Crate zai_rs 

ZAI-RS: Zhipu AI Rust SDK

zai-rs is a type-safe Rust SDK for the Zhipu AI (BigModel) API. Strongly-typed clients and models span chat completions, image generation, speech recognition, text embeddings, knowledge-base management, and more.

Capabilities

Capability	Description	Module
Chat completions	Sync / async text, vision, and voice	model
Image generation	Text-to-image	model::gen_image
Video generation	Async text-to-video	model::gen_video_async
Text-to-speech	Audio synthesis	model::text_to_audio
Speech-to-text	Audio transcription	model::audio_to_text
Voice cloning	Voice clone, list, delete	model::voice_clone
Text embeddings	Embeddings, reranking, tokenization	model::text_embedded
Content moderation	Safety analysis	model::moderation
OCR	Handwriting recognition	model::ocr
File management	Upload, list, content, delete	file
Batch processing	Create, list, retrieve, cancel	batches
Knowledge base	CRUD, document upload, retrieval	knowledge
Tool calling	Function calling, web search, file parsing	tool
MCP	Unified search, reader, repository, and vision capabilities	mcp
Agent	Agent creation & management	agent
Tool execution framework	Dynamic registration, execution, caching	toolkits
Real-time	WebSocket audio/video (GLM-Realtime)	realtime
Coding Plan usage	GLM Coding Plan quota / 余量查询	usage

Module Structure

• client — HTTP client, connection pool, retry strategy, error types
• model — Data models, request/response types, model definitions, SSE parsing
• file — File management (upload, list, content, delete)
• batches — Batch processing (create, list, retrieve, cancel)
• knowledge — Knowledge-base management (CRUD, document upload, retrieval)
• tool — Tool implementations (web search, file parsing)
• mcp — Unified MCP capabilities with automatic backend and transport selection (feature mcp)
• agent — Agent API (creation, chat, history)
• toolkits — Tool execution framework (registration, execution, caching, RMCP bridge)
• realtime — Real-time audio/video communication (WebSocket, experimental)
• usage — Coding Plan usage / quota query (GLM Coding Plan 余量查询)

Quick Start

use zai_rs::{client::ZaiClient, model::*};

#[tokio::main]
async fn main() -> Result<(), Box<dyn std::error::Error>> {
    let model = GLM4_5_flash {};
    let client = ZaiClient::from_env()?;
    let request = ChatCompletion::new(model, TextMessage::user("Hello"));
    let _resp = request.send_via(&client).await?;
    Ok(())
}

Configuration

ZaiClient owns credentials, validated endpoint families, connection pooling, timeouts, and retry policy. Clone the client to share the same transport safely across requests.

use zai_rs::client::{ApiFamily, ZaiClient};

let client = ZaiClient::builder("abc123.abcdefghijklmnopqrstuvwxyz")
    .endpoint(
        ApiFamily::CodingPaasV4,
        "https://open.bigmodel.cn/api/coding/paas/v4",
    )
    .build()?;

assert_eq!(
    client.endpoints().base(ApiFamily::CodingPaasV4).as_str(),
    "https://open.bigmodel.cn/api/coding/paas/v4"
);

Feature Flags

Feature	Default	Description
(default)	enabled	Core API functionality
realtime	disabled	Real-time audio/video over WebSocket (GLM-Realtime)
mcp	disabled	Unified high-level MCP capability client
rmcp-kits	disabled	Enable RMCP protocol bridge for MCP tool calling
toolkits	disabled	JSON-Schema validation for the tool-execution framework

Enable in Cargo.toml:

[dependencies]
zai-rs = { version = "0.6", features = ["mcp"] }

Error Handling

All API calls return ZaiResult<T>, unified under the ZaiError enum:

Error variants distinguish HTTP, authentication, account, API, rate-limit, content-policy, file, network, JSON, realtime, and unknown failures. Use ZaiError::category when recovery logic only needs a coarse classification.

Design Principles

• Compile-time type safety — trait bounds and type-state patterns ensure model/message compatibility at compile time
• Zero-cost abstractions — marker traits and type-state patterns impose no runtime overhead
• Consistent API style — request builders carry typed payloads and all network operations are dispatched with send_via(&ZaiClient)

Re-exports

pub use client::ZaiClient;

pub use client::error::*;

Modules

agent

Agent v1 wire contracts

batches

Batch Processing Module

client

HTTP client infrastructure: the shared ZaiClient, validated endpoints, transport policies and error types.

file

File Management Module

knowledge

Knowledge Base Module

mcpmcp

Unified MCP capability client. Unified API for MCP capabilities.

model

Model Module

prelude

Convenience imports for the most common client and chat APIs.

realtimerealtime

WebSocket realtime (GLM-Realtime) client — audio/video over a WebSocket. Gated behind the realtime Cargo feature (off by default).

services

Typed service facades for application, assistant, image, and document tools. Request and response types for service-oriented Z.AI endpoints.

tool

Tool Module

toolkits

Tool definition, execution, caching, and LLM tool-call parsing utilities. Supports both static tool definitions and dynamic registration at runtime.

usage

Coding Plan Usage / Quota Query