flutter_ai_provider_gemini

Native Google Gemini provider for flutter_ai — streaming, function calling, and Google Search grounding → web citations rendered inline.

Family: flutter_ai · core · client · elements · openai · anthropic
Recipes · Migrating from the Vercel AI SDK

_{With grounding enabled, answers stream source citations (AiSources / AiInlineCitation).}

A native Google Gemini LlmProvider for the flutter_ai family. It streams Gemini's streamGenerateContent endpoint and maps each chunk to AiStreamEvents.

Why native (vs. the OpenAI-compatible Gemini endpoint): it speaks Gemini's own protocol, which unlocks Google Search grounding — grounded answers stream their web sources back as SourcePart citations (rendered by AiSources / AiInlineCitation). It also supports function calling and thinking.

Usage

import 'package:flutter_ai_provider_gemini/flutter_ai_provider_gemini.dart';

final provider = GeminiProvider(
  apiKey: const String.fromEnvironment('GEMINI_API_KEY'),
  enableGrounding: true, // attach the Google Search tool → web citations
);

final controller = UseChatController(
  provider: provider,
  options: const AiRequestOptions(model: 'gemini-2.5-flash'),
);

Notes

• Grounding: set enableGrounding: true (or pass a googleSearch tool via AiRequestOptions.extra). Grounded responses emit SourceParts for each web source.
• Function calling: pass ToolDefinitions; the model's functionCalls stream as tool-call events, and tool results map to functionResponse parts (the function name is recovered from the matching call).
• Config: temperature / maxOutputTokens are sent under generationConfig; arbitrary extra request fields go via AiRequestOptions.extra.
• Images: user-message image attachments (FilePart with an image/* media type) are sent as inlineData/fileData. Other document types are not yet sent.
• Retry: transient failures (429/5xx, network) are retried with backoff honoring Retry-After (maxRetries, default 2).

Status

The request/response mapping is unit-tested against recorded SSE chunks. Supply an API key to use it against the live API.

If flutter_ai saves you time, you can buy me a coffee ☕.