From prompt to production RAG

Use Perplexity or Claude to spot-check 10 questions your corpus should answer. If they cannot answer them with the corpus pasted in, no RAG will save you.

Cursor scaffold: a script that chunks markdown by heading, embeds with text-embedding-3-large, stores in Postgres pgvector.

Cursor: a Netlify function that takes a query, retrieves top-k chunks, and returns them as JSON. Test independently of the LLM.

Stream the answer with citation markers. Reject answers that have no retrieved context — fall back to "I do not know."

Use ChatGPT to generate 50 question/answer pairs from the corpus. Score retrieval recall before you ship.

Cursor scaffold: a minimal chat UI on a /docs route. Source-of-truth citations as inline pills.

Log every query, the retrieved chunks, and the final answer. Review weekly — most RAG bugs are retrieval bugs.