Firebase ai logic

Firebase AI Logic (Gemini)

Firebase AI Logic lets you call Gemini directly from the client without managing a backend or API key. The SDK exchanges App Check tokens for short-lived Gemini credentials — abuse-resistant by default.

There are two providers under the same SDK:

• Vertex AI in Firebase — billed via your Google Cloud project, enterprise SLAs, available in more regions, supports all Gemini models including 2.5 Pro/Flash.
• Google AI (Gemini Developer API) — billed via Firebase, simpler quota model, good for early-stage apps and prototypes.

Pick one per project. Migrating later requires re-init code; pick deliberately.

Minimum viable example

import { getVertexAI, getGenerativeModel } from "@react-native-firebase/vertexai";

const model = getGenerativeModel(getVertexAI(), { model: "gemini-2.5-flash" });
const result = await model.generateContent("Write a haiku about Firebase");
console.log(result.response.text());

App Check must be initialized first — without it, anyone with your client bundle can spend your tokens. This is non-negotiable for production.

1. Enable the provider

Firebase Console → Build → AI Logic → Get Started → choose Vertex AI or Google AI. The console walks through enabling the underlying API.

2. Enable App Check (mandatory in production)

Without App Check, anyone with your web Firebase config can bill your project. See the firebase-app-check skill — initialize App Check before the AI SDK on every platform.

3. Install

Web

npm install firebase

import { getVertexAI, getGenerativeModel } from "firebase/vertexai";

const vertexAI = getVertexAI(app);
const model = getGenerativeModel(vertexAI, { model: "gemini-2.5-flash" });

Expo / React Native

npx expo install @react-native-firebase/vertexai

import { getVertexAI, getGenerativeModel } from "@react-native-firebase/vertexai";

const vertexAI = getVertexAI();
const model = getGenerativeModel(vertexAI, { model: "gemini-2.5-flash" });

4. Model selection

Model	When to use
gemini-2.5-flash	Default for most apps — fast, cheap, multimodal. Start here.
gemini-2.5-pro	Complex reasoning, long-context analysis, code generation.
gemini-2.5-flash-lite	Highest-volume / lowest-latency tasks. Classification, extraction.

Pin to a specific version in production (gemini-2.5-flash-002) — unversioned aliases roll forward and can change behavior overnight.

5. Single-turn generation

const result = await model.generateContent("Write a haiku about Firebase");
const text = result.response.text();

6. Multimodal (image + text)

import { fileToGenerativePart } from "firebase/vertexai";

const imagePart = await fileToGenerativePart(file); // File from <input>
const result = await model.generateContent([
  "Describe what's in this image",
  imagePart,
]);

On React Native, load the image as base64 and pass:

const imagePart = {
  inlineData: { data: base64String, mimeType: "image/jpeg" },
};

For videos and audio, the same inlineData shape works with the appropriate MIME (video/mp4, audio/mp3). Files > 20 MB must be uploaded via Cloud Storage and referenced by URI.

7. Streaming responses

const result = await model.generateContentStream("Write a 200-word story");
for await (const chunk of result.stream) {
  process.stdout.write(chunk.text());
}
const final = await result.response;

Streaming gives a much better UX — show tokens as they arrive instead of waiting 5+ seconds for a full response.

8. Structured JSON output

To get strictly-typed output instead of parsing free-form text:

const model = getGenerativeModel(vertexAI, {
  model: "gemini-2.5-flash",
  generationConfig: {
    responseMimeType: "application/json",
    responseSchema: {
      type: "object",
      properties: {
        sentiment: { type: "string", enum: ["positive", "neutral", "negative"] },
        confidence: { type: "number" },
        keywords:   { type: "array", items: { type: "string" } },
      },
      required: ["sentiment", "confidence"],
    },
  },
});

const result = await model.generateContent("I love this product!");
const parsed = JSON.parse(result.response.text()) as {
  sentiment: "positive" | "neutral" | "negative";
  confidence: number;
  keywords?: string[];
};

The model is constrained to emit valid JSON matching the schema — no try/catch ladders.

9. Chat sessions (multi-turn)

const chat = model.startChat({
  history: [
    { role: "user", parts: [{ text: "I'm planning a trip to Lima" }] },
    { role: "model", parts: [{ text: "Great choice! When are you going?" }] },
  ],
});

const response = await chat.sendMessage("Next month, 5 days");
const reply = response.response.text();

// History auto-accumulates
const fullHistory = await chat.getHistory();

Persist getHistory() to Firestore if you need conversations to survive app restarts.

For tool use / function calling — letting Gemini invoke your APIs or query Firestore as part of a conversation — see references/function-calling.md.

10. System instructions

Set persistent context that applies to every turn without consuming chat history:

const model = getGenerativeModel(vertexAI, {
  model: "gemini-2.5-flash",
  systemInstruction: "You are a terse code reviewer. Reply in at most 3 bullet points.",
});

11. Cost and quota

• All requests count against your provider's quota (Vertex AI quota for Vertex; Google AI free/paid tier for Google AI)
• Token-priced: input + output tokens. Multimodal inputs cost more per "token" because images/audio expand significantly.
• Cache long static prompts using context caching (Vertex AI feature) for repeated system prompts — drops cost by ~75%.
• Monitor in Firebase Console → AI Logic → Usage.

12. Common mistakes

• Shipping without App Check. Your Firebase web config + an AI Logic call = anyone can spend your money. Always enforce App Check before going to prod.
• Pinning to an unversioned model alias. Behavior changes silently. Pin to gemini-2.5-flash-002 or specific dates.
• Calling from a server. Use the Vertex AI Node SDK or firebase-admin directly — Firebase AI Logic is client-only.
• Treating JSON output as guaranteed valid. Even with responseSchema, schema misses are rare but possible. Wrap JSON.parse in try/catch and retry on failure.
• Loading 100MB videos as inlineData. Use Cloud Storage URIs for anything > 20 MB.
• Forgetting await on getHistory(). It's async.