Blog
Insights & Craft
Deep dives into technical problems, architecture patterns, and building production AI into real applications.
Building a RAG Agent That Knows When It Doesn't Know
Retrieval is a similarity guess and a language model handed a bad guess answers anyway, fluently and with total confidence. Here is the six-node agent I used to make a chatbot about me decline instead of invent, and what that architecture actually costs per question.
Three days, fifteen commits, a monorepo I barely typed. What made it work was 652 lines of rules written before any code existed: a spec gate, hard boundaries and a stop-on-failure rule. Here is what those rules caught, starting with a manifest that covered 38 of 66 files.
A Next.js client, an Express API, MongoDB and image uploads, onto Vercel. Four separate bugs produced the exact same symptom, a broken image, and not one of them logged anything. Here is all of it: the read-only disk, the 4.5MB ceiling, and the helper that deleted the only host that mattered.