axiom-foundation-models-diag（Axiom 基础模型诊断）

Foundation Models Diagnostics

Overview

Foundation Models issues manifest as context window exceeded errors, guardrail violations, slow generation, availability failures, and unexpected output. Core principle 80% of Foundation Models problems stem from misunderstanding model capabilities (3B parameter device-scale model, not world knowledge), context limits (4096 tokens), or availability requirements—not framework bugs.

Red Flags — Suspect Foundation Models Issue

If you see ANY of these, suspect a Foundation Models misunderstanding, not framework breakage:

• Generation takes >5 seconds
• Error: exceededContextWindowSize
• Error: guardrailViolation
• Error: unsupportedLanguageOrLocale
• Model gives hallucinated/wrong output
• UI freezes during generation
• Feature works in simulator but not on device
• ❌ FORBIDDEN "Foundation Models is broken, we need a different AI"

- Foundation Models powers Apple Intelligence across millions of devices - Wrong output = wrong use case (world knowledge vs summarization) - Do not rationalize away the issue—diagnose it

Critical distinction Foundation Models is a device-scale model (3B parameters) optimized for summarization, extraction, classification—NOT world knowledge or complex reasoning. Using it for the wrong task guarantees poor results.

Mandatory First Steps

ALWAYS run these FIRST (before changing code):

// 1. Check availability
let availability = SystemLanguageModel.default.availability

switch availability {
case .available:
    print("✅ Available")
case .unavailable(let reason):
    print("❌ Unavailable: \(reason)")
    // Possible reasons:
    // - Device not Apple Intelligence-capable
    // - Region not supported
    // - User not opted in
}

// Record: "Available? Yes/no, reason if not"

// 2. Check supported languages
let supported = SystemLanguageModel.default.supportedLanguages
print("Supported languages: \(supported)")
print("Current locale: \(Locale.current.language)")

if !supported.contains(Locale.current.language) {
    print("⚠️ Current language not supported!")
}

// Record: "Language supported? Yes/no"

// 3. Check context usage
let session = LanguageModelSession()
// After some interactions:
print("Transcript entries: \(session.transcript.entries.count)")

// Rough estimation (not exact):
let transcriptText = session.transcript.entries
    .map { $0.content }
    .joined()
print("Approximate chars: \(transcriptText.count)")
print("Rough token estimate: \(transcriptText.count / 3)")
// 4096 token limit ≈ 12,000 characters

// Record: "Approaching context limit? Yes/no"

// 4. Profile with Instruments
// Run with Foundation Models Instrument template
// Check:
// - Initial model load time
// - Token counts (input/output)
// - Generation time per request
// - Areas for optimization

// Record: "Latency profile: [numbers from Instruments]"

// 5. Inspect transcript for debugging
print("Full transcript:")
for entry in session.transcript.entries {
    print("Entry: \(entry.content.prefix(100))...")
}

// Record: "Any unusual entries? Repeated content?"

What this tells you

• Unavailable → Proceed to Pattern 1a/1b/1c (availability issues)
• Context exceeded → Proceed to Pattern 2a (token limit)
• Guardrail error → Proceed to Pattern 2b (content policy)
• Language error → Proceed to Pattern 2c (unsupported language)
• Wrong output → Proceed to Pattern 3a/3b/3c (output quality)
• Slow generation → Proceed to Pattern 4a/4b/4c/4d (performance)
• UI frozen → Proceed to Pattern 5a (main thread blocking)

MANDATORY INTERPRETATION

Before changing ANY code, identify ONE of these:

1. If availability =.unavailable → Device/region/opt-in issue (not code bug)
2. If error is exceededContextWindowSize → Too many tokens (condense transcript)
3. If error is guardrailViolation → Content policy triggered (not model failure)
4. If error is unsupportedLanguageOrLocale → Language not supported (check supported list)
5. If output is hallucinated → Wrong use case (world knowledge vs extraction)
6. If generation >5 seconds → Not streaming or need optimization
7. If UI frozen → Calling on main thread (use Task {})

If diagnostics are contradictory or unclear

• STOP. Do NOT proceed to patterns yet
• Add detailed logging to every respond() call
• Run with Instruments Foundation Models template
• Establish baseline: what's actually happening vs what you assumed

Decision Tree

Foundation Models problem?
│
├─ Won't start?
│  ├─ .unavailable → Availability issue
│  │  ├─ Device not capable? → Pattern 1a (device requirement)
│  │  ├─ Region restriction? → Pattern 1b (regional availability)
│  │  └─ User not opted in? → Pattern 1c (Settings check)
│  │
├─ Generation fails?
│  ├─ exceededContextWindowSize → Context limit
│  │  └─ Long conversation or verbose prompts? → Pattern 2a (condense)
│  │
│  ├─ guardrailViolation → Content policy
│  │  └─ Sensitive or inappropriate content? → Pattern 2b (handle gracefully)
│  │
│  ├─ unsupportedLanguageOrLocale → Language issue
│  │  └─ Non-English or unsupported language? → Pattern 2c (language check)
│  │
│  └─ Other error → General error handling
│     └─ Unknown error type? → Pattern 2d (catch-all)
│
├─ Output wrong?
│  ├─ Hallucinated facts → Wrong model use
│  │  └─ Asking for world knowledge? → Pattern 3a (use case mismatch)
│  │
│  ├─ Wrong structure → Parsing issue
│  │  └─ Manual JSON parsing? → Pattern 3b (use @Generable)
│  │
│  ├─ Missing data → Tool needed
│  │  └─ Need external information? → Pattern 3c (tool calling)
│  │
│  └─ Inconsistent output → Sampling issue
│     └─ Different results each time? → Pattern 3d (temperature/greedy)
│
├─ Too slow?
│  ├─ Initial delay (1-2s) → Model loading
│  │  └─ First request slow? → Pattern 4a (prewarm)
│  │
│  ├─ Long wait for results → Not streaming
│  │  └─ User waits 3-5s? → Pattern 4b (streaming)
│  │
│  ├─ Verbose schema → Token overhead
│  │  └─ Large @Generable type? → Pattern 4c (includeSchemaInPrompt)
│  │
│  └─ Complex prompt → Too much processing
│     └─ Massive prompt or task? → Pattern 4d (break down)
│
└─ UI frozen?
   └─ Main thread blocked → Async issue
      └─ App unresponsive during generation? → Pattern 5a (Task {})

Diagnostic Patterns

Pattern 1a: Device Not Capable

Symptom:

• SystemLanguageModel.default.availability =.unavailable
• Reason: Device not Apple Intelligence-capable

Diagnosis:

let availability = SystemLanguageModel.default.availability

switch availability {
case .available:
    print("✅ Available")
case .unavailable(let reason):
    print("❌ Reason: \(reason)")
    // Check if device-related
}

Fix:

// ❌ BAD - No availability UI
let session = LanguageModelSession() // Crashes on unsupported devices

// ✅ GOOD - Graceful UI
struct AIFeatureView: View {
    @State private var availability = SystemLanguageModel.default.availability

    var body: some View {
        switch availability {
        case .available:
            AIContentView()
        case .unavailable:
            VStack {
                Image(systemName: "cpu")
                Text("AI features require Apple Intelligence")
                    .font(.headline)
                Text("Available on iPhone 15 Pro and later")
                    .font(.caption)
                    .foregroundColor(.secondary)
            }
        }
    }
}

Time cost: 5-10 minutes to add UI

Pattern 1b: Regional Availability

Symptom:

• Feature works for some users, not others
• .unavailable due to region restrictions

Diagnosis: Foundation Models requires:

• Supported region (e.g., US, UK, Australia initially)
• May expand over time

Fix:

// ✅ GOOD - Clear messaging
switch SystemLanguageModel.default.availability {
case .available:
    // proceed
case .unavailable(let reason):
    // Show region-specific message
    Text("AI features not yet available in your region")
    Text("Check Settings → Apple Intelligence for availability")
}

Time cost: 5 minutes

Pattern 1c: User Not Opted In

Symptom:

• Device capable, region supported
• Still.unavailable

Diagnosis: User must opt in to Apple Intelligence in Settings

Fix:

// ✅ GOOD - Direct user to settings
switch SystemLanguageModel.default.availability {
case .available:
    // proceed
case .unavailable:
    VStack {
        Text("Enable Apple Intelligence")
        Text("Settings → Apple Intelligence → Enable")
        Button("Open Settings") {
            if let url = URL(string: UIApplication.openSettingsURLString) {
                UIApplication.shared.open(url)
            }
        }
    }
}

Time cost: 10 minutes

Pattern 2a: Context Window Exceeded

Symptom:

Error: LanguageModelSession.GenerationError.exceededContextWindowSize

Diagnosis:

• 4096 token limit (input + output)
• Long conversations accumulate tokens
• Verbose prompts eat into limit

Fix:

// ❌ BAD - Unhandled error
let response = try await session.respond(to: prompt)
// Crashes after ~10-15 turns

// ✅ GOOD - Condense transcript
var session = LanguageModelSession()

do {
    let response = try await session.respond(to: prompt)
} catch LanguageModelSession.GenerationError.exceededContextWindowSize {
    // Condense and continue
    session = condensedSession(from: session)
    let response = try await session.respond(to: prompt)
}

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

    guard entries.count > 2 else {
        return LanguageModelSession(transcript: previous.transcript)
    }

    // Keep: first (instructions) + last (recent context)
    var condensed = [entries.first!, entries.last!]

    let transcript = Transcript(entries: condensed)
    return LanguageModelSession(transcript: transcript)
}

Time cost: 15-20 minutes to implement condensing

Pattern 2b: Guardrail Violation

Symptom:

Error: LanguageModelSession.GenerationError.guardrailViolation

Diagnosis:

• User input triggered content policy
• Violence, hate speech, illegal activities
• Model refuses to generate

Fix:

// ✅ GOOD - Graceful handling
do {
    let response = try await session.respond(to: userInput)
    print(response.content)
} catch LanguageModelSession.GenerationError.guardrailViolation {
    // Show user-friendly message
    print("I can't help with that request")
    // Log for review (but don't show user input to avoid storing harmful content)
}

Time cost: 5-10 minutes

Pattern 2c: Unsupported Language

Symptom:

Error: LanguageModelSession.GenerationError.unsupportedLanguageOrLocale

Diagnosis: User input in language model doesn't support

Fix:

// ❌ BAD - No language check
let response = try await session.respond(to: userInput)
// Crashes if unsupported language

// ✅ GOOD - Check first
let supported = SystemLanguageModel.default.supportedLanguages

guard supported.contains(Locale.current.language) else {
    // Show disclaimer
    print("Language not supported. Currently supports: \(supported)")
    return
}

// Also handle errors
do {
    let response = try await session.respond(to: userInput)
} catch LanguageModelSession.GenerationError.unsupportedLanguageOrLocale {
    print("Please use English or another supported language")
}

Time cost: 10 minutes

Pattern 2d: General Error Handling

Symptom: Unknown error types

Fix:

// ✅ GOOD - Comprehensive error handling
do {
    let response = try await session.respond(to: prompt)
    print(response.content)
} catch LanguageModelSession.GenerationError.exceededContextWindowSize {
    // Handle context overflow
    session = condensedSession(from: session)
} catch LanguageModelSession.GenerationError.guardrailViolation {
    // Handle content policy
    showMessage("Cannot generate that content")
} catch LanguageModelSession.GenerationError.unsupportedLanguageOrLocale {
    // Handle language issue
    showMessage("Language not supported")
} catch {
    // Catch-all for unexpected errors
    print("Unexpected error: \(error)")
    showMessage("Something went wrong. Please try again.")
}

Time cost: 10-15 minutes

Pattern 3a: Hallucinated Output (Wrong Use Case)

Symptom:

• Model gives factually incorrect answers
• Makes up information

Diagnosis: Using model for world knowledge (wrong use case)

Fix:

// ❌ BAD - Wrong use case
let prompt = "Who is the president of France?"
let response = try await session.respond(to: prompt)
// Will hallucinate or give outdated info

// ✅ GOOD - Use server LLM for world knowledge
// Foundation Models is for:
// - Summarization
// - Extraction
// - Classification
// - Content generation

// OR: Use Tool calling with external data source
struct GetFactTool: Tool {
    let name = "getFact"
    let description = "Fetch factual information from verified source"

    @Generable
    struct Arguments {
        let query: String
    }

    func call(arguments: Arguments) async throws -> ToolOutput {
        // Fetch from Wikipedia API, news API, etc.
        let fact = await fetchFactFromAPI(arguments.query)
        return ToolOutput(fact)
    }
}

Time cost: 20-30 minutes to implement tool OR switch to appropriate AI

Pattern 3b: Wrong Structure (Not Using @Generable)

Symptom:

• Parsing errors
• Invalid JSON
• Wrong keys

Diagnosis: Manual JSON parsing instead of @Generable

Fix:

// ❌ BAD - Manual parsing
let prompt = "Generate person 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

// ✅ GOOD - @Generable
@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, guaranteed structure

Time cost: 10 minutes to convert to @Generable

Pattern 3c: Missing Data (Need Tool)

Symptom:

• Model doesn't have required information
• Output is vague or generic

Diagnosis: Need external data (weather, locations, contacts)

Fix:

// ❌ BAD - No external data
let response = try await session.respond(
    to: "What's the weather in Tokyo?"
)
// Will make up weather data

// ✅ GOOD - Tool calling
import WeatherKit

struct GetWeatherTool: Tool {
    let name = "getWeather"
    let description = "Get current weather for a city"

    @Generable
    struct Arguments {
        let city: String
    }

    func call(arguments: Arguments) async throws -> ToolOutput {
        // Fetch real weather
        let weather = await WeatherService.shared.weather(for: arguments.city)
        return ToolOutput("Temperature: \(weather.temperature)°F")
    }
}

let session = LanguageModelSession(tools: [GetWeatherTool()])
let response = try await session.respond(to: "What's the weather in Tokyo?")
// Uses real weather data

Time cost: 20-30 minutes to implement tool

Pattern 3d: Inconsistent Output (Sampling)

Symptom:

• Different output every time for same prompt
• Need consistent results for testing

Diagnosis: Random sampling (default behavior)

Fix:

// Default: Random sampling
let response1 = try await session.respond(to: "Write a haiku")
let response2 = try await session.respond(to: "Write a haiku")
// Different every time

// ✅ For deterministic output (testing/demos)
let response = try await session.respond(
    to: "Write a haiku",
    options: GenerationOptions(sampling: .greedy)
)
// Same output for same prompt (given same model version)

// ✅ For low variance
let response = try await session.respond(
    to: "Classify this article",
    options: GenerationOptions(temperature: 0.5)
)
// Slightly varied but focused

// ✅ For high creativity
let response = try await session.respond(
    to: "Write a creative story",
    options: GenerationOptions(temperature: 2.0)
)
// Very diverse output

Time cost: 2-5 minutes

Pattern 4a: Initial Latency (Prewarm)

Symptom:

• First generation takes 1-2 seconds to start
• Subsequent requests faster

Diagnosis: Model loading time

Fix:

// ❌ BAD - Load on user interaction
Button("Generate") {
    Task {
        let session = LanguageModelSession() // 1-2s delay here
        let response = try await session.respond(to: prompt)
    }
}

// ✅ GOOD - Prewarm on init
class ViewModel: ObservableObject {
    private var session: LanguageModelSession?

    init() {
        // Prewarm before user interaction
        Task {
            self.session = LanguageModelSession(instructions: "...")
        }
    }

    func generate(prompt: String) async throws -> String {
        guard let session = session else {
            // Fallback if not ready
            self.session = LanguageModelSession()
            return try await self.session!.respond(to: prompt).content
        }
        return try await session.respond(to: prompt).content
    }
}

Time cost: 10 minutes Latency saved: 1-2 seconds on first request

Pattern 4b: Long Generation (Streaming)

Symptom:

• User waits 3-5 seconds seeing nothing
• Then entire result appears at once

Diagnosis: Not streaming long generations

Fix:

// ❌ BAD - No streaming
let response = try await session.respond(
    to: "Generate 5-day itinerary",
    generating: Itinerary.self
)
// User waits 4 seconds seeing nothing

// ✅ GOOD - Streaming
@Generable
struct Itinerary {
    var destination: String
    var days: [DayPlan]
}

let stream = session.streamResponse(
    to: "Generate 5-day itinerary to Tokyo",
    generating: Itinerary.self
)

for try await partial in stream {
    // Update UI incrementally
    self.itinerary = partial
}
// User sees destination in 0.5s, then days progressively

Time cost: 15-20 minutes Perceived latency: 0.5s vs 4s

Pattern 4c: Large Schema Overhead

Symptom:

• Subsequent requests with same @Generable type slow

Diagnosis: Schema re-inserted into prompt every time

Fix:

// First request - schema inserted automatically
let first = try await session.respond(
    to: "Generate first person",
    generating: Person.self
)

// ✅ Subsequent requests - skip schema insertion
let second = try await session.respond(
    to: "Generate another person",
    generating: Person.self,
    options: GenerationOptions(includeSchemaInPrompt: false)
)

Time cost: 2 minutes Latency saved: 10-20% per request

Pattern 4d: Complex Prompt (Break Down)

Symptom:

• Generation takes >5 seconds
• Poor quality results

Diagnosis: Prompt too complex for single generation

Fix:

// ❌ BAD - One massive prompt
let prompt = """
    Generate complete 7-day itinerary with hotels, restaurants,
    activities, transportation, budget, tips, and local customs
    """
// 5-8 seconds, poor quality

// ✅ GOOD - Break into steps
let overview = try await session.respond(
    to: "Generate high-level 7-day plan for Tokyo"
)

var dayDetails: [DayPlan] = []
for day in 1...7 {
    let detail = try await session.respond(
        to: "Detail activities and restaurants for day \(day) in Tokyo",
        generating: DayPlan.self
    )
    dayDetails.append(detail.content)
}
// Total time similar, but better quality and progressive results

Time cost: 20-30 minutes Quality improvement: Significantly better

Pattern 5a: UI Frozen (Main Thread Blocking)

Symptom:

• App unresponsive during generation
• UI freezes for seconds

Diagnosis: Calling respond() on main thread synchronously

Fix:

// ❌ BAD - Blocking main thread
Button("Generate") {
    let response = try await session.respond(to: prompt)
    // UI frozen for 2-5 seconds!
}

// ✅ GOOD - Async task
Button("Generate") {
    Task {
        do {
            let response = try await session.respond(to: prompt)
            // Update UI on main thread
            await MainActor.run {
                self.result = response.content
            }
        } catch {
            print("Error: \(error)")
        }
    }
}

Time cost: 5 minutes UX improvement: Massive (no frozen UI)

Production Crisis Scenario

Context

Situation: You just launched an AI-powered feature using Foundation Models. Within 2 hours:

• 20% of users report "AI feature doesn't work"
• App Store reviews dropping: "New AI broken"
• VP of Product emailing: "What's the ETA on fix?"
• Engineering manager: "Should we roll back?"

Pressure Signals:

• 🚨 Revenue impact: Feature is key selling point for new app version
• ⏰ Time pressure: "Fix it NOW"
• 👔 Executive visibility: VP watching
• 📉 Public reputation: App Store reviews visible to all

Rationalization Traps

DO NOT fall into these traps:

1. "Disable the feature"

- Loses product differentiation - Admits defeat - Doesn't learn what went wrong

1. "Roll back to previous version"

- Loses weeks of work - Doesn't fix root cause - Users still angry

1. "It works for me"

- Simulator ≠ real devices - Your device ≠ all devices - Ignores real problem

1. "Switch to ChatGPT API"

- Violates privacy - Expensive at scale - Doesn't address availability issue

MANDATORY Protocol

Phase 1: Identify (5 minutes)

// Check error distribution
// What percentage seeing what error?

// Run this on test devices:
let availability = SystemLanguageModel.default.availability

switch availability {
case .available:
    print("✅ Available")
case .unavailable(let reason):
    print("❌ Unavailable: \(reason)")
}

// Hypothesis:
// - If 20% unavailable → Availability issue (device/region/opt-in)
// - If 20% getting errors → Code bug
// - If 20% seeing wrong results → Use case mismatch

Results: Discover that 20% of users have devices without Apple Intelligence support.

Phase 2: Confirm (5 minutes)

// Check which devices affected
// iPhone 15 Pro+ = ✅ Available
// iPhone 15 = ❌ Unavailable
// iPhone 14 = ❌ Unavailable

// Conclusion: Availability issue, not code bug

Root cause: Feature assumes all users have Apple Intelligence. 20% don't.

Phase 3: Device Requirements (5 minutes)

Verify:

• Apple Intelligence requires iPhone 15 Pro or later
• Or iPad with M1+ chip
• Or Mac with Apple silicon

20% of user base = older devices

Phase 4: Implement Fix (15 minutes)

// ✅ Add availability check + graceful UI
struct AIFeatureView: View {
    @State private var availability = SystemLanguageModel.default.availability

    var body: some View {
        switch availability {
        case .available:
            // Show AI feature
            AIContentView()

        case .unavailable:
            // Graceful fallback
            VStack {
                Image(systemName: "sparkles")
                    .font(.largeTitle)
                    .foregroundColor(.secondary)

                Text("AI-Powered Features")
                    .font(.headline)

                Text("Available on iPhone 15 Pro and later")
                    .font(.subheadline)
                    .foregroundColor(.secondary)
                    .multilineTextAlignment(.center)

                // Offer alternative
                Button("Use Standard Mode") {
                    // Show non-AI fallback
                }
            }
        }
    }
}

Phase 5: Deploy (20 minutes)

1. Test on multiple devices (15 min)

- iPhone 15 Pro: ✅ Shows AI feature - iPhone 14: ✅ Shows graceful message - iPad Pro M1: ✅ Shows AI feature

1. Submit hotfix build (5 min)

Communication Template

To VP of Product (immediate):

Root cause identified:

The AI feature requires Apple Intelligence (iPhone 15 Pro+).
20% of our users have older devices. We didn't check availability.

Fix: Added availability check with graceful fallback UI.

Timeline:
- Hotfix ready: Now
- TestFlight: 10 minutes
- App Store submission: 30 minutes
- Review: 24-48 hours (requesting expedited)

Impact mitigation:
- 80% of users see working AI feature
- 20% see clear message + standard mode fallback
- No functionality lost, just graceful degradation

To Engineering Team:

Post-mortem items:
1. Add availability check to launch checklist
2. Test on non-Apple-Intelligence devices
3. Document device requirements clearly
4. Add analytics for availability status

Time Saved

• Panic path (disable/rollback): 2 hours of meetings + lost work
• Proper diagnosis: 45 minutes root cause → fix → deploy

What We Learned

1. Always check availability before creating session
2. Test on real devices across device generations
3. Graceful degradation better than feature removal
4. Clear messaging to users about requirements

Quick Reference Table

Cross-References

Related Axiom Skills:

• axiom-foundation-models — Discipline skill for anti-patterns, proper usage patterns, pressure scenarios
• axiom-foundation-models-ref — Complete API reference with all WWDC 2025 code examples

Apple Resources:

• Foundation Models Framework Documentation
• WWDC 2025-286: Meet the Foundation Models framework
• WWDC 2025-301: Deep dive into the Foundation Models framework
• Instruments Foundation Models Template

Last Updated: 2025-12-03 Version: 1.0.0 Skill Type: Diagnostic