Skip to content

Ask my notes

This is my RAG That Doesn't Lie post, running live on itself: lexical retrieval over everything I've written here, answers that cite the exact section, and a hard gate that refuses before calling the model when retrieval confidence is below 3.6.

35/47evals passing

How it works

  1. Retrieve

    Posts are chunked by section at build time, about 60 sections. Each question is ranked with hand-rolled BM25, no vector database: at this corpus size lexical retrieval is simpler, free and explainable.

  2. Gate

    If the best section scores below 3.6, or the match is too thin to trust, Ask refuses before any model call. You get an honest refusal instead of a confident guess.

  3. Answer

    When the gate passes, the top sections go to Claude with instructions to answer only from them and cite each claim. Every answer shows its retrieval, so you can check my homework.