RAG on Paul Graham: search the startup canon

What it does Paul Graham GPT is a monolithic demo app that lets you search and chat with Paul Graham’s essays. It scrapes the full archive, chunks the text, and stores OpenAI-generated embeddings in a Postgres database with the pgvector extension. When you ask a question, it retrieves the most relevant passages via cosine similarity and feeds them into GPT-3.5-turbo to generate answers grounded in the text.

The interesting bit The author admits he crammed nearly everything into the homepage component to keep the code easy to follow. That makes it less a reusable framework and more a straightforward illustration of how embeddings, vector search, and LLM prompting fit together.

Key highlights

• Includes a downloadable CSV dataset with pre-computed embeddings for the full essay corpus.
• Vector search runs on Postgres with the pgvector extension, hosted on Supabase.
• Chat feeds retrieved passages into GPT-3.5-turbo to answer questions against the essay text.
• Provides both a scraping script to refresh the corpus and an embedding script with built-in rate-limit throttling.
• All code and data are open-source.

Caveats

• The app sacrifices composability for simplicity; most logic lives in a single homepage component.

Verdict Developers who want a concrete, easy-to-read example of retrieval-augmented generation should look here; those seeking a modular, production-ready framework should look elsewhere.