Vertex AI + Vercel (Use Gemini AI on Next.js App Router)

.env.local

// Base64 encoded string used for "Application Default Credentials".
// Alternative to GOOGLE_APPLICATION_CREDENTIALS env which accepts JSON file path.
//
// Use following command to generate base64 encoded string for given json file.
// $ openssl base64 < ~/Downloads/gcloud-service-key.json | tr -d '\n'
//
GOOGLE_SERVICE_KEY="<base64_encoded_json_file>"

route.ts

// /app/api/ai/route.ts
import { VertexAI } from "@google-cloud/vertexai";
import { NextResponse } from "next/server";

const credential = JSON.parse(
  Buffer.from(process.env.GOOGLE_SERVICE_KEY ?? "", "base64").toString()
);

const vertex_ai = new VertexAI({
  project: "glcoud-project-name",
  location: "us-central1",
  googleAuthOptions: {
    credentials: {
      client_email: credential.client_email,
      client_id: credential.client_id,
      private_key: credential.private_key,
    },
  },
});
const model = "gemini-1.5-pro-preview-0409";

const generativeModel = vertex_ai.preview.getGenerativeModel({ model });

async function generateContent(promptInput: string) {
  const result = await generativeModel.generateContent({
    contents: [
      {
        role: "user",
        parts: [
          {
            text: `your prompt here: ${promptInput}`,
          },
        ],
      },
    ],
  });
  return result.response.candidates?.[0]?.content.parts[0]?.text?.trim();
}

export async function POST(req: Request) {
  const { promptInput } = await req.json();
  // ... sanitize promptInput ...
  const text = await generateContent(personName);
  // ... sanitize output ...

  return NextResponse.json({ text });
}