llm_gemini #

Google Gemini backend for LLM interactions in Dart. Part of the dart-llm ecosystem.

Available on pub.flutter-io.cn.

Features #

• Streaming chat responses via the Gemini Interactions API (POST /v1beta/interactions)
• Tool/function calling with automatic multi-turn tool-loop execution
• Thinking mode with thought summaries surfaced separately from content
• Native structured output via response_format (JsonFormat, JsonSchemaFormat)
• Vision input, plus generated images surfaced on chunk.message.images
• Embeddings (single and batch)
• Builder pattern for fluent configuration
• Configurable retry and timeout policies
• extraHeaders for arbitrary headers on every request

Interactions API #

Chat runs on the Interactions API rather than the legacy generateContent endpoint. What this changes for callers:

• The API key travels in the x-goog-api-key header, never as a key= query parameter, so it cannot leak into request logs or proxies.
• streamChat stays stateless: requests send store: false and serialize the whole conversation into the input array instead of chaining previous_interaction_id. Pass backendOptions: {'previous_interaction_id': '…'} to opt into server-side continuation.
• Thinking arrives as thought_summary deltas and populates chunk.message.thinking; ordinary text populates chunk.message.content.
• chunk.providerMetadata carries interaction_id, thought_signatures (keyed by step index), and the usage counters with no first-class slot: total_tokens, total_thought_tokens, total_cached_tokens, total_tool_use_tokens.
• Tool calls carry the server-provided call id. When the model sent a thought signature, it rides inside that id (after ::sig::) so the tool loop echoes it back, which multi-turn function calling requires.

Embeddings still use the embedContent / batchEmbedContents endpoints.

Installation #

dependencies:
  llm_gemini: ^0.6.0

Prerequisites #

You need a Google API key with Gemini API access. Get one from aistudio.google.com.

Important: Never commit your API key to version control. Use environment variables or a .env file.

Usage #

Basic Chat #

import 'package:llm_gemini/llm_gemini.dart';

final repo = GeminiChatRepository(apiKey: 'your-api-key');

final stream = repo.streamChat('gemini-3.5-flash-lite', messages: [
  LLMMessage(role: LLMRole.user, content: 'Hello!'),
]);

await for (final chunk in stream) {
  stdout.write(chunk.message?.content ?? '');
}

System Message #

final stream = repo.streamChat('gemini-3.5-flash-lite', messages: [
  LLMMessage(role: LLMRole.system, content: 'You are a concise assistant.'),
  LLMMessage(role: LLMRole.user, content: 'Explain quantum entanglement.'),
]);

Tool Calling #

class WeatherTool extends LLMTool {
  @override
  String get name => 'get_weather';

  @override
  String get description => 'Get the current weather for a location.';

  @override
  List<LLMToolParam> get parameters => [
    LLMToolParam(
      name: 'location',
      type: 'string',
      description: 'City name',
      isRequired: true,
    ),
  ];

  @override
  Future<dynamic> execute(Map<String, dynamic> args, {dynamic extra}) async {
    return {'temperature': 22, 'condition': 'sunny'};
  }
}

final stream = repo.streamChat(
  'gemini-3.5-flash-lite',
  messages: [LLMMessage(role: LLMRole.user, content: 'What is the weather in Oslo?')],
  tools: [WeatherTool()],
);

Tools run automatically; pass LLMChatOptions(autoExecuteTools: false) to handle chunk.message?.toolCalls yourself. The tool name is reported on chunk.message?.toolCallDeltas from step.start, before any argument fragment.

Argument fragments are passed through verbatim. When they do not decode — an interaction cut short at max_output_tokens (finish reason LLMFinishReason.length) — the call arrives on chunk.message?.invalidToolCalls with its raw arguments and the parse error, and is never executed; the tool loop answers it with a tool error.

Structured Output #

Gemini supports native structured output via the response_format array. Use JsonFormat for any JSON, or JsonSchemaFormat to enforce a specific schema. Pass standard lowercase JSON Schema — unlike generateContent's responseSchema, the Interactions API takes the schema as written.

import 'package:llm_core/llm_core.dart';

// Simple JSON mode — any valid JSON output
final stream = repo.streamChat(
  'gemini-3.5-flash-lite',
  messages: [LLMMessage(role: LLMRole.user, content: 'List three fruits as JSON.')],
  options: const LLMChatOptions(responseFormat: JsonFormat()),
);

// JSON Schema mode — enforces schema structure
const schema = {
  'type': 'object',
  'properties': {
    'name': {'type': 'string'},
    'age': {'type': 'integer'},
  },
  'required': ['name', 'age'],
};

final stream = repo.streamChat(
  'gemini-3.5-flash-lite',
  messages: [LLMMessage(role: LLMRole.user, content: 'Return a person object.')],
  options: const LLMChatOptions(
    responseFormat: JsonSchemaFormat(name: 'Person', schema: schema),
  ),
);

Thinking Mode #

Extended reasoning is supported on compatible Gemini models. think: true sends generation_config.thinking_summaries: "auto", and thought summaries arrive on chunk.message.thinking rather than mixed into chunk.message.content.

final stream = repo.streamChat(
  'gemini-3.5-flash-lite',
  messages: [LLMMessage(role: LLMRole.user, content: 'Solve this step by step: ...')],
  think: true,
  options: const LLMChatOptions(
    reasoningBudget: 8192,
  ),
);

await for (final chunk in stream) {
  if (chunk.message?.thinking != null) stdout.write(chunk.message!.thinking!);
  if (chunk.message?.content != null) stdout.write(chunk.message!.content!);
}

The Interactions API has no raw thinking-token budget. reasoningBudget is mapped onto the discrete thinking_level field with these thresholds:

An explicit LLMChatOptions.reasoningEffort wins over the budget mapping (none/minimal → minimal, low → low, medium → medium, high/xhigh/max → high), and backendOptions: {'thinking_level': 'high'} bypasses both. Thought-token usage is surfaced as LLMUsage.reasoningTokens.

Embeddings #

// Single text
final embeddings = await repo.embed(
  model: 'gemini-embedding-001',
  messages: ['Hello world'],
);
print(embeddings.first.embedding); // List<double>

// Batch (uses batchEmbedContents endpoint)
final batchEmbeddings = await repo.batchEmbed(
  model: 'gemini-embedding-001',
  messages: ['Hello world', 'Goodbye world'],
);

Non-Streaming Response #

final response = await repo.chatResponse('gemini-3.5-flash-lite', messages: [
  LLMMessage(role: LLMRole.user, content: 'Hello!'),
]);

print(response.content);

Using LLMChatOptions #

final options = LLMChatOptions(
  tools: [WeatherTool()],
  toolAttempts: 5,
  // Portable fields are the easy path: `temperature` and `maxOutputTokens` are
  // mapped into generation_config for you.
  temperature: 0.7,
  maxOutputTokens: 2048,
);

final stream = repo.streamChat('gemini-3.5-flash-lite', messages: messages, options: options);

backendOptions keys are snake_case, matching the wire format. A camelCase key such as maxOutputTokens is not recognised as a generation-config field and is written to the top level of the request body instead — where the API either rejects it or ignores it. See Generation Config via backendOptions.

Advanced Configuration #

Builder Pattern #

final repo = GeminiChatRepository.builder()
  .apiKey('your-api-key')
  .baseUrl('https://generativelanguage.googleapis.com')
  .maxToolAttempts(10)
  .retryConfig(RetryConfig(
    maxAttempts: 3,
    initialDelay: Duration(seconds: 1),
    maxDelay: Duration(seconds: 30),
  ))
  .timeoutConfig(TimeoutConfig(
    connectionTimeout: Duration(seconds: 10),
    readTimeout: Duration(minutes: 5),
  ))
  .build();

Retry Configuration #

final repo = GeminiChatRepository(
  apiKey: 'your-api-key',
  retryConfig: RetryConfig(
    maxAttempts: 3,
    initialDelay: Duration(seconds: 1),
    maxDelay: Duration(seconds: 30),
    retryableStatusCodes: [429, 500, 502, 503, 504],
  ),
);

Timeout Configuration #

final repo = GeminiChatRepository(
  apiKey: 'your-api-key',
  timeoutConfig: TimeoutConfig(
    connectionTimeout: Duration(seconds: 10),
    readTimeout: Duration(minutes: 5),
    totalTimeout: Duration(minutes: 10),
  ),
);

Generation Config via backendOptions #

The Interactions API documents four generation_config fields: temperature, max_output_tokens, thinking_level, and thinking_summaries. temperature and maxOutputTokens from LLMChatOptions are mapped automatically. topP, topK, and stopSequences have no documented Interactions equivalent and are not sent.

Any documented field can be set directly, and generation_config is merged in wholesale for anything else a model turns out to accept:

final options = LLMChatOptions(
  backendOptions: {
    'temperature': 0.9,
    'max_output_tokens': 4096,
    'thinking_level': 'low',
    // Escape hatch — merged into generation_config as-is.
    'generation_config': {'top_p': 0.95},
  },
);

Keys outside that set are written to the top level of the request body, so undocumented request fields can be sent without a new release.

Models #

See Gemini Models for available models:

• gemini-3.6-flash — Newest Flash model
• gemini-3.5-flash — Larger Flash model
• gemini-3.5-flash-lite — Low-cost current Gemini chat model used by live tests
• gemini-embedding-001 — Embeddings used by live tests

Notes #

• embed with more than one input delegates to batchEmbed.
• providerMetadata carries interaction_id, status, thought_signatures, total_tokens, total_thought_tokens, total_cached_tokens and total_tool_use_tokens.
• Retries are off unless you pass a RetryConfig.

Example #

A runnable CLI lives in example/.