RAG 101: skip the vector DB drama until three basics hold

RAG sounds heavy, but v1 can stay lean: cleaned docs, simple splits, basic retrieval, LLM summary. Most failures come from these three gaps.

1. Chunks should respect document structure

Split on headings, clauses, and tables—not a blind 500‑character razor. Contracts and legal text especially—“half sentences” in results break trust.

2. Answers must point back to sources

Surface quotes or page numbers in UI or logs. RAG without provenance leaves you defenseless when the model hallucinates.

3. Keep a tiny eval set

A dozen real questions plus expected bullet answers—rerun after every model or chunk change.Vibes lie; spreadsheets don’t.

Add vector stores and rerankers later; without the trio above, engineering only amplifies noise.