README.md

DeepSeekSwift

🚨 Due to current server resource constraints, DeepSeek temporarily suspended API service recharges to prevent any potential impact on users operations. Existing balances can still be used for calls.

Overview

DeepSeek Swift SDK is a lightweight and efficient Swift-based client for interacting with the DeepSeek API. It provides support for chat message completion, streaming, error handling, and configurating DeepSeek LLM with advanced parameters.

Features

• Supports chat completion requests
• Supports fill in the middle completion requests
• Handles error responses with detailed error descriptions and recovery suggestions.
• streaming responses both for chat completion and as well fill in the middle responses
• Built-in support for different models and advanced parameters
• User balance fetchin and available LLM models fetching
• Uses Swift concurrency (async/await) for network calls

Installation

To integrate DeepSwiftSeek into your project, you can use Swift Package Manager (SPM):

let package = Package(
    dependencies: [
        .package(url: "https://github.com/tornikegomareli/DeepSwiftSeek.git", exact: "0.0.2")
    ]
)

Or add it via Xcode:

1. Open your project in Xcode.
2. Navigate to File > Swift Packages > Add Package Dependency.
3. Enter the repository URL.
4. Choose the latest version and click Next.

Usage

1. Initialize the Client

import DeepSwiftSeek

let configuration = Configuration(apiKey: "YOUR_API_KEY")
let deepSeekClient = DeepSeekClient(configuration: configuration)

2. Sending a Chat Completion Request

Task {
    do {
        let response = try await deepSeekClient.chatCompletions(
            messages: {
                ChatMessageRequest(role: .user, content: "Tell me a joke.", name: "User")
            },
            model: .deepSeekChat,
            parameters: .creative
        )
        print(response.choices.first?.message.content ?? "No response")
    } catch {
        print("Error: \(error.localizedDescription)")
    }
}

3. Streaming Chat Completions

Task {
    do {
        let stream = try await deepSeekClient.chatCompletionStream(
            messages: {
                ChatMessageRequest(role: .user, content: "Write a poem.", name: "User")
            },
            model: .deepSeekChat,
            parameters: .streaming
        )
        for try await chunk in stream {
            print(chunk) // Prints streamed responses
        }
    } catch {
        print("Streaming error: \(error.localizedDescription)")
    }
}

4. Streaming FIM Completion

Task {
    do {
        let stream = try await deepSeekClient.fimCompletionStream(
            messages: {
                [
                    ChatMessageRequest(
                      role: .user,
                      content: "function greet() {\n  /* FIM_START */\n  /* FIM_END */\n  return 'Hello world';\n}",
                      name: "User"
                    )
                ]
            },
            model: .deepSeekReasoner,
            parameters: .streaming
        )
        
        for try await chunk in stream {
            // Each chunk is a streamed part of the fill-in-the-middle response.
            print("FIM Stream Chunk:\n\(chunk)")
        }
    } catch {
        print("FIM Streaming Error: \(error.localizedDescription)")
    }
}

5. Sending FIM Completion Request

Task {
    do {
        let response = try await deepSeekClient.fimCompletions(
            messages: {
                [
                    ChatMessageRequest(
                      role: .user,
                      content: "function greet() {\n  // FIM_START\n  // FIM_END\n  return 'Hello world';\n}",
                      name: "User"
                    )
                ]
            },
            model: .deepSeekReasoner,
            parameters: .creative
        )
        if let content = response.choices.first?.message.content {
            print("FIM Completion:\n\(content)")
        }
    } catch {
        print("FIM Error: \(error.localizedDescription)")
    }
}

6. Getting List of Models

Task {
    do {
        let response = try await deepSeekClient.listModels()
    } catch {
        print("ListModels Error: \(error.localizedDescription)")
    }
}

7. Getting Balance of the user

Task {
    do {
        let response = try await deepSeekClient.fetchUserBalance()
    } catch {
        print("UserBalance Error: \(error.localizedDescription)")
    }
}

8. Handling Errors

The SDK provides detailed error handling:

catch let error as DeepSeekError {
    print("DeepSeek API Error: \(error.localizedDescription)")
    print("Recovery Suggestion: \(error.recoverySuggestion ?? "None")")
} catch {
    print("Unexpected error: \(error)")
}

Models

DeepSeek SDK supports multiple models:

public enum DeepSeekModel: String {
    case deepSeekChat = "deepseek-chat"
    case deepSeekReasoner = "deepseek-reasoner"
}

Available Parameters

You can configure chat completion parameters:

let parameters = ChatParameters(
    frequencyPenalty: 0.5,
    maxTokens: 512,
    presencePenalty: 0.5,
    temperature: 0.7,
    topP: 0.9
)

Predefined Parameter Sets

Mode	Temperature	Max Tokens	Top P
Creative	0.9	2048	0.9
Focused	0.3	2048	0.3
Streaming	0.7	4096	0.9
Code Generation	0.2	2048	0.95
Concise	0.5	256	0.5

Creating Custom Predefined Parameters

If you need specific configurations, you can define your own parameter presets:

extension ChatParameters {
    static let myCustomPreset = ChatParameters(
        frequencyPenalty: 0.4,
        maxTokens: 1024,
        presencePenalty: 0.6,
        temperature: 0.8,
        topP: 0.85
    )
}

Then use it in your requests:

let parameters = ChatParameters.myCustomPreset

This approach allows you to maintain reusable configurations tailored to different needs.

Error Handling

DeepSeek SDK has built-in error handling for various API failures:

Error Type	Description
invalidFormat	Invalid request body format.
authenticationFailed	Incorrect API key.
insufficientBalance	No balance remaining.
rateLimitReached	Too many requests sent.
serverOverloaded	High traffic on server.
encodingError	Failed to encode request body.

TODOs

• Improve documentation with more examples
• SwiftUI full demo based on chat, history and reasoning
• Reasoning model + OpenAI SDK

License

This project is available under the MIT License.

Disclaimer

This SDK is not affiliated with DeepSeek and is an independent implementation to interact with their API.