ai-lib 🦀✨

A unified, reliable, high-performance multi-provider AI SDK for Rust

A production-grade, provider-agnostic SDK that provides a unified Rust API for 20+ AI platforms and growing (OpenAI, Groq, Anthropic, Gemini, Mistral, Cohere, Azure OpenAI, Ollama, DeepSeek, Qwen, Baidu ERNIE, Tencent Hunyuan, iFlytek Spark, Kimi, HuggingFace, TogetherAI, xAI Grok, OpenRouter, Replicate, Perplexity, AI21, ZhipuAI, MiniMax, and more).
Eliminates fragmented authentication flows, streaming formats, error semantics, model naming differences, and inconsistent function calling. Scale from one-liner scripts to production systems without rewriting integration code.

Official Website

🚀 Core Value

ai-lib unifies AI provider complexity into a single, ergonomic Rust interface:

• Universal API: Chat, multimodal, and function calling across all providers
• Multimodal Content: Easy image and audio content creation with Content::from_image_file() and Content::from_audio_file()
• Unified Streaming: Consistent SSE/JSONL parsing with real-time deltas
• Reliability: Built-in retry, timeout, circuit breaker, and error classification
• Flexible Configuration: Environment variables, builder pattern, or explicit overrides
• Production Ready: Connection pooling, proxy support, observability hooks

Result: Focus on your product logic while ai-lib handles provider integration friction.

Import guidance: In application code, prefer use ai_lib::prelude::*; for a minimal set of common items. Library authors may use explicit imports by domain. See the module tree and import patterns guide: docs/MODULE_TREE_AND_IMPORTS.md.

⚙️ Quick Start

Installation

[dependencies]
ai-lib = "0.3.4"
tokio = { version = "1", features = ["full"] }
futures = "0.3"

One-liner Chat

use ai_lib::Provider;

#[tokio::main]
async fn main() -> anyhow::Result<()> {
    let reply = ai_lib::AiClient::quick_chat_text(Provider::Groq, "Hello!").await?;
    println!("Reply: {reply}");
    Ok(())
}

Standard Usage

// Application code can also use the prelude for minimal imports
use ai_lib::prelude::*;

#[tokio::main]
async fn main() -> anyhow::Result<()> {
    let client = AiClient::new(Provider::OpenAI)?;
    let req = ChatCompletionRequest::new(
        client.default_chat_model(),
        vec![Message {
            role: Role::User,
            content: Content::Text("Explain Rust ownership in one sentence.".to_string()),
            function_call: None,
        }]
    );
    let resp = client.chat_completion(req).await?;
    println!("Answer: {}", resp.choices[0].message.content.as_text());
    Ok(())
}

Streaming

use futures::StreamExt;
let mut stream = client.chat_completion_stream(req).await?;
while let Some(chunk) = stream.next().await {
    let c = chunk?;
    if let Some(delta) = c.choices[0].delta.content.clone() {
        print!("{delta}");
    }
}

🧠 Core Concepts

💡 Key Features

Core Capabilities

• Unified Provider Abstraction: Single API across all providers
• Universal Streaming: Consistent SSE/JSONL parsing with real-time deltas
• Multimodal Support: Text, image, and audio content handling
• Function Calling: Consistent tool patterns and OpenAI compatibility
• Batch Processing: Sequential and concurrent processing strategies

Reliability & Production

• Built-in Resilience: Retry with exponential backoff, circuit breakers
• Basic Failover (OSS): AiClient::with_failover([...]) to switch providers on retryable errors
• Error Classification: Distinguish transient vs permanent failures
• Connection Management: Pooling, timeouts, proxy support
• Observability: Pluggable metrics and tracing integration
• Security: No sensitive content logging by default

🌍 Supported Providers

17+ providers and growing - We continuously add new AI platforms to support the evolving ecosystem.

See examples/ for provider-specific usage patterns.

🔑 Configuration

Environment Variables

# API Keys (convention-based)
export OPENAI_API_KEY=...
export GROQ_API_KEY=...
export GEMINI_API_KEY=...
export ANTHROPIC_API_KEY=...
export OPENROUTER_API_KEY=...
export REPLICATE_API_TOKEN=...
export PERPLEXITY_API_KEY=...
export AI21_API_KEY=...
export ZHIPU_API_KEY=...
export MINIMAX_API_KEY=...

# Optional: Custom endpoints
export GROQ_BASE_URL=https://custom.groq.com

# Optional: Proxy and timeouts
export AI_PROXY_URL=http://proxy.internal:8080
export AI_TIMEOUT_SECS=30

# Optional: Connection pooling (enabled by default)
export AI_HTTP_POOL_MAX_IDLE_PER_HOST=32
export AI_HTTP_POOL_IDLE_TIMEOUT_MS=90000

Programmatic Configuration

use ai_lib::{AiClient, Provider, ConnectionOptions};
use std::time::Duration;

let client = AiClient::with_options(
    Provider::Groq,
    ConnectionOptions {
        base_url: Some("https://custom.groq.com".into()),
        proxy: Some("http://proxy.internal:8080".into()),
        api_key: Some("override-key".into()),
        timeout: Some(Duration::from_secs(45)),
        disable_proxy: false,
    }
)?;

Concurrency Control

use ai_lib::{AiClientBuilder, Provider};

let client = AiClientBuilder::new(Provider::Groq)
    .with_max_concurrency(64)
    .for_production()
    .build()?;

🔁 Failover (OSS)

Use with_failover to define an ordered fallback chain when a request fails with a retryable error (network/timeout/rate-limit/5xx).

use ai_lib::{AiClient, Provider};

let client = AiClient::new(Provider::OpenAI)?
    .with_failover(vec![Provider::Anthropic, Provider::Groq]);

When combined with routing features, the model selection is preserved across failover attempts.

🛡️ Reliability & Resilience

📊 Observability & Metrics

Custom Metrics Integration

struct CustomMetrics;
#[async_trait::async_trait]
impl ai_lib::metrics::Metrics for CustomMetrics {
    async fn incr_counter(&self, name: &str, value: u64) { /* ... */ }
    async fn start_timer(&self, name: &str) -> Option<Box<dyn ai_lib::metrics::Timer + Send>> { /* ... */ }
}
let client = AiClient::new_with_metrics(Provider::Groq, Arc::new(CustomMetrics))?;

Usage Tracking

match response.usage_status {
    UsageStatus::Finalized => println!("Accurate token counts: {:?}", response.usage),
    UsageStatus::Estimated => println!("Estimated tokens: {:?}", response.usage),
    UsageStatus::Pending => println!("Usage data not yet available"),
    UsageStatus::Unsupported => println!("Provider doesn't support usage tracking"),
}

Migration: Usage/UsageStatus are defined in ai_lib::types::response and re-exported at the root. Old imports from types::common are deprecated and will be removed before 1.0.

Optional Features

• interceptors: Retry, timeout, circuit breaker pipeline
• unified_sse: Common SSE parser for all providers
• unified_transport: Shared HTTP client factory
• cost_metrics: Basic cost accounting via environment variables
• routing_mvp: Model selection and routing capabilities

🗂️ Examples

📄 License

Dual-licensed under MIT or Apache License 2.0 - choose what works best for your project.

🤝 Contributing

1. Fork & clone repository
2. Create feature branch: git checkout -b feature/your-feature
3. Run tests: cargo test
4. Add examples for new features
5. Follow adapter patterns (prefer config-driven over custom)
6. Open PR with rationale + benchmarks (if performance impact)

We value: clarity, test coverage, minimal surface area, incremental composability.

📚 Citation

@software{ai-lib,
    title = {ai-lib: A Unified AI SDK for Rust},
    author = {Luqiang Wang},
    url = {https://github.com/hiddenpath/ai-lib},
    year = {2025}
}