundation-models-diag for systematic troubleshooting (context exceeded, guardrail violations, availability problems)
• Use axiom-foundation-models-ref for complete API reference with all WWDC code examples

---

Red Flags — Anti-Patterns That Will Fail

❌ Using for World Knowledge

Why it fails: The on-device model is 3 billion parameters, optimized for summarization, extraction, classification — NOT world knowledge or complex reasoning.

Example of wrong use:

// ❌ BAD - Asking for world knowledge
let session = LanguageModelSession()
let response = try await session.respond(to: "What's the capital of France?")

Why: Model will hallucinate or give low-quality answers. It's trained for content generation, not encyclopedic knowledge.

Correct approach: Use server LLMs (ChatGPT, Claude) for world knowledge, or provide factual data through Tool calling.

---

❌ Blocking Main Thread

Why it fails: session.respond() is async but if called synchronously on main thread, freezes UI for seconds.

Example of wrong use:

// ❌ BAD - Blocking main thread
Button("Generate") {
    let response = try await session.respond(to: prompt) // UI frozen!
}

Why: Generation takes 1-5 seconds. User sees frozen app, bad reviews follow.

Correct approach:

// ✅ GOOD - Async on background
Button("Generate") {
    Task {
        let response = try await session.respond(to: prompt)
        // Update UI with response
    }
}

---

❌ Manual JSON Parsing

Why it fails: Prompting for JSON and parsing with JSONDecoder leads to hallucinated keys, invalid JSON, no type safety.

Example of wrong use:

// ❌ BAD - Manual JSON parsing
let prompt = "Generate a person with name and age as JSON"
let response = try await session.respond(to: prompt)
let data = response.content.data(using: .utf8)!
let person = try JSONDecoder().decode(Person.self, from: data) // CRASHES!

Why: Model might output {firstName: "John"} when you expect {name: "John"}. Or invalid JSON entirely.

Correct approach:

// ✅ GOOD - @Generable guarantees structure
@Generable
struct Person {
    let name: String
    let age: Int
}

let response = try await session.respond(
    to: "Generate a person",
    generating: Person.self
)
// response.content is type-safe Person instance

---

❌ Ignoring Availability Check

Why it fails: Foundation Models only runs on Apple Intelligence devices in supported regions. App crashes or shows errors without check.

Example of wrong use:

// ❌ BAD - No availability check
let session = LanguageModelSession() // Might fail!

Correct approach:

// ✅ GOOD - Check first
switch SystemLanguageModel.default.availability {
case .available:
    let session = LanguageModelSession()
    // proceed
case .unavailable(let reason):
    // Show graceful UI: "AI features require Apple Intelligence"
}

---

❌ Single Huge Prompt

Why it fails: 4096 token context window (input + output). One massive prompt hits limit, gives poor results.

Example of wrong use:

// ❌ BAD - Everything in one prompt
let prompt = """
    Generate a 7-day itinerary for Tokyo including hotels, restaurants,
    activities for each day, transportation details, budget breakdown...
    """
// Exceeds context, poor quality

Correct approach: Break into smaller tasks, use tools for external data, multi-turn conversation.

---

❌ Not Handling Context Overflow

Why it fails: Multi-turn conversations grow transcript. Eventually exceeds 4096 tokens, throws error, conversation ends.

Must handle:

// ✅ GOOD - Handle overflow
do {
    let response = try await session.respond(to: prompt)
} catch LanguageModelSession.GenerationError.exceededContextWindowSize {
    // Condense transcript and create new session
    session = condensedSession(from: session)
}

---

❌ Not Handling Guardrail Violations

Why it fails: Model has content policy. Certain prompts trigger guardrails, throw error.

Must handle:

// ✅ GOOD - Handle guardrails
do {
    let response = try await session.respond(to: userInput)
} catch LanguageModelSession.GenerationError.guardrailViolation {
    // Show message: "I can't help with that request"
}

---

❌ Not Handling Unsupported Language

Why it fails: Model supports specific languages. User input might be unsupported, throws error.

Must check:

// ✅ GOOD - Check supported languages
let supported = SystemLanguageModel.default.supportedLanguages
guard supported.contains(Locale.current.language) else {
    // Show disclaimer
    return
}

---

Mandatory First Steps

Before writing any Foundation Models code, complete these steps:

1. Check Availability

switch SystemLanguageModel.default.availability {
case .available:
    // Proceed with implementation
    print("✅ Foundation Models available")
case .unavailable(let reason):
    // Handle gracefully - show UI message
    print("❌ Unavailable: \(reason)")
}

Why: Foundation Models requires:

• Apple Intelligence-enabled device
• Supported region
• User opted in to Apple Intelligence

Failure mode: App crashes or shows confusing errors without check.

---

2. Identify Use Case

Ask yourself: What is my primary goal?

Use Case	Foundation Models?	Alternative
Summarization	✅ YES
Extraction (key info from text)	✅ YES
Classification (categorize content)	✅ YES
Content tagging	✅ YES (built-in adapter!)
World knowledge	❌ NO	ChatGPT, Claude, Gemini
Complex reasoning	❌ NO	Server LLMs
Mathematical computation	❌ NO	Calculator, symbolic math

Critical: If your use case requires world knowledge or advanced reasoning, stop. Foundation Models is the wrong tool.

---

3. Design @Generable Schema

If you need structured output (not just plain text):

Bad approach: Prompt for "JSON" and parse manually Good approach: Define @Generable type

@Generable
struct SearchSuggestions {
    @Guide(description: "Suggested search terms", .count(4))
    var searchTerms: [String]
}

Why: Constrained decoding guarantees structure. No parsing errors, no hallucinated keys.

---

4. Consider Tools for External Data

If your feature needs external information:

• Weather → WeatherKit tool
• Locations → MapKit tool
• Contacts → Contacts API tool
• Calendar → EventKit tool

Don't try to get this information from the model (it will hallucinate). Do define Tool protocol implementations.

---

5. Plan Streaming for Long Generations

If generation takes >1 second, use streaming:

let stream = session.streamResponse(
    to: prompt,
    generating: Itinerary.self
)

for try await partial in stream {
    // Update UI incrementally
    self.itinerary = partial
}

Why: Users see progress immediately, perceived latency drops dramatically.

---

Decision Tree

Need on-device AI?
│
├─ World knowledge/reasoning?
│  └─ ❌ NOT Foundation Models
│     → Use ChatGPT, Claude, Gemini, etc.
│     → Reason: 3B parameter model, not trained for encyclopedic knowledge
│
├─ Summarization?
│  └─ ✅ YES → Pattern 1 (Basic Session)
│     → Example: Summarize article, condense email
│     → Time: 10-15 minutes
│
├─ Structured extraction?
│  └─ ✅ YES → Pattern 2 (@Generable)
│     → Example: Extract name, date, amount from invoice
│     → Time: 15-20 minutes
│
├─ Content tagging?
│  └─ ✅ YES → Pattern 3 (contentTagging use case)
│     → Example: Tag article topics, extract entities
│     → Time: 10 minutes
│
├─ Need external data?
│  └─ ✅ YES → Pattern 4 (Tool calling)
│     → Example: Fetch weather, query contacts, get locations
│     → Time: 20-30 minutes
│
├─ Long generation?
│  └─ ✅ YES → Pattern 5 (Streaming)
│     → Example: Generate itinerary, create story
│     → Time: 15-20 minutes
│
└─ Dynamic schemas (runtime-defined structure)?
   └─ ✅ YES → Pattern 6 (DynamicGenerationSchema)
      → Example: Level creator, user-defined forms
      → Time: 30-40 minutes

---

Pattern 1: Basic Session (~1500 words)

Use when: Simple text generation, summarization, or content analysis.

Core Concepts

LanguageModelSession:

• Stateful — retains transcript of all interactions
• Instructions vs prompts:
  • Instructions (from developer): Define model's role, static guidance
  • Prompts (from user): Dynamic input for generation
• Model trained to obey instructions over prompts (security feature)

Implementation

import FoundationModels

func respond(userInput: String) async throws -> String {
    let session = LanguageModelSession(instructions: """
        You are a friendly barista in a pixel art coffee shop.
        Respond to the player's question concisely.
        """
    )
    let response = try await session.respond(to: userInput)
    return response.content
}

// WWDC 301:1:05

Key Points

1. Instructions are optional — Reasonable defaults if omitted
2. Never interpolate user input into instructions — Security risk (prompt injection)
3. Keep instructions concise — Each token adds latency

Multi-Turn Interactions

let session = LanguageModelSession()

// First turn
let first = try await session.respond(to: "Write a haiku about fishing")
print(first.content)
// "Silent waters gleam,
//  Casting lines in morning mist—
//  Hope in every cast."

// Second turn - model remembers context
let second = try await session.respond(to: "Do another one about golf")
print(second.content)
// "Silent morning dew,
//  Caddies guide with gentle words—
//  Paths of patience tread."

// Inspect full transcript
print(session.transcript)

// WWDC 286:17:46

Why this works: Session retains transcript automatically. Model uses context from previous turns.

Transcript Inspection

let transcript = session.transcript
// Use for:
// - Debugging generation issues
// - Showing conversation history in UI
// - Exporting chat logs

Error Handling (Basic)

do {
    let response = try await session.respond(to: prompt)
} catch LanguageModelSession.GenerationError.guardrailViolation {
    // Content policy triggered
    print("Cannot generate that content")
} catch LanguageModelSession.GenerationError.unsupportedLanguageOrLocale {
    // Language not supported
    print("Please use English or another supported language")
}

When to Use This Pattern

✅ Good for:

• Simple Q&A
• Text summarization
• Content analysis
• Single-turn generation

❌ Not good for:

• Structured output (use Pattern 2)
• Long conversations (will hit context limit)
• External data needs (use Pattern 4)

Time Cost

Implementation: 10-15 minutes for basic usage Debugging: +5-10 minutes if hitting errors

---

Pattern 2: @Generable Structured Output (~2000 words)

Use when: You need structured data from model, not just plain text.

The Problem

Without @Generable:

// ❌ BAD - Unreliable
let prompt = "Generate a person with name and age as JSON"
let response = try await session.respond(to: prompt)
// Might get: {"firstName": "John"} when you expect {"name": "John"}
// Might get invalid JSON entirely
// Must parse manually, prone to crashes

The Solution: @Generable

@Generable
struct Person {
    let name: String
    let age: Int
}

let session = LanguageModelSession()
let response = try await session.respond(
    to: "Generate a person",
    generating: Person.self
)

let person = response.content // Type-safe Person instance!

// WWDC 301:8:14

How It Works (Constrained Decoding)

1. @Generable macro generates schema at compile-time
2. Schema passed to model automatically
3. Model generates tokens constrained by schema
4. Framework parses output into Swift type
5. Guaranteed structural correctness — No hallucinated keys, no parsing errors

"Constrained decoding masks out invalid tokens. Model can only pick tokens valid according to schema."

Supported Types

Primitives:

• String, Int, Float, Double, Bool

Arrays:

@Generable
struct SearchSuggestions {
    var searchTerms: [String]
}

Nested/Composed:

@Generable
struct Itinerary {
    var destination: String
    var days: [DayPlan] // Composed type
}

@Generable
struct DayPlan {
    var activities: [String]
}

// WWDC 286:6:18

Enums with Associated Values:

@Generable
struct NPC {
    let name: String
    let encounter: Encounter

    @Generable
    enum Encounter {
        case orderCoffee(String)
        case wantToTalkToManager(complaint: String)
    }
}

// WWDC 301:10:49

Recursive Types:

@Generable
struct Itinerary {
    var destination: String
    var relatedItineraries: [Itinerary] // Recursive!
}

@Guide Constraints

Control generated values with @Guide:

Natural Language Description:

@Generable
struct NPC {
    @Guide(description: "A full name with first and last")
    let name: String
}

Numeric Ranges:

@Generable
struct Character {
    @Guide(.range(1...10))
    let level: Int
}

// WWDC 301:11:20

Array Count:

@Generable
struct Suggestions {
    @Guide(description: "Suggested search terms", .count(4))
    var searchTerms: [String]
}

// WWDC 286:5:32

Maximum Count:

@Generable
struct Result {
    @Guide(.maximumCount(3))
    let topics: [String]
}

Regex Patterns:

@Generable
struct NPC {
    @Guide(Regex {
        Capture {
            ChoiceOf {
                "Mr"
                "Mrs"
            }
        }
        ". "
        OneOrMore(.word)
    })
    let name: String
}

// Output: {name: "Mrs. Brewster"}

// WWDC 301:13:40

Property Order Matters

Properties generated in declaration order:

@Generable
struct Itinerary {
    var destination: String // Generated first
    var days: [DayPlan]     // Generated second
    var summary: String     // Generated last
}

"You may find model produces best summaries when they're last property."

Why: Later properties can reference earlier ones. Put most important properties first for streaming.

---

Pattern 3: Streaming with PartiallyGenerated (~1500 words)

Use when: Generation takes >1 second and you want progressive UI updates.

The Problem

Without streaming:

// User waits 3-5 seconds seeing nothing
let response = try await session.respond(to: prompt, generating: Itinerary.self)
// Then entire result appears at once

User experience: Feels slow, frozen UI.

The Solution: Streaming

@Generable
struct Itinerary {
    var name: String
    var days: [DayPlan]
}

let stream = session.streamResponse(
    to: "Generate a 3-day itinerary to Mt. Fuji",
    generating: Itinerary.self
)

for try await partial in stream {
    print(partial) // Incrementally updated
}

// WWDC 286:9:40

PartiallyGenerated Type

@Generable macro automatically creates PartiallyGenerated type:

// Compiler generates:
extension Itinerary {
    struct PartiallyGenerated {
        var name: String?        // All properties optional!
        var days: [DayPlan]?
    }
}

Why optional: Properties fill in as model generates them.

SwiftUI Integration

struct ItineraryView: View {
    let session: LanguageModelSession
    @State private var itinerary: Itinerary.PartiallyGenerated?

    var body: some View {
        VStack {
            if let name = itinerary?.name {
                Text(name)
                    .font(.title)
            }

            if let days = itinerary?.days {
                ForEach(days, id: \.self) { day in
                    DayView(day: day)
                }
            }

            Button("Generate") {
                Task {
                    let stream = session.streamResponse(
                        to: "Generate 3-day itinerary to Tokyo",
                        generating: Itinerary.self
                    )

                    for try await partial in stream {
                        self.itinerary = partial
                    }
                }
            }
        }
    }
}

// WWDC 286:10:05

Animations & Transitions

Add polish:

if let name = itinerary?.name {
    Text(name)
        .transition(.opacity)
}

if let days = itinerary?.days {
    ForEach(days, id: \.self) { day in
        DayView(day: day)
            .transition(.slide)
    }
}

"Get creative with SwiftUI animations to hide latency. Turn waiting into delight."

View Identity

Critical for arrays:

// ✅ GOOD - Stable identity
ForEach(days, id: \.id) { day in
    DayView(day: day)
}

// ❌ BAD - Identity changes, animations break
ForEach(days.indices, id: \.self) { index in
    DayView(day: days[index])
}

Property Order for Streaming UX

// ✅ GOOD - Title appears first, summary last
@Generable
struct Itinerary {
    var name: String        // Shows first
    var days: [DayPlan]     // Shows second
    var summary: String     // Shows last (can reference days)
}

// ❌ BAD - Summary before content
@Generable
struct Itinerary {
    var summary: String     // Doesn't make sense before days!
    var days: [DayPlan]
}

// WWDC 286:11:00

When to Use Streaming

✅ Use for:

• Itineraries
• Stories
• Long descriptions
• Multi-section content

❌ Skip for:

• Simple Q&A (< 1 sentence)
• Quick classification
• Content tagging

Time Cost

Implementation: 15-20 minutes with SwiftUI Polish (animations): +5-10 minutes

---

Pattern 4: Tool Calling (~2000 words)

Use when: Model needs external data (weather, locations, contacts) to generate response.

The Problem

// ❌ BAD - Model will hallucinate
let response = try await session.respond(
    to: "What's the temperature in Cupertino?"
)
// Output: "It's about 72°F" (completely made up!)

Why: 3B parameter model doesn't have real-time weather data.

The Solution: Tool Calling

Let model autonomously call your code to fetch external data.

import FoundationModels
import WeatherKit
import CoreLocation

struct GetWeatherTool: Tool {
    let name = "getWeather"
    let description = "Retrieve latest weather for a city"

    @Generable
    struct Arguments {
        @Guide(description: "The city to fetch weather for")
        var city: String
    }

    func call(arguments: Arguments) async throws -> ToolOutput {
        let places = try await CLGeocoder().geocodeAddressString(arguments.city)
        let weather = try await WeatherService.shared.weather(for: places.first!.location!)
        let temp = weather.currentWeather.temperature.value

        return ToolOutput("\(arguments.city)'s temperature is \(temp) degrees.")
    }
}

// WWDC 286:13:42

Attaching Tool to Session

let session = LanguageModelSession(
    tools: [GetWeatherTool()],
    instructions: "Help user with weather forecasts."
)

let response = try await session.respond(
    to: "What's the temperature in Cupertino?"
)

print(response.content)
// "It's 71°F in Cupertino!"

// WWDC 286:15:03

Model autonomously:

1. Recognizes it needs weather data
2. Calls GetWeatherTool
3. Receives real temperature
4. Incorporates into natural response

Tool Protocol Requirements

protocol Tool {
    var name: String { get }
    var description: String { get }

    associatedtype Arguments: Generable

    func call(arguments: Arguments) async throws -> ToolOutput
}

Name: Short, verb-based (e.g. getWeather, findContact) Description: One sentence explaining purpose Arguments: Must be @Generable (guarantees valid input) call: Your code — fetch data, process, return

ToolOutput

Two forms:

1. Natural language (String):

return ToolOutput("Temperature is 71°F")

2. Structured (GeneratedContent):

let content = GeneratedContent(properties: ["temperature": 71])
return ToolOutput(content)

Multiple Tools Example

let session = LanguageModelSession(
    tools: [
        GetWeatherTool(),
        FindRestaurantTool(),
        FindHotelTool()
    ],
    instructions: "Plan travel itineraries."
)

let response = try await session.respond(
    to: "Create a 2-day plan for Tokyo"
)

// Model autonomously decides:
// - Calls FindRestaurantTool for dining
// - Calls FindHotelTool for accommodation
// - Calls GetWeatherTool to suggest activities

Stateful Tools

Tools can maintain state across calls:

class FindContactTool: Tool {
    let name = "findContact"
    let description = "Find contact from age generation"

    var pickedContacts = Set<String>() // State!

    @Generable
    struct Arguments {
        let generation: Generation

        @Generable
        enum Generation {
            case babyBoomers
            case genX
            case millennial
            case genZ
        }
    }

    func call(arguments: Arguments) async throws -> ToolOutput {
        // Use Contacts API
        var contacts = fetchContacts(for: arguments.generation)

        // Remove already picked
        contacts.removeAll(where: { pickedContacts.contains($0.name) })

        guard let picked = contacts.randomElement() else {
            return ToolOutput("No more contacts")
        }

        pickedContacts.insert(picked.name) // Update state
        return ToolOutput(picked.name)
    }
}

// WWDC 301:21:55

Why class, not struct: Need to mutate state from call method.

Tool Calling Flow

1. Session initialized with tools
2. User prompt: "What's Tokyo's weather?"
3. Model analyzes: "Need weather data"
4. Model generates tool call: getWeather(city: "Tokyo")
5. Framework calls your tool's call() method
6. Your tool fetches real data from API
7. Tool output inserted into transcript
8. Model generates final response using tool output

"Model decides autonomously when and how often to call tools. Can call multiple tools per request, even in parallel."

Tool Calling Guarantees

✅ Guaranteed:

• Valid tool names (no hallucinated tools)
• Valid arguments (via @Generable)
• Structural correctness

❌ Not guaranteed:

• Tool will be called (model might not need it)
• Specific argument values (model decides based on context)

Real-World Example: Itinerary Planner

struct FindPointsOfInterestTool: Tool {
    let name = "findPointsOfInterest"
    let description = "Find restaurants, museums, parks near a landmark"

    let landmark: String

    @Generable
    struct Arguments {
        let category: Category

        @Generable
        enum Category {
            case restaurant
            case museum
            case park
            case marina
        }
    }

    func call(arguments: Arguments) async throws -> ToolOutput {
        // Use MapKit
        let request = MKLocalSearch.Request()
        request.naturalLanguageQuery = "\(arguments.category) near \(landmark)"

        let search = MKLocalSearch(request: request)
        let response = try await search.start()

        let names = response.mapItems.prefix(5).map { $0.name ?? "" }
        return ToolOutput(names.joined(separator: ", "))
    }
}

From WWDC 259 summary: "Tool fetches points of interest from MapKit. Model uses world knowledge to determine promising categories."

When to Use Tools

✅ Use for:

• Weather data
• Map/location queries
• Contact information
• Calendar events
• External APIs

❌ Don't use for:

• Data model already has
• Information in prompt/instructions
• Simple calculations (model can do these)

Time Cost

Simple tool: 20-25 minutes Complex tool with state: 30-40 minutes

---

Pattern 5: Context Management (~1500 words)

Use when: Multi-turn conversations that might exceed 4096 token limit.

The Problem

// Long conversation...
for i in 1...100 {
    let response = try await session.respond(to: "Question \(i)")
    // Eventually...
    // Error: exceededContextWindowSize
}

Context window: 4096 tokens (input + output combined) Average: ~3 characters per token in English

Rough calculation:

• 4096 tokens ≈ 12,000 characters
• ≈ 2,000-3,000 words total

Long conversation or verbose prompts/responses → Exceed limit

Handling Context Overflow

Basic: Start fresh session

var session = LanguageModelSession()

do {
    let response = try await session.respond(to: prompt)
    print(response.content)
} catch LanguageModelSession.GenerationError.exceededContextWindowSize {
    // New session, no history
    session = LanguageModelSession()
}

// WWDC 301:3:37

Problem: Loses entire conversation history.

Better: Condense Transcript

var session = LanguageModelSession()

do {
    let response = try await session.respond(to: prompt)
} catch LanguageModelSession.GenerationError.exceededContextWindowSize {
    // New session with condensed history
    session = condensedSession(from: session)
}

func condensedSession(from previous: LanguageModelSession) -> LanguageModelSession {
    let allEntries = previous.transcript.entries
    var condensedEntries = [Transcript.Entry]()

    // Always include first entry (instructions)
    if let first = allEntries.first {
        condensedEntries.append(first)

        // Include last entry (most recent context)
        if allEntries.count > 1, let last = allEntries.last {
            condensedEntries.append(last)
        }
    }

    let condensedTranscript = Transcript(entries: condensedEntries)
    return LanguageModelSession(transcript: condensedTranscript)
}

// WWDC 301:3:55

Why this works:

• Instructions always preserved
• Recent context retained
• Total tokens drastically reduced

Advanced: Summarize Middle Entries

For long conversations where recent context isn't enough:

func condensedSession(from previous: LanguageModelSession) -> LanguageModelSession {
    let entries = previous.transcript.entries

    guard entries.count > 3 else {
        return LanguageModelSession(transcript: previous.transcript)
    }

    // Keep first (instructions) and last (recent)
    var condensedEntries = [entries.first!]

    // Summarize middle entries
    let middleEntries = Array(entries[1..<entries.count-1])
    let summaryPrompt = """
        Summarize this conversation in 2-3 sentences:
        \(middleEntries.map { $0.content }.joined(separator: "\n"))
        """

    // Use Foundation Models itself to summarize!
    let summarySession = LanguageModelSession()
    let summary = try await summarySession.respond(to: summaryPrompt)

    condensedEntries.append(Transcript.Entry(content: summary.content))
    condensedEntries.append(entries.last!)

    return LanguageModelSession(transcript: Transcript(entries: condensedEntries))
}

"You could summarize parts of transcript with Foundation Models itself."

Preventing Context Overflow

1. Keep prompts concise:

// ❌ BAD
let prompt = """
    I want you to generate a comprehensive detailed analysis of this article
    with multiple sections including summary, key points, sentiment analysis,
    main arguments, counter arguments, logical fallacies, and conclusions...
    """

// ✅ GOOD
let prompt = "Summarize this article's key points"

2. Use tools for data: Instead of putting entire dataset in prompt, use tools to fetch on-demand.

3. Break complex tasks into steps:

// ❌ BAD - One massive generation
let response = try await session.respond(
    to: "Create 7-day itinerary with hotels, restaurants, activities..."
)

// ✅ GOOD - Multiple smaller generations
let overview = try await session.respond(to: "Create high-level 7-day plan")
for day in 1...7 {
    let details = try await session.respond(to: "Detail activities for day \(day)")
}

Monitoring Context Usage

"Each token in instructions and prompt adds latency. Longer outputs take longer."

Use Instruments (Foundation Models template) to:

• See token counts
• Identify verbose prompts
• Optimize context usage

Time Cost

Basic overflow handling: 5-10 minutes Condensing strategy: 15-20 minutes Advanced summarization: 30-40 minutes

---

Pattern 6: Sampling & Generation Options (~1000 words)

Use when: You need control over output randomness/determinism.

Understanding Sampling

Model generates output one token at a time:

1. Creates probability distribution for next token
2. Samples from distribution
3. Picks token
4. Repeats

Default: Random sampling → Different output each time

Deterministic Output (Greedy)

let response = try await session.respond(
    to: prompt,
    options: GenerationOptions(sampling: .greedy)
)

// WWDC 301:6:14

Use cases:

• Repeatable demos
• Testing/debugging
• Consistent results required

Caveat: Only holds for same model version. OS updates may change output.

Temperature Control

Low variance (conservative, focused):

let response = try await session.respond(
    to: prompt,
    options: GenerationOptions(temperature: 0.5)
)

High variance (creative, diverse):

let response = try await session.respond(
    to: prompt,
    options: GenerationOptions(temperature: 2.0)
)

// WWDC 301:6:14

Temperature scale:

• 0.1-0.5: Very focused, predictable
• 1.0 (default): Balanced
• 1.5-2.0: Creative, varied

Example use cases:

• Low temp: Fact extraction, classification
• High temp: Creative writing, brainstorming

When to Adjust Sampling

✅ Greedy for:

• Unit tests
• Demos
• Consistency critical

✅ Low temperature for:

• Factual tasks
• Classification
• Extraction

✅ High temperature for:

• Creative content
• Story generation
• Varied NPC dialog

Time Cost

Implementation: 2-3 minutes (one line change)

---

Pressure Scenarios

Scenario 1: "Just Use ChatGPT API" (~1000 words)

Context: You're implementing a new AI feature. PM suggests using ChatGPT API for "better results."

Pressure signals:

• 👔 Authority: PM outranks you
• 💸 Existing integration: Team already uses OpenAI for other features
• ⏰ Speed: "ChatGPT is proven, Foundation Models is new"

Rationalization traps:

• "PM knows best"
• "ChatGPT gives better answers"
• "Faster to implement with existing code"

Why this fails:

1. Privacy violation: User data sent to external server

   • Medical notes, financial docs, personal messages
   • Violates user expectation of on-device privacy
   • Potential GDPR/privacy law issues

2. Cost: Every API call costs money

   • Foundation Models is free
   • Scale to millions of users = massive costs

3. Offline unavailable: Requires internet

   • Airplane mode, poor signal → feature broken
   • Foundation Models works offline

4. Latency: Network round-trip adds 500-2000ms

   • Foundation Models: On-device, <100ms startup

When ChatGPT IS appropriate:

• World knowledge required (e.g. "Who is the president of France?")
• Complex reasoning (multi-step logic, math proofs)
• Very long context (>4096 tokens)

Mandatory response:

"I understand ChatGPT delivers great results for certain tasks. However,
for this feature, Foundation Models is the right choice for three critical reasons:

1. **Privacy**: This feature processes [medical notes/financial data/personal content].
   Users expect this data stays on-device. Sending to external API violates that trust
   and may have compliance issues.

2. **Cost**: At scale, ChatGPT API calls cost $X per 1000 requests. Foundation Models
   is free. For Y million users, that's $Z annually we can avoid.

3. **Offline capability**: Foundation Models works without internet. Users in airplane
   mode or with poor signal still get full functionality.

**When to use ChatGPT**: If this feature required world knowledge or complex reasoning,
ChatGPT would be the right choice. But this is [summarization/extraction/classification],
which is exactly what Foundation Models is optimized for.

**Time estimate**: Foundation Models implementation: 15-20 minutes.
Privacy compliance review for ChatGPT: 2-4 weeks."

Time saved: Privacy compliance review vs correct implementation: 2-4 weeks vs 20 minutes

---

Scenario 2: "Parse JSON Manually" (~1000 words)

Context: Teammate suggests prompting for JSON, parsing with JSONDecoder. Claims it's "simple and familiar."

Pressure signals:

• ⏰ Deadline: Ship in 2 days
• 📚 Familiarity: "Everyone knows JSON"
• 🔧 Existing code: Already have JSON parsing utilities

Rationalization traps:

• "JSON is standard"
• "We parse JSON everywhere already"
• "Faster than learning new API"

Why this fails:

1. Hallucinated keys: Model outputs {firstName: "John"} when you expect {name: "John"}

   • JSONDecoder crashes: keyNotFound
   • No compile-time safety

2. Invalid JSON: Model might output:

   Here's the person: {name: "John", age: 30}
   

   • Not valid JSON (preamble text)
   • Parsing fails

3. No type safety: Manual string parsing, prone to errors

Real-world example:

// ❌ BAD - Will fail
let prompt = "Generate a person with name and age as JSON"
let response = try await session.respond(to: prompt)

// Model outputs: {"firstName": "John Smith", "years": 30}
// Your code expects: {"name": ..., "age": ...}
// CRASH: keyNotFound(name)

Debugging time: 2-4 hours finding edge cases, writing parsing hacks

Correct approach:

// ✅ GOOD - 15 minutes, guaranteed to work
@Generable
struct Person {
    let name: String
    let age: Int
}

let response = try await session.respond(
    to: "Generate a person",
    generating: Person.self
)
// response.content is type-safe Person, always valid

Mandatory response:

"I understand JSON parsing feels familiar, but for LLM output, @Generable is objectively
better for three technical reasons:

1. **Constrained decoding guarantees structure**: Model can ONLY generate valid Person
   instances. Impossible to get wrong keys, invalid JSON, or missing fields.

2. **No parsing code needed**: Framework handles parsing automatically. Zero chance of
   parsing bugs.

3. **Compile-time safety**: If we change Person struct, compiler catches all issues.
   Manual JSON parsing = runtime crashes.

**Real cost**: Manual JSON approach will hit edge cases. Debugging 'keyNotFound' crashes
takes 2-4 hours. @Generable implementation takes 15 minutes and has zero parsing bugs.

**Analogy**: This is like choosing Swift over Objective-C for new code. Both work, but
Swift's type safety prevents entire categories of bugs."

Time saved: 4-8 hours debugging vs 15 minutes correct implementation

---

Scenario 3: "One Big Prompt" (~1000 words)

Context: Feature requires extracting name, date, amount, category from invoice. Teammate suggests one prompt: "Extract all information."

Pressure signals:

• 🏗️ Architecture: "Simpler with one API call"
• ⏰ Speed: "Why make it complicated?"
• 📉 Complexity: "More prompts = more code"

Rationalization traps:

• "Simpler is better"
• "One prompt means less code"
• "Model is smart enough"

Why this fails:

1. Context overflow: Complex prompt + large invoice → Exceeds 4096 tokens
2. Poor results: Model tries to do too much at once, quality suffers
3. Slow generation: One massive response takes 5-8 seconds
4. All-or-nothing: If one field fails, entire generation fails

Better approach: Break into tasks + use tools

// ❌ BAD - One massive prompt
let prompt = """
    Extract from this invoice:
    - Vendor name
    - Invoice date
    - Total amount
    - Line items (description, quantity, price each)
    - Payment terms
    - Due date
    - Tax amount
    ...
    """
// 4 seconds, poor quality, might exceed context

// ✅ GOOD - Structured extraction with focused prompts
@Generable
struct InvoiceBasics {
    let vendor: String
    let date: String
    let amount: Double
}

let basics = try await session.respond(
    to: "Extract vendor, date, and amount",
    generating: InvoiceBasics.self
) // 0.5 seconds, axiom-high quality

@Generable
struct LineItem {
    let description: String
    let quantity: Int
    let price: Double
}

let items = try await session.respond(
    to: "Extract line items",
    generating: [LineItem].self
) // 1 second, axiom-high quality

// Total: 1.5 seconds, better quality, graceful partial failures

Mandatory response:

"I understand the appeal of one simple API call. However, this specific task requires
a different approach:

1. **Context limits**: Invoice + complex extraction prompt will likely exceed 4096 token
   limit. Multiple focused prompts stay well under limit.

2. **Better quality**: Model performs better with focused tasks. 'Extract vendor name'
   gets 95%+ accuracy. 'Extract everything' gets 60-70%.

3. **Faster perceived performance**: Multiple prompts with streaming show progressive
   results. Users see vendor name in 0.5s, not waiting 5s for everything.

4. **Graceful degradation**: If line items fail, we still have basics. All-or-nothing
   approach means total failure.

**Implementation**: Breaking into 3-4 focused extractions takes 30 minutes. One big
prompt takes 2-3 hours debugging why it hits context limit and produces poor results."

Time saved: 2-3 hours debugging vs 30 minutes proper design

---

Performance Optimization

1. Prewarm Session (~200 words)

Problem: First generation takes 1-2 seconds just to load model.

Solution: Create session before user interaction.

class ViewModel: ObservableObject {
    private var session: LanguageModelSession?

    init() {
        // Prewarm on init, not when user taps button
        Task {
            self.session = LanguageModelSession(instructions: "...")
        }
    }

    func generate(prompt: String) async throws -> String {
        let response = try await session!.respond(to: prompt)
        return response.content
    }
}

"Prewarming session before user interaction reduces initial latency."

Time saved: 1-2 seconds off first generation

---

2. includeSchemaInPrompt: false (~200 words)

Problem: @Generable schemas inserted into prompt, increases token count.

Solution: For subsequent requests with same schema, skip insertion.

let firstResponse = try await session.respond(
    to: "Generate first person",
    generating: Person.self
    // Schema inserted automatically
)

// Subsequent requests with SAME schema
let secondResponse = try await session.respond(
    to: "Generate another person",
    generating: Person.self,
    options: GenerationOptions(includeSchemaInPrompt: false)
)

"Setting includeSchemaInPrompt to false decreases token count and latency for subsequent requests."

When to use: Multi-turn with same @Generable type

Time saved: 10-20% latency reduction per request

---

3. Property Order for Streaming UX (~200 words)

Problem: User waits for entire generation.

Solution: Put important properties first, stream to show early.

// ✅ GOOD - Title shows immediately
@Generable
struct Article {
    var title: String      // Shows in 0.2s
    var summary: String    // Shows in 0.8s
    var fullText: String   // Shows in 2.5s
}

// ❌ BAD - Wait for everything
@Generable
struct Article {
    var fullText: String   // User waits 2.5s
    var title: String
    var summary: String
}

UX impact: Perceived latency drops from 2.5s to 0.2s

---

4. Foundation Models Instrument (~100 words)

Use Instruments app with Foundation Models template to:

• Profile latency of each request
• See token counts (input/output)
• Identify optimization opportunities
• Quantify improvements

"New Instruments profiling template lets you observe areas of optimization and quantify improvements."

Access: Instruments → Create → Foundation Models template

---

Checklist

Before shipping Foundation Models features:

Required Checks

• Availability checked before creating session
• Using @Generable for structured output (not manual JSON)
• Handling context overflow (exceededContextWindowSize)
• Handling guardrail violations (guardrailViolation)
• Handling unsupported language (unsupportedLanguageOrLocale)
• Streaming for long generations (>1 second)
• Not blocking UI (using Task {} for async)
• Tools for external data (not prompting for weather/locations)
• Prewarmed session if latency-sensitive

Best Practices

• Instructions are concise (not verbose)
• Never interpolating user input into instructions
• Property order optimized for streaming UX
• Using appropriate temperature/sampling
• Tested on real device (not just simulator)
• Profiled with Instruments (Foundation Models template)
• Error handling shows graceful UI messages
• Tested offline (airplane mode)
• Tested with long conversations (context handling)

Model Capability

• Not using for world knowledge
• Not using for complex reasoning
• Use case is: summarization, extraction, classification, or generation
• Have fallback if unavailable (show message, disable feature)

---

Resources

WWDC: 286, 259, 301

Skills: axiom-foundation-models-diag, axiom-foundation-models-ref

---

Last Updated: 2025-12-03 Version: 1.0.0 Target: iOS 26+, macOS 26+, iPadOS 26+, axiom-visionOS 26+