A 200-line RAG that beats most demos

Most ‘RAG demos’ ship 1,200 lines of LangChain when 200 lines of Postgres do the job. Here is the pattern we keep using on Day 2 of prototypes.

Schema: one table for chunks, one column for the embedding, an HNSW index, and a simple top-k query with cosine distance. That is it.

Where this loses: very large corpora (>10M chunks), cross-encoder reranking, and exotic hybrid retrieval. For most studio prototypes, none of those matter on Day 7.