My RAG evaluation was lying to me, twice!!

I built the usual thing: a CLI that ingests documents, embeds them, and answers questions with citations. Bun, TypeScript, SQLite with sqlite-vec for the vector index, Voyage for embeddings, Claude for generation. About 700 lines.

Source: [Dev.to](https://dev.to/jerry_qian_72ad9b2386d3d0/my-rag-evaluation-was-lying-to-me-twice-45n9)