yls_agi_sdk_dart

Flutter/Dart SDK for the YLS AGI gateway, bridged from yls-agi-rust-sdk with flutter_rust_bridge.

It provides:

• Unified OpenAI, Gemini, and Claude chat APIs
• Streaming chat responses
• Multimodal inputs with text, image URL, and base64 image parts
• OpenAI gpt-image-2 image generation and reference-image editing
• Gemini image generation and reference-image editing
• High-level provider facades and official-SDK-style namespaces
• Access to the generated low-level FRB bindings when needed

Install

dependencies:
  yls_agi_sdk_dart: ^0.1.0

Initialize

import 'package:flutter/widgets.dart';
import 'package:yls_agi_sdk_dart/yls_agi_sdk_dart.dart';

Future<void> main() async {
  WidgetsFlutterBinding.ensureInitialized();
  await YlsAgi.init();
}

Create Config

final config = YlsConfigFactory.ylsAgi(
  apiKey: 'your-api-key',
  chatgptImageApiKey: 'your-codex-key',
  proxy: YlsConfigFactory.noProxy(),
);

You can also use YlsConfigFactory.gateway(...) if you need custom base URLs or auth modes.

Quick Start

final openai = YlsAgi.openai(config);

final response = await openai
    .request(YlsModels.openai.gpt41)
    .system('You are a concise assistant.')
    .user('用一句话介绍 Rust')
    .generationOptions(
      temperature: 0.2,
      maxTokens: 256,
    )
    .chat();

print(response.contentText);

Official-SDK-Style APIs

responses.create(...)

final openai = YlsAgi.openai(config);

final response = await openai.responses.create(
  input: 'Explain Rust ownership in one sentence.',
  system: 'You are a concise assistant.',
  options: YlsOptionPresets.precise(maxTokens: 128),
);

print(response.contentText);

responses.stream(...)

final gemini = YlsAgi.gemini(config);

await for (final chunk in gemini.responses.stream(
  input: '请流式输出一段简短介绍',
  options: YlsOptionPresets.balanced(),
)) {
  print(chunk.textDelta);
  if (chunk.isDone) break;
}

chat.completions.create(...)

final claude = YlsAgi.claude(config);

final response = await claude.chat.completions.create(
  messages: [
    YlsMessageBuilder.systemText('You are a concise assistant.'),
    YlsMessageBuilder.userText('Summarize Dart isolates in one sentence.'),
  ],
  options: YlsOptionPresets.precise(),
);

chat.completions.stream(...)

await for (final chunk in openai.chat.completions.stream(
  messages: [
    YlsMessageBuilder.userText('Stream a short product description.'),
  ],
)) {
  print(chunk.textDelta);
}

Provider Facades

final openai = YlsAgi.openai(config);
final gemini = YlsAgi.gemini(config);
final claude = YlsAgi.claude(config);

These facades bind the provider up front, so you do not need to pass Provider.openAi, Provider.gemini, or Provider.claude on every call.

Chain-Style Builders

final request = openai
    .request()
    .system('You are a concise assistant.')
    .user('Say hello in Chinese.')
    .generationOptions(
      temperature: 0.2,
      maxTokens: 64,
    )
    .build();

final response = await openai.chat(request);

You can also send directly from the builder:

final response = await openai
    .request()
    .system('You are a concise assistant.')
    .user('Say hello in Chinese.')
    .chat();

Multimodal Input

final imageMessage = YlsMessageBuilder
    .userText('描述这张图片')
    .addImageBase64(
      mimeType: YlsMimeTypes.png,
      dataBase64: imageBase64,
    );

final response = await YlsAgi.gemini(config).chat.completions.create(
  model: YlsModels.gemini.gemini3ProPreview,
  messages: [imageMessage],
);

You can also use responses.create(inputParts: [...]):

final response = await YlsAgi.gemini(config).responses.create(
  inputParts: [
    YlsMessageBuilder.text('描述这张图片'),
    YlsMessageBuilder.imageBase64(
      mimeType: YlsMimeTypes.png,
      dataBase64: imageBase64,
    ),
  ],
);

Gemini Image Generation

final gemini = YlsAgi.gemini(config);

final imageResponse = await gemini.images.generate(
  prompt: 'Create a photorealistic orange cat wearing sunglasses.',
  options: YlsOptionPresets.image(),
);

final firstImage = imageResponse.images.first;
print(firstImage.mimeType);
await firstImage.saveToFile('/tmp/cat.png');

OpenAI gpt-image-2 Image Generation

final openai = YlsAgi.openai(config);

final imageResponse = await openai.images.generate(
  prompt: '一个打磨完整的 2D 塔防炮台，透明背景。',
);

await imageResponse.image.saveToFile('/tmp/turret.png');

By default, the high-level layer uses:

• outer model: YlsDefaultModels.openaiImageOuter
• image model: YlsDefaultModels.openaiImageModel

You can override either of them:

final imageResponse = await openai.images.generate(
  model: YlsModels.openai.gpt54,
  imageModel: YlsModels.openai.gptImage2,
  prompt: 'A polished fantasy potion icon, transparent background.',
);

OpenAI gpt-image-2 Reference Images

final edited = await openai.imageRequest()
    .prompt('把这张草图转成精致的游戏图标')
    .referenceUrl('https://example.com/sketch.png')
    .referenceFileId('file-123')
    .generate();

You can also attach base64 reference images:

final edited = await openai.imageRequest()
    .prompt('把这张草图转成精致的游戏图标')
    .referenceBase64(
      mimeType: YlsMimeTypes.png,
      dataBase64: imageBase64,
    )
    .generate();

Gemini Image Editing

final edited = await gemini.imageRequest()
    .prompt('把它改成赛博朋克风格')
    .referenceImage(
      mimeType: YlsMimeTypes.png,
      dataBase64: imageBase64,
    )
    .generate();

Defaults and Presets

The high-level layer includes:

• YlsDefaultModels.openaiChat
• YlsDefaultModels.openaiImageOuter
• YlsDefaultModels.openaiImageModel
• YlsDefaultModels.geminiChat
• YlsDefaultModels.claudeChat
• YlsDefaultModels.geminiImage
• YlsOptionPresets.precise()
• YlsOptionPresets.balanced()
• YlsOptionPresets.creative()
• YlsOptionPresets.image()

Low-Level API

If you need full control, the generated FRB bindings are also exported, including:

• ClientConfig
• ChatRequest
• ChatMessage
• MessagePart
• YlsAgiClient
• chat(...)
• chatStream(...)
• generateGeminiImage(...)