LLM Kit DeepSeek

DeepSeek provider for LLM Kit - Complete integration with DeepSeek's chat and reasoning models.

Note: This provider uses the standardized builder pattern. See the Quick Start section for the recommended usage.

Features

• Text Generation: Generate text using DeepSeek models (deepseek-chat, deepseek-reasoner)
• Streaming: Stream responses in real-time for immediate feedback
• Tool Calling: Support for function/tool calling with custom tools
• Reasoning Models: Advanced reasoning capabilities with deepseek-reasoner (R1)
• Prompt Caching: Automatic prompt caching with cache hit/miss token tracking
• Cache Metadata: Track prompt cache efficiency for optimization

Installation

Add this to your Cargo.toml:

[dependencies]
llm-kit-deepseek = "0.1"
llm-kit-core = "0.1"
llm-kit-provider = "0.1"
tokio = { version = "1", features = ["full"] }

Quick Start

Using the Client Builder (Recommended)

use llm_kit_deepseek::DeepSeekClient;
use llm_kit_core::{GenerateText, Prompt};

#[tokio::main]
async fn main() -> Result<(), Box<dyn std::error::Error>> {
    // Create provider using the client builder
    let provider = DeepSeekClient::new()
        .api_key("your-api-key")  // Or use DEEPSEEK_API_KEY env var
        .build();
    
    // Create a language model
    let model = provider.chat_model("deepseek-chat");
    
    // Generate text
    let result = GenerateText::new(std::sync::Arc::new(model), Prompt::text("Hello, DeepSeek!"))
        .temperature(0.7)
        .max_output_tokens(100)
        .execute()
        .await?;
    
    println!("{}", result.text);
    Ok(())
}

Using Settings Directly (Alternative)

use llm_kit_deepseek::{DeepSeekProvider, DeepSeekProviderSettings};
use llm_kit_core::{GenerateText, Prompt};

#[tokio::main]
async fn main() -> Result<(), Box<dyn std::error::Error>> {
    // Create provider with settings
    let provider = DeepSeekProvider::new(DeepSeekProviderSettings::default());
    
    let model = provider.chat_model("deepseek-chat");
    
    let result = GenerateText::new(std::sync::Arc::new(model), Prompt::text("Hello, DeepSeek!"))
        .execute()
        .await?;
    
    println!("{}", result.text);
    Ok(())
}

Configuration

Environment Variables

Set your DeepSeek API key as an environment variable:

export DEEPSEEK_API_KEY=your-api-key
export DEEPSEEK_BASE_URL=https://api.deepseek.com/v1  # Optional

Using the Client Builder

use llm_kit_deepseek::DeepSeekClient;

let provider = DeepSeekClient::new()
    .api_key("your-api-key")
    .base_url("https://api.deepseek.com/v1")
    .header("Custom-Header", "value")
    .build();

Using Settings Directly

use llm_kit_deepseek::{DeepSeekProvider, DeepSeekProviderSettings};

let settings = DeepSeekProviderSettings::new()
    .with_api_key("your-api-key")
    .with_base_url("https://api.deepseek.com/v1")
    .add_header("Custom-Header", "value");

let provider = DeepSeekProvider::new(settings);

Loading from Environment

use llm_kit_deepseek::DeepSeekClient;

// Reads from DEEPSEEK_API_KEY environment variable
let provider = DeepSeekClient::new()
    .load_api_key_from_env()
    .build();

Builder Methods

The DeepSeekClient builder supports:

• .api_key(key) - Set the API key
• .base_url(url) - Set custom base URL
• .header(key, value) - Add a single custom header
• .headers(map) - Add multiple custom headers
• .load_api_key_from_env() - Load API key from DEEPSEEK_API_KEY environment variable
• .build() - Build the provider

Reasoning Models

DeepSeek's reasoner model (R1) provides advanced reasoning capabilities for complex problem-solving:

use llm_kit_deepseek::DeepSeekClient;
use llm_kit_core::{GenerateText, Prompt};

let provider = DeepSeekClient::new()
    .load_api_key_from_env()
    .build();

let model = provider.chat_model("deepseek-reasoner");

let result = GenerateText::new(std::sync::Arc::new(model), 
    Prompt::text("Solve this complex logic puzzle: ..."))
    .execute()
    .await?;

// Access reasoning and answer separately
for output in result.experimental_output.iter() {
    if let llm_kit_provider::language_model::Output::Reasoning(reasoning) = output {
        println!("Reasoning: {}", reasoning.text);
    }
}

println!("Answer: {}", result.text);

Streaming

Stream responses for real-time output:

use llm_kit_deepseek::DeepSeekClient;
use llm_kit_core::{StreamText, Prompt};
use futures_util::StreamExt;

let provider = DeepSeekClient::new()
    .load_api_key_from_env()
    .build();

let model = provider.chat_model("deepseek-chat");

let result = StreamText::new(std::sync::Arc::new(model), 
    Prompt::text("Write a story"))
    .temperature(0.8)
    .execute()
    .await?;

let mut text_stream = result.text_stream();
while let Some(text_delta) = text_stream.next().await {
    print!("{}", text_delta);
}

Supported Models

All DeepSeek models are supported, including:

• deepseek-chat - Main chat model for general tasks and conversations
• deepseek-reasoner - Advanced reasoning model (R1) for complex problem-solving and logic tasks

For a complete list of available models, see the DeepSeek documentation.

Provider-Specific Features

Prompt Cache Statistics

DeepSeek provides detailed information about prompt cache hits and misses to help optimize performance:

use llm_kit_deepseek::DeepSeekClient;
use llm_kit_core::{GenerateText, Prompt};

let provider = DeepSeekClient::new().build();
let model = provider.chat_model("deepseek-chat");

let result = GenerateText::new(std::sync::Arc::new(model), Prompt::text("Hello!"))
    .execute()
    .await?;

// Access cache metadata
if let Some(metadata) = result.provider_metadata {
    if let Some(deepseek) = metadata.get("deepseek") {
        println!("Cache hit tokens: {:?}", 
            deepseek.get("promptCacheHitTokens"));
        println!("Cache miss tokens: {:?}",
            deepseek.get("promptCacheMissTokens"));
    }
}

This helps you understand cache efficiency and optimize your prompts for better performance and cost savings.

Examples

See the examples/ directory for complete examples:

• chat.rs - Basic chat completion with DeepSeek
• stream.rs - Streaming responses
• chat_tool_calling.rs - Tool calling with custom tools
• stream_tool_calling.rs - Streaming with tool calls
• reasoning.rs - Using the deepseek-reasoner model for complex reasoning

Run examples with:

cargo run --example chat
cargo run --example stream
cargo run --example reasoning
cargo run --example chat_tool_calling

Make sure to set your DEEPSEEK_API_KEY environment variable first.

Documentation

• API Documentation
• LLM Kit Documentation
• DeepSeek API Reference

License

Licensed under:

• MIT license (LICENSE-MIT)

Contributing

Contributions are welcome! Please see the Contributing Guide for more details.