The assistant on this page has no LLM

Jul 2026 · Shaaf Kashan

The chat assistant on this site answers questions about me and my work. There's no large language model behind it and no API key; your questions are answered in the browser. It runs entirely in your browser on a few kilobytes of maths.

That was deliberate. The knowledge is small and fixed: a curated set of answers about my projects and background. For that job a retrieval engine is enough (TF-IDF ranking, with fuzzy matching so typos still land). It's instant, it costs nothing to run, it can't invent a claim I never made, and your questions stay in your browser unless you rate an answer, which sends that question so I can improve the assistant.

An LLM would have been the easy thing to reach for, and worse on every axis I cared about here: cost per message, latency, privacy, and the odds of it confidently making something up about me. The real question is usually not "which model" but "does this need a model at all".

Knowing when AI earns its place in a system, and when a simpler and more auditable approach wins, is most of what I think about.