How I built a RAG chatbot for my portfolio

Visitors to a portfolio often ask the same questions: what have you built, what can you do, how do they reach you. A chatbot can answer those — but only if it is grounded in real context.

The idea

I stored short chunks about my projects, skills, and experience in Supabase. When someone asks a question, the app embeds it, finds the closest chunks, and passes them to an LLM as context.

That is classic RAG: Retrieve, Augment, Generate.

The stack

• Embeddings: Hugging Face all-MiniLM-L6-v2
• Vector store: Supabase with a match_documents RPC
• Generation: Groq (llama-3.3-70b-versatile)

Contact info stays in the system prompt so email and location are never lost when retrieval misses.

What I learned

Retrieval quality matters more than model size. A clear knowledge file and a sensible similarity threshold did more for answer quality than swapping models.