foundation-models

Foundation Models

Apple's on-device LLM. Its context window (input + output combined) is model-version-dependent and can differ across OS releases — query it at runtime with SystemLanguageModel.default.contextSize rather than hardcoding a number. (At the time of writing, Apple documents the base model at 4,096 tokens, but treat that as a current value, not a fixed contract.) Optimized for summarization, extraction, classification, and generation. No network, no cost, no data leaves device.

Worked Example

Request: "Add article summarization with streaming to my app."

1. Check availability:

guard case .available = SystemLanguageModel.default.availability else {
    showUnavailableMessage()
    return
}

2. Define output type:

@Generable
struct ArticleSummary {
    @Guide(description: "One-sentence summary of the article's main point")
    var headline: String

    @Guide(.count(2...5), description: "Key takeaways in order of importance")
    var takeaways: [String]

    @Guide(.range(1...10), description: "Reading complexity score")
    var complexity: Int
}

3. Stream with progressive UI:

let session = LanguageModelSession(instructions: "Summarize articles concisely and accurately")

let stream = session.streamResponse(
    to: Prompt { "Summarize this article:"; articleText },
    generating: ArticleSummary.self
)

for try await partial in stream {
    withAnimation { self.summary = partial }
}
// partial.headline appears first, then takeaways fill in, then complexity

4. Handle errors:

catch LanguageModelSession.GenerationError.exceededContextWindowSize {
    // Article too long — chunk it or truncate
    session = LanguageModelSession(instructions: originalInstructions)
}

---

Anti-Patterns

1: Manual JSON Parsing

// WRONG: Manual JSON parsing — fragile, model might produce malformed output
let json = try JSONSerialization.jsonObject(with: response.content.data(using: .utf8)!)

// RIGHT: @Generable with constrained decoding — model cannot produce invalid structure
let person = try await session.respond(to: "Generate a person", generating: Person.self).content

2: Blocking UI with Synchronous Generation

// WRONG: User waits for entire response
self.text = try await session.respond(to: prompt).content

// RIGHT: Streaming for progressive display
for try await partial in session.streamResponse(to: prompt) {
    withAnimation { self.text = partial.content }
}

3: Context Overflow from Unbounded Conversations

// WRONG: Endless multi-turn — crashes when the context window fills

// RIGHT: Catch and recover
catch LanguageModelSession.GenerationError.exceededContextWindowSize {
    session = LanguageModelSession(instructions: originalInstructions)
}

The context window is TOTAL — instructions + schema + transcript + new prompt + output all count against it. Read the actual limit from SystemLanguageModel.default.contextSize; don't assume a fixed number.

4: User Input in Instructions (Prompt Injection)

// WRONG: User input in instructions
let session = LanguageModelSession(instructions: "Summarize: \(userInput)")

// RIGHT: User input in prompt only
let session = LanguageModelSession(instructions: "You summarize text concisely")
let response = try await session.respond(to: userInput)

Instructions are developer-controlled. Model trained to prioritize instructions over prompts.

5: Struct for Stateful Tools

Use class (not struct) when tools track state across calls. Struct copies lose mutations between calls.

6: Over-Complex Instructions Duplicating @Generable

@Generable and @Guide encode output structure at the decoding level. Don't repeat the schema in instructions — it wastes tokens (critical given the limited context window). Use instructions for tone and behavioral constraints only.

---

Decision Trees

Foundation Models vs Server API

Question	Choice
Privacy required / offline needed / avoid per-request cost?	FM
Summarization, extraction, or classification?	FM
World knowledge, complex reasoning, math, or translation?	Server API
Need more context than SystemLanguageModel.default.contextSize allows?	Server API

Both can coexist in one app.

Other Decisions

• @Generable vs plain text: Use @Generable when you need structured data or type safety. Plain text when just displaying to user.
• Tools vs prompt context: Tools for dynamic/real-time data, device data (contacts, calendar), external APIs. Prompt context for static, short data.
• New session vs reuse: Reuse for same topic. New session when context fills, topic changes, or you need different instructions/tools.

---

Pressure Scenarios

"Use ChatGPT API Instead"

FM: private, offline, no latency, no per-request cost, no API keys. Server API: world knowledge, complex reasoning, larger context, translation. Both can coexist — not either/or.

"One Big Prompt for Everything"

The context window is TOTAL (query it via SystemLanguageModel.default.contextSize). Keep instructions concise, use @Generable instead of describing format, chunk large inputs, monitor with tokenUsage(for:), catch exceededContextWindowSize for multi-turn.

"Skip Availability Checks"

Three unavailable states — .deviceNotEligible (permanent), .appleIntelligenceNotEnabled (user action), .modelNotReady (temporary). Not checking = crashes on unsupported devices. Check on scenePhase activation to catch state changes.

---

Diagnostics Quick Reference

Symptom	Error	Key Fix
Context too long	.exceededContextWindowSize	Fresh/condensed session
Content policy error	.guardrailViolation	Rephrase prompt, filter input
Language not supported	.unsupportedLanguageOrLocale	Fall back to server
Structured output fails	.decodingError	Verify nested @Generable, add @Guide
Too many requests	.rateLimited	Backoff, queue requests
Tool not called	Inspect session.transcript	Strengthen instructions and tool description
Slow response	Profile with Instruments	Pre-warm, reduce tokens, stream
Wrong output	Check @Guide constraints	Add descriptions, constrain range/count

Full triage procedures and production crisis playbook: references/diagnostics.md

---

References

• API Reference — Complete API with WWDC code examples
• Diagnostics — Error triage, Instruments workflow, production crisis defense
• WWDC Sessions: 286, 259, 301 (2025, historical anchors). Newer Foundation Models sessions ship with later releases (2026 / Xcode 27) — check the current WWDC catalog for updated guidance.