Crate rai_sdk 

A unified Rust SDK for backend AI workflows across OpenAI, Anthropic, and OpenRouter.

rai-sdk wraps three provider APIs behind one typed client so switching models does not mean rewriting request-building, streaming, or tool-calling code.

Capabilities

• Typed providers and models — construct models with Model::gpt4o_mini, Model::claude_sonnet_46, or Model::openrouter_auto, or pass any provider model ID directly.
• Typestate request builders — RequestBuilder::generate only exists once a prompt and a model are present, so incomplete requests fail to compile rather than at runtime.
• Structured output — derive JsonSchema and call RequestBuilder::generate_structured to validate the response against a generated schema and deserialize it into your own type.
• Tool calling — register typed async tools with Tool; RequestBuilder::generate runs the tool loop, feeding results back until the model produces a final answer.
• Streaming — consume raw provider events, or use RequestBuilder::stream_accumulated to stream internally and return a complete Response.
• Proxyable streams — RequestBuilder::stream_wire_events yields serializable WireStreamEvents so a server can re-emit a generation to its own clients over SSE, and StreamAccumulator reassembles them on the far end. See the wire module.
• Retries — transient rate-limit, timeout, and HTTP failures are retried with configurable exponential backoff and jitter via RetryConfig.
• Multimodal prompts — build prompts from text, image, audio, video, and file ContentBlocks. Provider support varies.
• Local and self-hosted models — point a client at any endpoint speaking the OpenAI Chat Completions format with ClientBuilder::openai_compatible_base_url (or ClientBuilder::ollama) and name models with Model::openai_compatible. See the provider::openai_compatible module, which the openai feature gates.

Quickstart

use rai_sdk::{ClientBuilder, Model};

let client = ClientBuilder::new()
    .from_env()
    .model(Model::gpt4o_mini())
    .build()?;

let response = client
    .request()
    .prompt("Explain Rust ownership in two sentences.")
    .generate()
    .await?;

println!("{}", response.text());

Configuration

ClientBuilder::from_env reads OPENAI_API_KEY, ANTHROPIC_API_KEY, and OPENROUTER_API_KEY, along with optional base-URL, timeout, and retry overrides. Everything can also be set explicitly on ClientBuilder or Config, and explicit values take precedence over the environment.

Cargo features

The openai, anthropic, and openrouter features are all enabled by default and gate the corresponding provider support. Disable the defaults to compile against only the providers you use.

openai additionally gates the OpenAI-compatible provider, which reuses that module’s request builder and stream parser rather than duplicating them. A build that talks only to local models therefore enables openai and nothing else.

Enabling a provider also requires at least one TLS backend:

• rustls-tls (default) needs no system OpenSSL, but builds aws-lc-rs, which requires cmake and a C compiler.
• native-tls uses the platform TLS stack instead, avoiding that build requirement.

Because a TLS backend is part of the default feature set, disabling default features means re-enabling one explicitly:

rai-sdk = { version = "0.1", default-features = false, features = ["anthropic", "native-tls"] }

Cargo features are additive, so dependency feature unification can enable both backends. That configuration is supported and uses rustls; select only native-tls as shown above to avoid compiling aws-lc-rs.

Further reading

The guide covers each capability in task-oriented chapters. Its examples are compile-checked against this crate, so they stay in sync with the API you see here.

Re-exports

pub use client::Client;

pub use client::ClientBuilder;

pub use client::RequestBuilder;

pub use config::Config;

pub use config::EndpointCapabilities;

pub use error::Capability;

pub use error::Error;

pub use error::ProviderKind;

pub use error::Result;

pub use error::ToolArgumentIssue;

pub use generation::GenerationConfig;

pub use message::ContentBlock;

pub use message::ImageSource;

pub use message::Message;

pub use message::Prompt;

pub use message::Response;

pub use message::Role;

pub use message::StreamChunk;

pub use message::StructuredOutput;

pub use message::ToolCall;

pub use message::ToolDefinition;

pub use message::Usage;

pub use model::AnthropicModel;

pub use model::Model;

pub use model::OpenAICompatibleModel;

pub use model::OpenAIModel;

pub use model::OpenRouterModel;

pub use retry::RetryConfig;

pub use tool::Tool;

pub use tool::ToolContext;

pub use wire::StreamAccumulator;

pub use wire::WIRE_PROTOCOL_VERSION;

pub use wire::WireError;

pub use wire::WireErrorKind;

pub use wire::WireStreamEvent;

pub use schemars;

Modules

client

The client and request builders that drive generation.

config

Client-level configuration: credentials, endpoints, timeouts, retries.

error

Error types shared by every provider.

generation

Per-request generation settings.

message

Prompts, messages, and responses.

model

Typed model catalogs for every supported provider.

provider

Provider implementations and the low-level streaming event they emit.

retry

Retry policy for transient provider failures.

tool

Typed tools that the model can call during generation.

wire

Serializable stream events for proxying a generation across a network hop.

Traits

JsonSchema

A type which can be described as a JSON Schema document.

Derive Macros

JsonSchema

Derive macro for JsonSchema trait.