An AI-powered Formula One assistant built with Retrieval-Augmented Generation (RAG). Ask anything about F1 — race results, driver standings, technical regulations, history — and get answers grounded in real F1 data retrieved from a vector database.

Live demo: f1-gpt-app.bhaskarg.workers.dev

Home — suggestion chips	Chat — RAG context panel

User question
     │
     ▼
Embed question (text-embedding-3-small)
     │
     ▼
Query Cloudflare Vectorize (top-5 nearest chunks)
     │
     ▼
Inject retrieved context into system prompt
     │
     ▼
Stream response via GPT-4o-mini
     │
     ▼
Display in chat UI with RAG metrics badge

1. Scrape & chunk — F1 data is scraped with Puppeteer, split into chunks, and embedded
2. Store — Embeddings are stored in Cloudflare Vectorize (f1-index)
3. Retrieve — At query time, the user's question is embedded and the top-5 most similar chunks are fetched
4. Generate — Retrieved chunks are injected into the system prompt; GPT-4o-mini streams the answer
5. Observe — A RAG metrics badge shows how many vectors were retrieved and their similarity scores

• RAG pipeline — grounded answers from a live F1 vector index, not just LLM hallucinations
• Streaming responses — token-by-token streaming via AI SDK
• RAG metrics UI — expandable badge shows retrieved chunk count and per-chunk similarity scores
• Rate limiting — server-side per-IP limiter (5 req / 60s) with live client-side countdown
• Suggestion chips — one-click starter questions on the empty state
• Markdown rendering — formatted responses with react-markdown

src/
├── app/
│   ├── api/
│   │   ├── chat/route.ts        # Streaming RAG chat endpoint + rate limiter
│   │   └── rag-stats/route.ts   # RAG metrics endpoint
│   ├── page.tsx                 # Chat UI
│   └── globals.css
scripts/
└── loadDb.ts                    # Scrape → embed → upload to Vectorize

# Install dependencies
npm install

# Add your OpenAI key
echo "OPENAI_API_KEY=sk-..." >> .env.local

# Start dev server (no Vectorize — GPT base knowledge only)
npm run dev

# Start with Cloudflare Workers runtime (Vectorize available)
npm run preview

Note: Vectorize is a Cloudflare-only service. Run npm run preview to test the full RAG pipeline locally. npm run dev falls back to GPT's base knowledge.

# Scrape F1 data, generate embeddings, write vectors.ndjson
npm run seed

# Upload vectors to Cloudflare Vectorize
npm run db:upload

npm run deploy

Requires a Cloudflare account with a Vectorize index named f1-index and OPENAI_API_KEY set as a Workers secret:

wrangler secret put OPENAI_API_KEY

Building this project gave me hands-on experience with the full RAG stack:

• Chunking strategy — how chunk size affects retrieval quality
• Embedding models — using text-embedding-3-small for semantic search
• Vector similarity — interpreting cosine similarity scores (≥ 0.8 = strong match)
• Prompt engineering — injecting retrieved context without confusing the LLM
• Streaming with AI SDK v6 — streamText + toUIMessageStreamResponse() + useChat
• Edge deployment — running a full RAG pipeline on Cloudflare Workers (no Node.js runtime)