← All projects

Ask My Portfolio — RAG Chat API

Solo — AI / backend engineer · 2026

The grounded RAG chat that answers questions about me on this very site — a single Cloudflare Worker with SSE token streaming, Supabase pgvector retrieval, and Turnstile + KV abuse guards.

  • TypeScript
  • Cloudflare Workers
  • OpenAI
  • Supabase + pgvector
  • Turnstile
  • Workers KV
  • SSE
View code Live

Problem

The “Ask my portfolio” chat box on this page is this project. A generic chatbot will happily answer questions about me — and make half of them up. I wanted the opposite: a chat that answers only from a curated knowledge base about my background and work, says “I don’t know” when the answer isn’t there, and is safe to leave running on a public URL without waking up to a surprise API bill.

How it works

A POST /api/chat request runs through a short, deliberate pipeline:

  1. Verify the visitor with Cloudflare Turnstile, so bots can’t hammer the endpoint.
  2. Throttle per IP and check a daily token budget — both backed by Workers KV. When the day’s budget is spent, the API declines gracefully instead of running up a bill.
  3. Embed the question and retrieve the most relevant chunks from a Supabase pgvector store.
  4. Assemble a tightly scoped prompt — a system message that only answers questions about me and refuses everything else — plus the retrieved context and a little chat history.
  5. Stream the answer back token by token over SSE, then the list of sources it used.

Role & decisions

  • One Worker, no framework. A single route doesn’t need one. The whole service runs at the edge — routing, auth, retrieval, and streaming in one deployable.
  • Grounded or silent. Answers are constrained to hand-written markdown in knowledge/. If something isn’t there, the assistant is told to say so and hand out my email rather than invent an answer.
  • Abuse guards as first-class code, not an afterthought. Turnstile, a per-IP rate limit, and a daily spend cap are the boring-but-necessary parts that separate a demo from something you can safely make public.
  • Prompt-injection resistance. A deterministic eval suite checks that the assembled prompt keeps the refusal guard and holds injected instructions below the system message — no live API calls, so it runs in CI.
  • Observability. Conversations are logged to Supabase (country, city, and a salted, truncated IP hash — never raw IPs) and viewable through a key-gated admin vault.

Outcome

It’s live, and you’re looking at it — this page’s chat is served by the Worker. Embeddings use text-embedding-3-small at 1536 dimensions; generation is streamed. It’s a small, honest example of production RAG: grounded answers, token streaming, and the guardrails that make a public AI endpoint boring to operate.