Foundation Models Reference

Complete API reference for Apple's Foundation Models framework covering on-device AI with LanguageModelSession, @Generable structured output, streaming, tool calling, and context management.

Overview

Comprehensive guide to Apple Intelligence based on WWDC 2025 sessions 286, 259, and 301, covering all 26 official code examples, @Generable structured output, streaming with PartiallyGenerated, Tool protocol, and dynamic schemas.

What This Reference Covers

LanguageModelSession

• Session creation and configuration
• Text generation with respond()
• Streaming with respond(stream:)
• Context management and limits
• Generation options (temperature, sampling)

@Generable Structured Output

• Automatic Swift type generation
• @Guide constraints and enums
• Nested generable types
• Custom descriptions
• Array generation

Streaming

• AsyncSequence patterns
• PartiallyGenerated for progressive UI
• Error handling mid-stream
• Cancellation

Tool Calling

• Tool protocol implementation
• Parameter passing
• Tool result handling
• Multi-turn conversations

Dynamic Schemas

• Runtime schema construction
• Conditional field inclusion
• Validation patterns

When to Use This Reference

Use this reference when:

• Implementing on-device AI features
• Defining @Generable output types
• Adding streaming responses to UI
• Building tools for the model
• Managing context window limits
• Debugging guardrail violations

Key Patterns

Basic Text Generation

import FoundationModels

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

@Generable Structured Output

@Generable
struct MovieReview {
    @Guide(description: "1-5 star rating")
    var rating: Int

    var summary: String
    var pros: [String]
    var cons: [String]
}

let review: MovieReview = try await session.respond(
    to: "Review the movie Inception",
    generating: MovieReview.self
)

Streaming Responses

for try await partial in session.respond(
    to: prompt,
    generating: Summary.self,
    stream: true
) {
    if case .partial(let summary) = partial {
        // Update UI with partial.content
    }
}

Tool Protocol

@Tool
struct SearchTool: Tool {
    static let description = "Search the web"

    @Parameter(description: "Search query")
    var query: String

    func call() async throws -> String {
        // Perform search, return results
    }
}

Complete API Coverage

This reference includes:

• All 26 WWDC 2025 code examples with annotations
• Complete LanguageModelSession API
• @Generable macro with all options
• @Guide constraints and validation
• Tool protocol implementation
• Streaming patterns with PartiallyGenerated
• Context management strategies
• Error handling and guardrails
• Performance profiling with Instruments

Built-in Use Cases

// Content tagging
let tags = try await session.contentTagging(for: article)

// Summarization
let summary = try await session.summarize(content)

Error Handling

do {
    let response = try await session.respond(to: prompt)
} catch let error as LanguageModelError {
    switch error {
    case .contextExceeded:
        // Reduce input size
    case .guardrailViolation:
        // Content not allowed
    case .unavailable:
        // Model not ready
    }
}

Related Resources

• foundation-models — Discipline-enforcing skill with anti-patterns
• foundation-models-diag — Systematic troubleshooting
• WWDC 2025/286 — Introducing Foundation Models
• WWDC 2025/259 — Build intelligent apps
• WWDC 2025/301 — Advanced Foundation Models

Documentation Scope

This is a reference skill — comprehensive API guide without mandatory workflows.

Reference includes

• Complete Foundation Models API (iOS 26+)
• All WWDC 2025 code examples
• @Generable and @Guide documentation
• Tool protocol patterns
• Streaming and context management

Vs Diagnostic: Reference skills provide information. Diagnostic skills enforce workflows and handle pressure scenarios.