AI Agent Usage Guide

Core Architecture

Leap.load() (iOS/macOS)
    ↓
ModelRunner
    ↓
Conversation
    ↓
MessageResponse (streaming)

Installation

Swift Package Manager (Recommended)

// In Xcode: File → Add Package Dependencies
// Repository: https://github.com/Liquid4All/leap-ios.git
// Version: 0.9.2

dependencies: [
    .package(url: "https://github.com/Liquid4All/leap-ios.git", from: "0.9.2")
]

targets: [
    .target(
        name: "YourApp",
        dependencies: [
            .product(name: "LeapSDK", package: "leap-ios"),
            .product(name: "LeapModelDownloader", package: "leap-ios") // Optional
        ]
    )
]

CocoaPods

pod 'Leap-SDK', '~> 0.9.2'
pod 'Leap-Model-Downloader', '~> 0.9.2' # Optional

Loading Models

Method 1: Automatic Download (Recommended)

The simplest approach - specify model name and quantization, SDK handles everything:

import LeapSDK

// Load model with automatic download and caching
let modelRunner = try await Leap.load(
    model: "LFM2.5-1.2B-Instruct",
    quantization: "Q4_K_M"
) { progress, speed in
    // progress: Double (0.0 to 1.0)
    // speed: Int64 (bytes per second)
    print("Download progress: \(Int(progress * 100))% at \(speed) bytes/s")
}

Available models and quantizations: LEAP Model Library

Method 2: Download Without Loading

Separate download from loading for better control:

import LeapModelDownloader

let downloader = ModelDownloader()

// Download model to cache
let manifest = try await downloader.downloadModel(
    "LFM2.5-1.2B-Instruct",
    quantization: "Q4_K_M"
) { progress, speed in
    print("Progress: \(Int(progress * 100))%")
}

// Later, load from cache (no download)
let modelRunner = try await Leap.load(
    model: "LFM2.5-1.2B-Instruct",
    quantization: "Q4_K_M"
)

Method 3: Custom Manifest URL

Load from a custom manifest:

let manifestURL = URL(string: "https://your-server.com/model-manifest.json")!

let modelRunner = try await Leap.load(
    manifestURL: manifestURL,
    downloadProgressHandler: { progress, speed in
        print("Progress: \(Int(progress * 100))%")
    }
)

Method 4: Local Bundle (Legacy)

Load from a local .bundle or .gguf file:

guard let bundleURL = Bundle.main.url(forResource: "model", withExtension: "bundle") else {
    fatalError("Model bundle not found")
}

let modelRunner = try await Leap.load(
    url: bundleURL,
    options: LiquidInferenceEngineOptions(
        bundlePath: bundleURL.path,
        cpuThreads: 6,
        contextSize: 8192,
        nGpuLayers: 8  // Metal GPU acceleration on macOS
    )
)

Core Classes

`ModelRunner`

The loaded model instance. Create conversations from this:

protocol ModelRunner {
    func createConversation(systemPrompt: String?) -> Conversation
    func createConversationFromHistory(history: [ChatMessage]) -> Conversation
}

Usage:

let conversation = modelRunner.createConversation(
    systemPrompt: "Explain it to me like I'm 5 years old"
)

// Or restore from saved history
let savedHistory: [ChatMessage] = loadHistoryFromDisk()
let conversation = modelRunner.createConversationFromHistory(history: savedHistory)

`Conversation`

Manages chat history and generation:

class Conversation {
    let modelRunner: ModelRunner
    private(set) var history: [ChatMessage]
    private(set) var isGenerating: Bool

    func generateResponse(
        message: ChatMessage,
        generationOptions: GenerationOptions?
    ) -> AsyncThrowingStream<MessageResponse, Error>

    func generateResponse(
        userTextMessage: String,
        generationOptions: GenerationOptions?
    ) -> AsyncThrowingStream<MessageResponse, Error>

    func registerFunction(_ function: LeapFunction)
    func exportToJSON() throws -> [[String: Any]]
}

Properties:

history - Array of ChatMessage objects representing the conversation
isGenerating - Boolean indicating if generation is in progress

Key Methods:

generateResponse(message:generationOptions:) - Start generation, returns async stream
registerFunction(_:) - Register functions for function calling
exportToJSON() - Export conversation history to JSON

`ChatMessage`

Represents a single message in the conversation:

struct ChatMessage {
    var role: ChatMessageRole  // .system, .user, .assistant, .tool
    var content: [ChatMessageContent]
    var reasoningContent: String?  // For reasoning models
    var functionCalls: [LeapFunctionCall]?
}

enum ChatMessageRole: String {
    case user = "user"
    case system = "system"
    case assistant = "assistant"
    case tool = "tool"
}

Creating Messages:

// Text message
let message = ChatMessage(
    role: .user,
    content: [.text("Hello, how are you?")]
)

// Multimodal message with image
let message = ChatMessage(
    role: .user,
    content: [
        .text("Describe this image:"),
        .image(jpegData)  // Data containing JPEG bytes
    ]
)

// Audio message
let message = ChatMessage(
    role: .user,
    content: [
        .text("Transcribe this audio:"),
        .audio(wavData)  // Data containing WAV bytes (16kHz mono)
    ]
)

`ChatMessageContent`

Content types supported in messages:

enum ChatMessageContent {
    case text(String)
    case image(Data)  // JPEG encoded
    case audio(Data)  // WAV encoded (16kHz, mono, PCM)
}

Audio Requirements:

Format: WAV only
Sample Rate: 16 kHz (mono channel required)
Encoding: PCM (Float32, Int16, Int24, or Int32)
Channels: Mono (1 channel) - stereo will be rejected

Creating Audio Content:

// From WAV file
let wavData = try Data(contentsOf: audioFileURL)
let audioContent = ChatMessageContent.audio(wavData)

// From float samples
let samples: [Float] = [0.1, 0.2, 0.15, ...]  // Normalized to -1.0 to 1.0
let audioContent = ChatMessageContent.fromFloatSamples(samples, sampleRate: 16000)

`MessageResponse`

Streaming response types from generation:

enum MessageResponse {
    case chunk(String)  // Text chunk
    case reasoningChunk(String)  // Reasoning text (thinking models only)
    case audioSample(samples: [Float], sampleRate: Int)  // Audio output (24kHz typically)
    case functionCall([LeapFunctionCall])  // Function call requests
    case complete(MessageCompletion)  // Generation complete
}

MessageCompletion Fields:

struct MessageCompletion {
    let message: ChatMessage  // Complete assistant message
    let finishReason: GenerationFinishReason  // .stop or .exceed_context
    let stats: GenerationStats?  // Token counts and speed
}

struct GenerationStats {
    var promptTokens: UInt64
    var completionTokens: UInt64
    var totalTokens: UInt64
    var tokenPerSecond: Float
}

`GenerationOptions`

Control generation behavior:

struct GenerationOptions {
    var temperature: Float?  // Randomness (0.0 to 2.0)
    var topP: Float?  // Nucleus sampling
    var minP: Float?  // Minimum probability
    var repetitionPenalty: Float?  // Reduce repetition
    var rngSeed: UInt64?  // Seed for deterministic output
    var jsonSchemaConstraint: String?  // JSON schema for structured output
    var functionCallParser: LeapFunctionCallParserProtocol?
    var resetHistory: Bool  // Default true
    var sequenceLength: UInt32?  // Override context length
    var maxOutputTokens: UInt32?  // Limit output length
    var enableThinking: Bool  // Surface <think> blocks
    var cacheControl: CacheControl?
}

Example:

var options = GenerationOptions(
    temperature: 0.7,
    maxOutputTokens: 512,
    enableThinking: false
)

// For structured output
try options.setResponseFormat(type: MyStruct.self)

Generation Patterns

Basic Text Generation

import LeapSDK

@MainActor
final class ChatViewModel: ObservableObject {
    @Published var messages: [ChatMessage] = []
    @Published var isGenerating = false
    @Published var currentResponse = ""

    private var modelRunner: ModelRunner?
    private var conversation: Conversation?
    private var generationTask: Task<Void, Never>?

    func loadModel() async {
        do {
            modelRunner = try await Leap.load(
                model: "LFM2.5-1.2B-Instruct",
                quantization: "Q4_K_M"
            ) { progress, _ in
                print("Loading: \(Int(progress * 100))%")
            }
            conversation = modelRunner?.createConversation(
                systemPrompt: "Explain it to me like I'm 5 years old"
            )
        } catch {
            print("Failed to load model: \(error)")
        }
    }

    func send(_ text: String) {
        guard let conversation else { return }

        // Cancel any ongoing generation
        generationTask?.cancel()

        let userMessage = ChatMessage(role: .user, content: [.text(text)])
        currentResponse = ""
        isGenerating = true

        generationTask = Task {
            do {
                for try await response in conversation.generateResponse(
                    message: userMessage,
                    generationOptions: GenerationOptions(temperature: 0.7)
                ) {
                    await handleResponse(response)
                }
            } catch {
                print("Generation error: \(error)")
            }
            isGenerating = false
        }
    }

    func stopGeneration() {
        generationTask?.cancel()
        generationTask = nil
        isGenerating = false
    }

    @MainActor
    private func handleResponse(_ response: MessageResponse) {
        switch response {
        case .chunk(let text):
            currentResponse += text

        case .reasoningChunk(let reasoning):
            print("Thinking: \(reasoning)")

        case .audioSample(let samples, let sampleRate):
            // Handle audio output (typically 24kHz)
            playAudio(samples: samples, sampleRate: sampleRate)

        case .functionCall(let calls):
            // Handle function calls
            handleFunctionCalls(calls)

        case .complete(let completion):
            if let stats = completion.stats {
                print("Generated \(stats.completionTokens) tokens at \(stats.tokenPerSecond) tok/s")
            }
            // Final message is already in conversation.history
            messages = conversation?.history ?? []
            currentResponse = ""
        }
    }
}

Multimodal Input (Vision)

func sendImageMessage(image: UIImage, prompt: String) {
    guard let jpegData = image.jpegData(compressionQuality: 0.8) else { return }

    let message = ChatMessage(
        role: .user,
        content: [
            .text(prompt),
            .image(jpegData)
        ]
    )

    Task {
        for try await response in conversation.generateResponse(message: message) {
            await handleResponse(response)
        }
    }
}

Audio Input

import AVFoundation

func transcribeAudio(audioFileURL: URL) async {
    // Load WAV file (must be 16kHz, mono, PCM)
    guard let wavData = try? Data(contentsOf: audioFileURL) else { return }

    let message = ChatMessage(
        role: .user,
        content: [
            .text("Transcribe this audio:"),
            .audio(wavData)
        ]
    )

    Task {
        for try await response in conversation.generateResponse(message: message) {
            await handleResponse(response)
        }
    }
}

// Recording audio with AVAudioRecorder
class AudioRecorder {
    private var audioRecorder: AVAudioRecorder?

    func startRecording(to url: URL) throws {
        let settings: [String: Any] = [
            AVFormatIDKey: Int(kAudioFormatLinearPCM),
            AVSampleRateKey: 16000.0,  // 16 kHz required
            AVNumberOfChannelsKey: 1,  // Mono required
            AVEncoderBitDepthKey: 16,  // 16-bit PCM
            AVEncoderAudioQualityKey: AVAudioQuality.high.rawValue
        ]

        audioRecorder = try AVAudioRecorder(url: url, settings: settings)
        audioRecorder?.record()
    }

    func stopRecording() -> URL? {
        audioRecorder?.stop()
        return audioRecorder?.url
    }
}

Function Calling

// Define function
let weatherFunction = LeapFunction(
    name: "get_weather",
    description: "Get the current weather for a location",
    parameters: [
        LeapFunctionParameter(
            name: "location",
            description: "City name",
            type: .string,
            required: true
        ),
        LeapFunctionParameter(
            name: "unit",
            description: "Temperature unit",
            type: .string,
            required: false,
            enumValues: ["celsius", "fahrenheit"]
        )
    ]
)

// Register with conversation
conversation.registerFunction(weatherFunction)

// Handle function calls in response
func handleResponse(_ response: MessageResponse) {
    switch response {
    case .functionCall(let calls):
        for call in calls {
            if call.name == "get_weather" {
                let location = call.arguments["location"] as? String ?? "Unknown"
                let result = getWeather(location: location)

                // Add tool result back to conversation
                let toolMessage = ChatMessage(
                    role: .tool,
                    content: [.text(result)]
                )

                // Create new conversation with updated history
                let updatedHistory = conversation.history + [toolMessage]
                conversation = modelRunner.createConversationFromHistory(
                    history: updatedHistory
                )
            }
        }
    default:
        break
    }
}

Structured Output

Use @Generatable macro for type-safe JSON output:

import LeapSDK

@Generatable
struct Recipe {
    let name: String
    let ingredients: [String]
    let steps: [String]
    let cookingTime: Int
}

// Configure generation
var options = GenerationOptions()
try options.setResponseFormat(type: Recipe.self)

// Generate
for try await response in conversation.generateResponse(
    message: ChatMessage(role: .user, content: [.text("Give me a pasta recipe")]),
    generationOptions: options
) {
    if case .complete(let completion) = response {
        // Parse JSON response into Recipe struct
        if case .text(let json) = completion.message.content.first {
            let recipe = try JSONDecoder().decode(Recipe.self, from: json.data(using: .utf8)!)
            print("Recipe: \(recipe.name)")
        }
    }
}

Conversation Persistence

Save and restore conversation history:

import Foundation

// Save conversation
func saveConversation() throws {
    let jsonArray = try conversation.exportToJSON()
    let data = try JSONSerialization.data(withJSONObject: jsonArray)
    try data.write(to: conversationFileURL)
}

// Restore conversation
func restoreConversation() throws {
    let data = try Data(contentsOf: conversationFileURL)
    let jsonArray = try JSONSerialization.jsonObject(with: data) as! [[String: Any]]

    let history = try jsonArray.map { json in
        try ChatMessage(from: json)
    }

    conversation = modelRunner.createConversationFromHistory(history: history)
}

// Using Codable (alternative)
extension ChatMessage: Codable {
    // Already implemented in SDK
}

func saveWithCodable() throws {
    let encoder = JSONEncoder()
    let data = try encoder.encode(conversation.history)
    try data.write(to: conversationFileURL)
}

func restoreWithCodable() throws {
    let data = try Data(contentsOf: conversationFileURL)
    let decoder = JSONDecoder()
    let history = try decoder.decode([ChatMessage].self, from: data)
    conversation = modelRunner.createConversationFromHistory(history: history)
}

Model Download Management

Query download status and manage cached models:

import LeapModelDownloader

let downloader = ModelDownloader()

// Check download status
let status = downloader.queryStatus("LFM2.5-1.2B-Instruct", quantization: "Q4_K_M")

switch status {
case .notOnLocal:
    print("Model not downloaded")
case .downloadInProgress(let progress):
    print("Downloading: \(Int(progress * 100))%")
case .downloaded:
    print("Model ready")
}

// Get model size before downloading
let sizeInBytes = try await downloader.getModelSize(
    modelName: "LFM2.5-1.2B-Instruct",
    quantization: "Q4_K_M"
)
print("Model size: \(sizeInBytes / 1_000_000) MB")

// Remove downloaded model
try downloader.removeModel("LFM2.5-1.2B-Instruct", quantization: "Q4_K_M")

// Cancel ongoing download
downloader.requestStopDownload(model)

Complete ViewModel Example

import SwiftUI
import LeapSDK
import LeapModelDownloader

@MainActor
final class ChatViewModel: ObservableObject {
    @Published var messages: [ChatMessage] = []
    @Published var currentResponse = ""
    @Published var isGenerating = false
    @Published var isLoadingModel = false
    @Published var downloadProgress: Double = 0.0
    @Published var error: String?

    private var modelRunner: ModelRunner?
    private var conversation: Conversation?
    private var generationTask: Task<Void, Never>?

    func loadModel() async {
        isLoadingModel = true
        downloadProgress = 0.0
        error = nil

        do {
            modelRunner = try await Leap.load(
                model: "LFM2.5-1.2B-Instruct",
                quantization: "Q4_K_M"
            ) { [weak self] progress, speed in
                Task { @MainActor in
                    self?.downloadProgress = progress
                }
            }

            conversation = modelRunner?.createConversation(
                systemPrompt: "Explain it to me like I'm 5 years old"
            )

        } catch {
            self.error = "Failed to load model: \(error.localizedDescription)"
        }

        isLoadingModel = false
    }

    func send(_ text: String) {
        guard let conversation, !text.isEmpty else { return }

        generationTask?.cancel()

        let userMessage = ChatMessage(role: .user, content: [.text(text)])
        messages.append(userMessage)
        currentResponse = ""
        isGenerating = true

        generationTask = Task {
            do {
                for try await response in conversation.generateResponse(
                    message: userMessage,
                    generationOptions: GenerationOptions(
                        temperature: 0.7,
                        maxOutputTokens: 512
                    )
                ) {
                    await handleResponse(response)
                }
            } catch is CancellationError {
                // Generation was cancelled
            } catch {
                self.error = "Generation failed: \(error.localizedDescription)"
            }

            isGenerating = false
        }
    }

    func stopGeneration() {
        generationTask?.cancel()
        generationTask = nil
        isGenerating = false
    }

    @MainActor
    private func handleResponse(_ response: MessageResponse) {
        switch response {
        case .chunk(let text):
            currentResponse += text

        case .reasoningChunk(let reasoning):
            print("Thinking: \(reasoning)")

        case .audioSample(let samples, let sampleRate):
            // Handle audio playback
            break

        case .functionCall(let calls):
            // Handle function calls
            break

        case .complete(let completion):
            if let stats = completion.stats {
                print("Stats: \(stats.totalTokens) tokens, \(stats.tokenPerSecond) tok/s")
            }
            messages = conversation?.history ?? []
            currentResponse = ""
        }
    }

    deinit {
        generationTask?.cancel()
    }
}

Error Handling

enum LeapError: Error {
    case modelLoadingFailure(String, Error?)
    case generationFailure(String, Error?)
    case serializationFailure(String, Error?)
    case invalidInput(String)
}

// Handling errors
do {
    let modelRunner = try await Leap.load(model: "LFM2.5-1.2B-Instruct", quantization: "Q4_K_M")
} catch let error as LeapError {
    switch error {
    case .modelLoadingFailure(let message, _):
        print("Model loading failed: \(message)")
    case .generationFailure(let message, _):
        print("Generation failed: \(message)")
    case .serializationFailure(let message, _):
        print("Serialization failed: \(message)")
    case .invalidInput(let message):
        print("Invalid input: \(message)")
    }
} catch {
    print("Unexpected error: \(error)")
}

Common Imports

// Core SDK
import LeapSDK

// Optional model downloader
import LeapModelDownloader

// SwiftUI integration
import SwiftUI
import Combine

// Audio handling
import AVFoundation

// Image processing
import UIKit  // iOS
import AppKit  // macOS

Troubleshooting

Model Fails to Load

// Check available disk space
let downloader = ModelDownloader()
if let freeSpace = downloader.getAvailableDiskSpace() {
    print("Free space: \(freeSpace / 1_000_000_000) GB")
}

// Check model size
let modelSize = try await downloader.getModelSize(
    modelName: "LFM2.5-1.2B-Instruct",
    quantization: "Q4_K_M"
)
print("Model requires: \(modelSize / 1_000_000) MB")

Generation is Slow

Test on physical device (simulator is much slower)
Use smaller quantization (Q4_K_M instead of Q8_0)
Reduce context size in options
For macOS: increase nGpuLayers for Metal acceleration

Audio Not Working

Verify WAV format (16kHz, mono, PCM)
Check that model supports audio (LFM2.5-Audio models)
For input: ensure mono channel, stereo will fail
For output: audio is typically 24kHz (different from input)

Memory Issues

// Explicitly unload model when done
modelRunner = nil
conversation = nil

// Use smaller models on devices with limited RAM
// LFM2-350M for 3GB devices
// LFM2.5-1.2B for 6GB+ devices

Platform Requirements

iOS: 15.0+
macOS: 12.0+
Xcode: 15.0+ with Swift 5.9+
Device RAM: 3GB minimum (6GB+ recommended for larger models)
Storage: 500MB - 2GB depending on model and quantization

Getting Started

On-Device

GPU Inference

Tools

Core Architecture

Installation

Swift Package Manager (Recommended)

CocoaPods

Loading Models

Method 1: Automatic Download (Recommended)

Method 2: Download Without Loading

Method 3: Custom Manifest URL

Method 4: Local Bundle (Legacy)

Core Classes

`ModelRunner`

`Conversation`

`ChatMessage`

`ChatMessageContent`

`MessageResponse`

`GenerationOptions`

Generation Patterns

Basic Text Generation

Multimodal Input (Vision)

Audio Input

Function Calling

Structured Output

Conversation Persistence

Model Download Management

Complete ViewModel Example

Error Handling

Common Imports

Troubleshooting

Model Fails to Load

Generation is Slow

Audio Not Working

Memory Issues

Platform Requirements

​Core Architecture

​Installation

​Swift Package Manager (Recommended)

​CocoaPods

​Loading Models

​Method 1: Automatic Download (Recommended)

​Method 2: Download Without Loading

​Method 3: Custom Manifest URL

​Method 4: Local Bundle (Legacy)

​Core Classes

​ModelRunner

​Conversation

​ChatMessage

​ChatMessageContent

​MessageResponse

​GenerationOptions

​Generation Patterns

​Basic Text Generation

​Multimodal Input (Vision)

​Audio Input

​Function Calling

​Structured Output

​Conversation Persistence

​Model Download Management

​Complete ViewModel Example

​Error Handling

​Common Imports

​Troubleshooting

​Model Fails to Load

​Generation is Slow

​Audio Not Working

​Memory Issues

​Platform Requirements

Core Architecture

Installation

Swift Package Manager (Recommended)

CocoaPods

Loading Models

Method 1: Automatic Download (Recommended)

Method 2: Download Without Loading

Method 3: Custom Manifest URL

Method 4: Local Bundle (Legacy)

Core Classes

`ModelRunner`

`Conversation`

`ChatMessage`

`ChatMessageContent`

`MessageResponse`

`GenerationOptions`

Generation Patterns

Basic Text Generation

Multimodal Input (Vision)

Audio Input

Function Calling

Structured Output

Conversation Persistence

Model Download Management

Complete ViewModel Example

Error Handling

Common Imports

Troubleshooting

Model Fails to Load

Generation is Slow

Audio Not Working

Memory Issues

Platform Requirements