ai_sdk_openai 3.0.0

OpenAI provider for the AI SDK Dart port.

ai_sdk_openai

OpenAI provider for AI SDK Dart. Supports language models, embeddings, and image generation via the OpenAI API.

Installation

dependencies:
  ai_sdk_dart: ^3.0.0
  ai_sdk_openai: ^3.0.0

Usage

The top-level openai factory reads const String.fromEnvironment('OPENAI_API_KEY'). Use it with fvm dart run --define=OPENAI_API_KEY=sk-... bin/app.dart, or read Platform.environment['OPENAI_API_KEY'] yourself and pass apiKey: to OpenAIProvider in server and CLI apps.

Language model

import 'package:ai_sdk_dart/ai_sdk_dart.dart';
import 'package:ai_sdk_openai/ai_sdk_openai.dart';

final result = await generateText(
  model: openai('gpt-4.1-mini'),
  prompt: 'Say hello!',
);
print(result.text);

Streaming

import 'dart:io';

final result = await streamText(
  model: openai('gpt-4.1'),
  prompt: 'Tell me a joke.',
);
await for (final chunk in result.textStream) {
  stdout.write(chunk);
}

Embeddings

final result = await embed(
  model: openai.embedding('text-embedding-3-small'),
  value: 'Hello, world!',
);
print(result.embedding); // List<double>

Image generation

import 'package:ai_sdk_dart/ai_sdk_dart.dart';
import 'package:ai_sdk_openai/ai_sdk_openai.dart';

final result = await generateImage(
  model: openai.image('dall-e-3'),
  prompt: 'A futuristic city at sunset.',
);

Reasoning models (o1, o3, o4-mini)

import 'package:ai_sdk_dart/ai_sdk_dart.dart';
import 'package:ai_sdk_openai/ai_sdk_openai.dart';

final result = await generateText(
  model: openai('o4-mini'),
  prompt: 'Solve the Tower of Hanoi for 4 disks.',
  providerOptions: {
    'openai': OpenAILanguageModelOptions(
      reasoningEffort: 'high',
      reasoningSummary: 'detailed',
    ).toMap(),
  },
);
print(result.text);

Responses API + hosted tools

openai.responses(modelId) returns a LanguageModelV4 backed by the OpenAI Responses API instead of Chat Completions — same generateText/streamText usage as openai(modelId). It's required for hosted, server-side tools, passed via providerDefinedTools: (a List<LanguageModelV4ProviderDefinedTool>, separate from the tools: map):

final result = await generateText(
  model: openai.responses('gpt-4.1-mini'),
  prompt: 'What happened in the news today?',
  providerDefinedTools: [OpenAIWebSearchTool()],
);

OpenAIWebSearchTool, OpenAIFileSearchTool (vector-store search), OpenAICodeInterpreterTool, and OpenAIMcpTool (a remote MCP server) are all available — see example/example.md for the full set, plus runnable file/batch lifecycle examples.

Native structured output

When using generateObject or Output.object() with generateText, the OpenAI provider automatically uses response_format: json_schema for schema-validated responses:

final result = await generateObject(
  model: openai('gpt-4.1-mini'),
  schema: Schema<Map<String, dynamic>>(
    jsonSchema: const {
      'type': 'object',
      'properties': {
        'city': {'type': 'string'},
        'country': {'type': 'string'},
      },
      'required': ['city', 'country'],
    },
    fromJson: (json) => json,
  ),
  prompt: 'Capital of France?',
);
print(result.object); // {city: Paris, country: France}

Computer tool output

Computer screenshots use a provider-neutral, camelCase result shape at the public LanguageModelV4 boundary. Supply the screenshot under output and include safety acknowledgements under acknowledgedSafetyChecks:

import 'package:ai_sdk_provider/ai_sdk_provider.dart';

const computerToolResult = LanguageModelV4ToolResultPart(
  toolCallId: 'call-computer',
  toolName: 'computer',
  output: ToolResultOutputJson({
    'output': {
      'type': 'computer_screenshot',
      'imageUrl': 'https://example.test/screenshot.png',
      'detail': 'high',
    },
    'acknowledgedSafetyChecks': [
      {
        'id': 'safety-1',
        'code': 'external_side_effect',
        'message': 'Reviewed by the user',
      },
    ],
  }),
);

The OpenAI Responses adapter converts this public shape to the wire shape (image_url, acknowledged_safety_checks, and related fields) when it sends the computer result. Older provisional integrations that put type and image_url directly at the top level should migrate to the wrapper above and use imageUrl/fileId and acknowledgedSafetyChecks at the public boundary.

Custom API key / base URL

final myOpenAi = OpenAIProvider(
  apiKey: 'sk-...',
  baseUrl: 'https://my-proxy.example.com/v1',
);

final result = await generateText(
  model: myOpenAi('gpt-4.1-mini'),
  prompt: 'Hello!',
);

With provider registry

final registry = createProviderRegistry({
  'openai': RegistrableProvider(
    languageModelFactory: openai.call,
    embeddingModelFactory: openai.embedding,
  ),
});

final model = registry.languageModel('openai:gpt-4.1-mini');

License

MIT