QAI SDK

A modular, type-safe Rust SDK for AI providers. One unified API across OpenAI, Anthropic Claude, Google Gemini, DeepSeek, xAI Grok, and any OpenAI-compatible endpoint.

Features

Unified API Demo

The playground.html showcase demonstrates the lightning-fast API flexibility. Open it locally to interact with it directly.

Quick Start

Add to your Cargo.toml:

[dependencies]
qai-sdk = "0.1"
tokio = { version = "1", features = ["full"] }

By default, all providers are enabled. To optimize compile times, disable default features and select only the providers you need:

[dependencies]
qai-sdk = { version = "0.1", default-features = false, features = ["openai", "anthropic"] }

Basic Usage

use qai_sdk::prelude::*;

#[tokio::main]
async fn main() -> anyhow::Result<()> {
    // Create a provider
    let provider = create_openai(ProviderSettings {
        api_key: Some("sk-...".to_string()),
        ..Default::default()
    });

    // Get a chat model
    let model = provider.chat("gpt-4o");

    // Generate a response
    let result = model.generate(
        Prompt {
            messages: vec![Message {
                role: Role::User,
                content: vec![Content::Text {
                    text: "Hello, world!".to_string(),
                }],
            }],
        },
        GenerateOptions {
            model_id: "gpt-4o".to_string(),
            max_tokens: Some(100),
            temperature: Some(0.7),
            top_p: None,
            stop_sequences: None,
            tools: None,
        },
    ).await?;

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

Streaming

use qai_sdk::prelude::*;
use futures::StreamExt;

let model = provider.chat("gpt-4o");
let mut stream = model.generate_stream(prompt, options).await?;

while let Some(part) = stream.next().await {
    match part {
        StreamPart::TextDelta { delta } => print!("{delta}"),
        StreamPart::Finish { finish_reason } => println!("\n[{finish_reason}]"),
        _ => {}
    }
}

Switch Providers in One Line

// OpenAI
let provider = create_openai(settings.clone());
// Anthropic
let provider = create_anthropic(settings.clone());
// Google Gemini
let provider = create_google(settings.clone());
// DeepSeek
let provider = create_deepseek(settings.clone());
// xAI Grok
let provider = create_xai(settings.clone());
// Any OpenAI-compatible API
let provider = create_openai_compatible(settings);

Provider Registry — Resolve Models by String

use qai_sdk::core::registry::ProviderRegistry;

let registry = ProviderRegistry::new()
    .register("openai", openai_provider)
    .register("anthropic", anthropic_provider);

let model = registry.language_model("openai:gpt-4o")?;
let result = model.generate(prompt, options).await?;

Structured Output — Force JSON Schema Conformance

use qai_sdk::core::structured::*;

let result = generate_object(
    &model,
    "Generate a user profile for John Doe, age 30",
    ObjectGenerateOptions {
        model_id: "gpt-4o".to_string(),
        schema: serde_json::json!({
            "type": "object",
            "properties": {
                "name": { "type": "string" },
                "age": { "type": "integer" }
            },
            "required": ["name", "age"]
        }),
        mode: OutputMode::Json,
        ..Default::default()
    },
).await?;
println!("{}", result.object); // {"name": "John Doe", "age": 30}

Middleware — Composable Model Wrappers

use qai_sdk::core::middleware::*;

let wrapped = wrap_language_model(
    model,
    vec![Box::new(DefaultSettingsMiddleware {
        temperature: Some(0.7),
        max_tokens: Some(2048),
        top_p: None,
    })],
);
// Every call now uses temperature=0.7 if not explicitly set

Universal Agent — Multi-Step Tool Loop

use qai_sdk::core::agent::Agent;

let agent = Agent::builder()
    .model(model)
    .tools(vec![weather_tool, search_tool])
    .tool_handler(|name, args| async move {
        match name.as_str() {
            "get_weather" => Ok(serde_json::json!({"temp": "22°C"})),
            _ => Err(anyhow::anyhow!("Unknown tool")),
        }
    })
    .max_steps(10)
    .system("You are a helpful assistant.")
    .build()
    .expect("agent build");

let result = agent.run("What's the weather?").await?;
println!("{}  ({} steps)", result.text, result.total_steps);

Documentation

Dive deep into specific provider features and initialization parameters in our comprehensive module docs:

• Core Interoperability qai_sdk::core
• OpenAI Provider qai_sdk::openai
• Anthropic Provider qai_sdk::anthropic
• Google Gemini Provider qai_sdk::google
• DeepSeek Provider qai_sdk::deepseek
• xAI Grok Provider qai_sdk::xai
• OpenAI Compatible Provider qai_sdk::openai_compatible
• Model Context Protocol qai_sdk::mcp
• Structured Output qai_sdk::core::structured
• Provider Registry qai_sdk::core::registry
• Middleware qai_sdk::core::middleware
• Universal Agent qai_sdk::core::agent

Architecture

qai-sdk is a single, monolithic crate designed with zero-cost abstractions. Providers are organically separated via modular architecture and gated by Cargo features, keeping compile times fast when you only need specific integrations:

qai-sdk
├── core
│   ├── traits          — LanguageModel, EmbeddingModel, ImageModel, SpeechModel, TranscriptionModel
│   ├── structured      — generate_object() / stream_object() with JSON Schema validation
│   ├── registry        — ProviderRegistry for "provider:model" string resolution
│   ├── middleware      — Composable LanguageModelMiddleware (DefaultSettings, ExtractReasoning)
│   └── agent           — Universal Agent with builder pattern & max_steps tool loop
├── openai              — OpenAI API (GPT, DALL-E, Whisper, TTS, Responses)
├── anthropic           — Anthropic API (Claude)
├── google              — Google API (Gemini)
├── deepseek            — DeepSeek API (via OpenAI-compatible pipeline)
├── xai                 — xAI API (Grok, via OpenAI-compatible pipeline)
├── openai_compatible   — Any OpenAI-compatible endpoint (Ollama, LM Studio)
└── mcp                 — Model Context Protocol (JSON-RPC, Stdio/SSE, resources, prompts)

Examples

See the examples/ directory for 17 comprehensive examples covering:

• Basic chat, streaming, and multimodal conversations
• Tool calling / function calling
• Embeddings, image generation, speech, and transcription
• OpenAI Responses API
• Error handling patterns
• Provider factory pattern
• OpenAI-compatible endpoints (Ollama, LM Studio, etc.)

Run an example:

cp .env.example .env
# Fill in your API keys
cargo run --example chat_basic

Environment Variables

Contributing

See CONTRIBUTING.md for guidelines.

License

Licensed under either of:

• MIT License
• Apache License, Version 2.0

at your option.

Author

Keyvan Arasteh — @keyvanarasteh