Capstone Project: PrepMind
AI RAG Assistant

Phase 1 — Firebase Setup (30 min)

Create a Firebase project (free Spark plan). Enable Firestore, Firebase Auth (Google provider), and Firebase Hosting. Install Firebase Genkit.

npm install -g firebase-tools
firebase login
firebase init  # select Hosting, Firestore, Functions
npm install @genkit-ai/firebase @genkit-ai/googleai

Phase 2 — Indexing Pipeline (45 min)

Write a Node.js script that reads documents, chunks them into 512-token pieces, embeds each chunk with Gemini's text-embedding-004, and stores in Firestore with the vector field.

// index-documents.js
import { GoogleGenerativeAI } from "@google/generative-ai";
import { initializeApp } from "firebase-admin/app";
import { getFirestore, FieldValue } from "firebase-admin/firestore";

const genai = new GoogleGenerativeAI(process.env.GOOGLE_API_KEY);
const db = getFirestore();

async function indexDocument(text, source) {
  const chunks = chunkText(text, 512, 50); // 512 tokens, 50 overlap
  for (const chunk of chunks) {
    const result = await genai.getGenerativeModel({ model: "text-embedding-004" })
      .embedContent(chunk);
    await db.collection("knowledge_base").add({
      text: chunk,
      source: source,
      embedding: FieldValue.vector(result.embedding.values),
      indexed_at: new Date(),
    });
  }
  console.log(`Indexed ${chunks.length} chunks from ${source}`);
}

Phase 3 — Genkit RAG Flow (45 min)

Build the Genkit flow that handles query embedding, Firestore vector search, context assembly, and Gemini generation with streaming.

// rag-flow.ts  (Firebase Genkit)
import { defineFlow, run } from "@genkit-ai/flow";
import { gemini15Flash } from "@genkit-ai/googleai";
import { textEmbedding004 } from "@genkit-ai/googleai";

export const prepMindFlow = defineFlow(
  { name: "prepMindFlow", inputSchema: z.string(), outputSchema: z.string() },
  async (userQuery) => {
    // 1. Embed the query
    const queryEmbedding = await run("embed-query", () =>
      ai.embed({ embedder: textEmbedding004, content: userQuery })
    );

    // 2. Firestore vector search
    const snapshot = await db.collection("knowledge_base")
      .findNearest("embedding", queryEmbedding, {
        limit: 5,
        distanceMeasure: "COSINE",
      })
      .get();

    // 3. Assemble context
    const context = snapshot.docs
      .map(d => `[Source: ${d.data().source}]\n${d.data().text}`)
      .join("\n\n---\n\n");

    // 4. Generate answer
    const { text } = await ai.generate({
      model: gemini15Flash,
      system: `Answer using ONLY the context below. Cite sources.
If the answer is not in the context, say "I don't have information on that."
Context:\n${context}`,
      prompt: userQuery,
    });

    return text;
  }
);

Phase 4 — Deploy & Present (30 min)

Deploy to Firebase Hosting. Test with 10 real questions. Add a feedback button (thumbs up/down stored in Firestore) so you can show the evaluation loop in your interview.

firebase deploy --only hosting,functions
# Your live URL: https://your-project.web.app

What to say in interviews: "I built PrepMind — a RAG chatbot deployed on Firebase that uses Firestore's native vector search and Gemini API. I can walk you through the indexing pipeline, the query flow, and the evaluation loop I added."