Swift GenKit

GenKit is a library that abstracts away all the differences across generative AI platforms. It's sort of like a lightweight LangChain for Swift. The goal is to make native Swift development with generative AI fast, easy and fun!

Swift Package Manager

...
dependencies: [
    .package(url: "https://github.com/nathanborror/swift-gen-kit", branch: "main"),
],
targets: [
    .target(
        name: "YOUR_TARGET",
        dependencies: [
            .product(name: "GenKit", package: "swift-gen-kit"),
        ]
    ),
],
...

Usage

Establish the service and model you want to use:

let service = AnthropicService(configuration: .init(token: "ANTHROPIC_API_TOKEN"))
let model = Model(id: "claude-3-5-sonnet-20240620")

An example chat completion that just generates a single response:

let request = ChatServiceRequest(
    model: model,
    messages: [
        .system(content: "You are a helpful assistant."),
        .user(content: "Hello!"),
    ]
)
let message = try await service.completion(request)
print(message)

An streaming session example that may perform multiple generations in a loop if tools are present:

var request = ChatSessionRequest(service: service, model: model)
request.with(system: "You are a helpful assistant.")
request.with(history: [.user(content: "Hello!")])

let stream = ChatSession.shared.stream(request)
for try await message in stream {
    print(message.content)
}

Demo App

Heat is a good example of how GenKit can be used.

Features

Sessions

Sessions are the highest level of abstraction and the easiest to use. They run in a loop and call out to tools as needed and send tool responses back to the model until it completes its work.

Services

Services are a common interface for working across many platforms. They allow you to seamlessly switch out the underlying platform without changing any code.

• Chat Service
• Embedding Service
• Image Service
• Model Service
• Transcription Service
• Tool Service
• Vision Service
• Speech Service

Supported Providers

Provider packages are swift interfaces that talk directly to model provider REST APIs. You can use these directly but their APIs vary slightly.

• Mistral (chats, embeddings, models)
• Ollama (chats, embeddings, models, vision)
• OpenAI (chats, embeddings, models, images, transcriptions, vision, speech)
• Perplexity (chats, models)
• Anthropic (chats, models)
• ElevenLabs (speech, models)
• Google (chat, models)

Todos

• Consolodate the VisionService and ToolService into the ChatService
• Consolodate ChatServiceRequest and ChatSessionRequest
• Add workflows or chaining