Using Jev for just-in-time context classification instead of RAG

Josh Rosen @JoshARosenOriginally posted 2026.09.18

Open original on X

I predict Jev revolutionizes context management. Picture a big pile of context, memory, markdown files, too much to fit in the context window. Now imagine a just-in-time classifier of that context using Jev. The harness around the LLM provides a built-in “classify context” tool that takes classification instructions from the LLM and returns a subset of context. Way better than conventional embeddings/RAG because you can change the classification semantics per request. Meaning can be derived on read, not write.