fake_firebase_ai 0.1.0

Fake for Firebase AI Logic (firebase_ai). Unit test Gemini calls, streaming, chat and function calling without network or API keys.

fake_firebase_ai

A fake for firebase_ai (Firebase AI Logic) for unit and widget tests. No network, no API key, same answers every run.

FakeFirebaseAI swaps out the HTTP layer, not the SDK. You get a real GenerativeModel, so the SDK's own code still runs in your tests: streaming, ChatSession history, automatic function calling and error mapping all behave as they do in production.

Install

dev_dependencies:
  fake_firebase_ai: ^0.1.0

Usage

import 'package:fake_firebase_ai/fake_firebase_ai.dart';
import 'package:firebase_ai/firebase_ai.dart';
import 'package:flutter_test/flutter_test.dart';

void main() {
  test('replies', () async {
    final ai = FakeFirebaseAI()..reply('Hello!');
    final model = await ai.generativeModel('gemini-2.5-flash');

    final response = await model.generateContent([Content.text('Hi')]);

    expect(response.text, 'Hello!');
    expect(ai.requests.single.text, 'Hi');
  });
}

generativeModel() mocks firebase_core for you. You don't need Firebase.initializeApp or any platform channel setup.

Replies

Replies are queued and used in order, one per generateContent / generateContentStream call.

ai.reply('text');                                   // a plain reply
ai.replyStream(['Hel', 'lo']);                      // one SSE chunk each
ai.replyFunctionCall('getWeather', {'city': 'Seoul'});
ai.replyBlocked();                                  // promptFeedback.blockReason = SAFETY
ai.replyError(429, 'Quota exceeded');               // throws QuotaExceeded
ai.replyError(400, 'Bad request');                  // throws ServerException
ai.enqueue(FakeResponse.raw([{...}]));              // any raw JSON response

Use a handler for dynamic replies. It runs when the queue is empty:

ai.handler = (request) => FakeResponse.text('echo: ${request.text}');

With an empty queue and no handler, the call throws a StateError that names the request.

Function calling

AutoFunctionDeclaration runs inside the real ChatSession, so queue the call and then the final answer:

final model = await ai.generativeModel('gemini-2.5-flash', tools: [
  Tool.functionDeclarations([getWeather]),
]);
ai
  ..replyFunctionCall('getWeather', {'city': 'Seoul'})
  ..reply('It is sunny in Seoul.');

final response = await model.startChat().sendMessage(Content.text('Weather?'));

Inspecting requests

Every request is recorded in ai.requests as a FakeRequest:

Field
model	gemini-2.5-flash
method	generateContent, streamGenerateContent or countTokens
text	every text part in the request, joined
contents	the prompt and chat history as raw JSON
body	the full request body as raw JSON

countTokens

countTokens counts whitespace-separated words by default. Set ai.tokenCount = 42 to return a fixed value. It doesn't use the reply queue.

Agent Platform (Vertex AI) or a custom setup

FakeFirebaseAI is an http.Client. Pass it to any model you build yourself. Call generativeModel() once first, or set up the Firebase mocks yourself:

FirebaseAI.agentPlatform().generativeModel(model: 'gemini-2.5-flash', httpClient: ai);

Not supported yet

Live API (WebSocket), Imagen and TemplateGenerativeModel. Please open an issue if you need one of these.